Field operations, asset models and the cost of a wrong record

Physical AI and Logistics Management for Telecommunications

Telecom operators run one of the largest distributed physical estates in any industry. This review states ten hypotheses about where Physical AI is carrying value across that estate, tests each against graded evidence, and resolves each to a position on the trajectory and the measured change that would move it, nine to a binding constraint and the tenth to the measurement that would name one.

Ronnie Gomez, Industrial Research Fellow · The Charlot Lab

1. Headline numbers

The table below collects the figures this review returns to. Grades follow the Institute's convention: verified means independently replicated or drawn from a regulator, standards body or peer-reviewed source; reported means a named analyst house or operator stated it without independent confirmation; self-published means the number comes from a party selling the thing being measured; modelled means it was computed rather than observed. Where a figure is the author's arithmetic on a cited primary, it carries index 42 and is graded modelled throughout the paper. Assumptions used inside a modelled figure are labelled as assumptions at the point of use and in the table cell that carries them.

Two of these entries frame the rest. The first is the ratio of recurring operating cost to expansion capital in US wireless, which WIA's own numbers put at 5.4 to 1 4,42, placing the larger addressable pool in operations. The second is the share of BVLOS flights in the FAA's flagship BEYOND programme flown without a visual observer at the site, which was under 2% through October 2024 23, placing the crewless configuration that carries the largest published savings early on its regulatory curve rather than at scale. Section 9 sets out what governs the rate of change on that curve, and section 10 sets out the physical conditions that bound the flight itself.

Table 1. Reference figures for the US and global telecom physical estate, with verification grade.

2. What is in scope, and how was it measured?

Physical AI in this review means machines that sense, move, manipulate or navigate in the field and decide onboard what to do next: uncrewed aircraft flying tower orbits, ground vehicles working inside exchanges and data halls, robotic boring and trenching equipment, and the perception and planning software that turns sensor output into a usable model of the plant. It also covers the diagnosis and dispatch systems that decide whether anything gets sent at all, because on the current evidence those carry the largest measured value of any application in telecom. Logistics management means the allocation and movement of physical resources against that plant: crew routing and scheduling, spares and consumables, work sequencing, and restoration after storm or excavation damage.

The review excludes network-layer autonomy that never produces a physical action. Self-optimising RAN, traffic steering and energy-saving sleep modes enter here only where the output is a decision to dispatch or not to dispatch. Customer-premises equipment and consumer self-install are out of scope except where they bear on truck-roll avoidance.

The fellowship framing comes from Ronnie Gomez's project statement 43, which names four capabilities: locate and inspect assets, route field crews, track inventory, and coordinate the people and machines that service the network. The statement also names the scientific problem correctly, which is giving those systems a reliable model of the physical plant and the judgment to act on it when conditions get messy. Sections 4 and 11 take that problem as the organising question, and sections 6 through 10 test the four capabilities against what the public record supports. Inventory carries the least public evidence, which is itself a finding; it is stated as hypothesis 10 and resolved in sections 13 and 16.

Method. Two evidence tracks were compiled on 4-5 August 2026, one on economics and one on technology, and merged into 43 sources; a third track on the physical conditions bounding an inspection sortie was compiled on 6 August 2026 and added 19 further sources, numbered 44 to 62. Regulator, standards-body and audit documents were preferred where they exist, which here means the US DOT Office of Inspector General 23, OSHA 11, ANSI/TIA-222 7,60, the FCC exposure rule and bulletin 48,49 and the FCC base-station power limit 51, and federal fuel and highway series 54,55,56,59. Trade-association datasets are used where nothing better was located and are marked member-supplied. Where a figure could only be found in material published by a party selling the product, it is graded self-published and the seller is named in the same sentence. Each source's own limitations are recorded in its reference entry.

How findings are stated. Every finding resolves to three things rather than to a verdict. The first is where on the trajectory the application currently sits, stated in degrees: what works now, at what scale and under what conditions. The second is which constraint is currently binding, named from five technical classes, material science, actual engineering implementation, energy, computation efficiency and AI algorithms, or named as regulatory or workforce where that is what binds. The third is the measured change that would move the boundary, stated as a threshold, a factor, a date or a rule. Section 3 states the hypotheses in that form and section 16 resolves each into a table with those columns.

Physical quantities enter as test conditions on the usage hypotheses rather than as a separate subject. Section 10 computes the energy, exposure and optical conditions that bound an aerial inspection sortie from manufacturer datasheets, federal regulations, government series and standard textbook results, and folds them into hypotheses 1, 3 and 4. Those computations are graded modelled and the assumptions inside them are labelled where used.

Limits worth stating up front. bls.gov blocked automated retrieval, so reference 3 rests on search-index extracts rather than a direct read. The TSIA benchmark data in 15 sits behind membership and is quoted from secondary description. The peer-reviewed review at 41 was truncated in the compiled evidence record and its citation needs completing before publication; it is used only for the qualitative claim that contact servicing of towers remains at feasibility stage. ANSI/TIA-222 and the ASME visual examination code are paywalled and were not read directly 53,60. nrel.gov did not resolve, so no figure was taken from Fleet DNA 62. Nothing here is investment advice.

One tie-break is needed because the two middle grades overlap. Where a party is both the operator of a programme and the seller of the thing being measured, or is stating a result about its own operations, the claim is graded self-published regardless of which outlet relayed it. Trade-press repetition of a company statement is not independent confirmation and does not raise a grade.

3. Which hypotheses does this test?

Ten hypotheses organise the evidence that follows. Each is stated as a claim about the market or about usage, with the reason it matters and the evidence that would settle it. Each asks where on the trajectory an application currently sits, and section 16 answers that in degrees, with the constraint that binds, or for hypothesis 10 the measurement that would name one, and the measured change that would move it.

Hypothesis 1. Aerial structural inspection is becoming a material line in operator cost. It matters because it is the application with the widest vendor coverage and the one most often used to justify a drone programme. Settling it needs a scope-matched, operator-audited comparison of an uncrewed pass against a climbing crew; the guyed to self-supporting mix of the estate, which sets national inspection volume; and the share of inspection flights flown without a posted visual observer. Sections 8, 9 and 10 supply the evidence.

Hypothesis 2. The asset digital twin is becoming the system of record for the physical plant, displacing the RF data sheet and the structural drawing as the authority for what stands on a structure. It matters because a system of record is a durable position with contractual obligations attached, and because every autonomous application here consumes that record. Settling it needs per-parameter accuracy against a physically verified baseline, a published cost of keeping a twin current over five years, and evidence of the twin being referenced in lease, co-location or structural-analysis documents. Sections 8 and 11 supply the evidence.

Hypothesis 3. Crewless aerial autonomy is available at national scale under the current regulatory gate, meaning an operator can fly inspection missions with no person at the site. It matters because the labour term in every drone inspection business case depends on it. Settling it needs the share of flights flown without a visual observer, the number and industry mix of BVLOS approvals, and the detect-and-avoid and operator-certification requirements in the Part 108 final rule. Section 9 supplies the evidence and section 10 the physical conditions the flight must also meet.

Hypothesis 4. A plant model of the quality autonomous action requires can be built and kept current with today's methods, meaning a record good enough, and explicit enough about its own uncertainty, for a machine to plan physical work against. It matters because it is the scientific problem the fellowship names and because every application here inherits the error in that record. Settling it needs registry error measured against physical truth at each level of the hierarchy, per-parameter accuracy budgets tied to task tolerance, and an abstention rate reported alongside every accuracy figure. Sections 4, 10 and 11 supply the evidence.

Hypothesis 5. Dispatch decision systems carry more measured value today than embodied field systems. It matters because it orders where an operator's first increment of capability lands and because it requires no new hardware in the field. Settling it needs precision and recall at the chosen operating point rather than accuracy alone, a truck-roll cost decomposed by fault class, and a field-operations labour figure separated from total telecom labour. Section 7 supplies the evidence.

Hypothesis 6. Labour scarcity is the mechanism that funds telecom field automation, through rising wages against a shrinking crew base. It matters because most business cases here are denominated in avoided labour hours. Settling it needs a neutral series separating fiber construction crews from wireless tower crews, resolved regionally, with wages for each. Section 5 supplies the evidence.

Hypothesis 7. Automation of buildout, meaning boring, trenching and method selection, moves the fiber cost stack. It matters because labour is 67.9% of median total underground cost and buildout carries the largest unit costs in this review 1,42. Settling it needs a published bore rate, cost per drop and failure rate for the robotic systems in trial, and a comparison against manual methods on matched ground. Sections 6 and 8 supply the evidence.

Hypothesis 8. Buried-plant locate accuracy is the largest measurable damage-prevention target in the domain. It matters because it is the only place in this review where the cost of an inaccurate plant model is counted directly, in units of broken infrastructure. Settling it needs the telecom cross-tabulation of the damage root-cause data, positional uncertainty captured at time of installation, and a sensing performance figure across the soil conditions the work meets. Section 4 supplies the evidence.

Hypothesis 9. Robotics enters telecom first as a carrier revenue line rather than an internal cost line, with operators selling connectivity and services to other people's robots before fielding robots against their own plant. It matters because it changes which part of the business sponsors the capability and which metrics govern it. Settling it needs a reported robotics services revenue line at a carrier, and independent verification of units in service, uptime and task success at the announced programmes. Section 8 supplies the evidence.

Hypothesis 10. Field inventory and spares availability is a measurable driver of failed first visits, with parts rather than diagnosis or access accounting for a material share. It matters because inventory is one of the four capabilities in the fellowship scope and because the remedy for a parts failure differs entirely from the remedy for a diagnosis failure. Settling it needs a decomposition of first-visit failure into diagnosis, access and parts causes, producible from work-order data an operator already holds. Sections 7 and 13 supply what the public record contains.

4. How big is the estate?

Any autonomous system that acts on telecom plant needs to know what is there. The industry does not agree on the answer even in aggregate. CTIA's 2025 survey, drawn from its own members, reports 447,605 cell sites in operation at year-end 2024, of which 166,264 were small cells, up 6% from 156,787 the year before 6. WIA's 2025 report, modelled by iGR, gives 254,850 macrocell sites, 158,500 purpose-built towers, 198,100 outdoor small cells and 830,350 indoor small cell nodes 4. WIA's indoor node count on its own exceeds CTIA's entire site total. This review did not locate published definitions from either body that would allow the two counts to be reconciled, and both are association surveys of member-supplied figures with no independent audit.

This is the top of a hierarchy of registry error rather than a footnote about survey methodology. CTIA's 447,605 and WIA's 254,850 macrocells differ by 1.76x, and against WIA's 1,283,300 all-node total the ratio is 2.9x; the two also differ in reference date and in scope. The same problem appears at the level of a single asset. In a Bentley-sponsored validation of the OpenTower iQ digital twin, one of four towers examined had a RAD centre documented at 180 ft that the twin measured at 175 ft 28. A five-foot difference on a parameter that drives structural loading and RF design, in the one case of four the study details, is the concrete version of the same problem, and it is the quantity hypothesis 4 turns on. The study does not establish that the other three sites were correct on that parameter.

Below ground the evidence is much better and the measured cost is larger. The Common Ground Alliance's 2024 DIRT Report, published 28 August 2025, analysed 196,977 unique damage reports to buried utilities across North America, with telecommunications facilities accounting for close to half of reported damages 31. This is the largest dataset bearing on the fellow's problem and the only one located by this review that measures the cost of a wrong plant model directly, in units of broken infrastructure. It is the evidence base for hypothesis 8.

The root-cause breakdown is where the operational signal sits. CGA attributes 24.5% of damages to failure to notify 811, 16.1% to an excavator failing to maintain clearance, 11.9% to a facility not being marked at all, 8.6% to inaccurate marking, and 6.8% to improper excavation 31. Two of those categories, together 20.5% of damages, are locator error, which works out to roughly 40,400 damages a year traceable to the buried-asset model being wrong or unfindable in the field 31,42. If telecom's share of those follows its share of damages overall, the telecom-specific figure is on the order of 20,000 a year; the DIRT report does not contain the cross-tabulation that would confirm it, and producing that cross-tabulation is one of the measurements that would settle hypothesis 8 42.

The five root causes listed above are a subset of the top ten to which CGA attributes 85% of reported damages, and they sum to 67.9% 31,42. CGA itself treats DIRT as an undercount, since submission is voluntary and the reporting population shifts year to year, and the CGA Index moving from 94.0 to 96.7 partly reflects who reported rather than what happened.

There is a second, schedule-shaped cost in the same report. Across eight 811 centres, excavators faced an average 38% chance of being unable to start work on time because locate responses were incomplete 31. That figure covers eight centres and is not nationally representative, and it puts a number on what an untrusted registry does to a construction programme before anything is dug at all. The 20.5% locate-error share and the 38% late-start rate are the two published series in which an improvement to the subsurface record would register.

5. What is the field labour supply, and why do sources disagree?

The standard case for automating telecom field work opens with a labour shortage. That is hypothesis 6, and the public numbers behind it point in different directions by a wide margin.

On one side, Continuum Capital, commissioned by the Fiber Broadband Association and the Power & Communication Contractors Association and presented at Fiber Connect in July 2024, put the requirement at about 28,000 additional broadband construction workers and 30,000 additional broadband technicians to absorb planned federal and state funding, plus roughly 119,000 more over ten years to cover retirement and attrition, with the shortfall pushing projects two to three years out 2. It is a forecast, commissioned by two trade associations whose members benefit from workforce funding, published before both the BEAD restructuring and the 2025 contraction in wireless construction. The Communications Workers of America has publicly called US telecom workforce shortage claims misinformation 2.

On the other side, WIA's own 2025 data shows the US wireless infrastructure workforce falling to 342,350 full-time equivalents from 368,750 in 2024, a decline of 26,400 jobs or 7.2% 4. Wireless Estimator, tracking against its own twenty-year record, reported tower construction hiring at a two-decade low in Q4 2025, with help-wanted postings down about 60% against recent years while jobseeker registrations rose more than 300% 5. That source is a single proprietary job board reporting on its own commercial base, and postings volume is a noisy proxy, and the direction is unambiguous and opposite to the fiber trade bodies' direction in the same period.

These are not strictly the same labour pool. Fiber construction and wireless tower work draw on overlapping but distinct crews, and a fiber shortage can coexist with wireless slack. On the public data the scarcity is regional and segment-specific, so an automation case denominated on wage scarcity carries an input that no source in this review resolves. The constraint binding hypothesis 6 is workforce, compounded by the absence of a neutral series, and what moves it is a neutral series separating the two crew types regionally, which this review did not locate published in that form by any of the bodies surveyed here.

The BLS projection supplies a term all five sources can carry. Employment of telecommunications equipment installers and repairers, excluding line installers, is projected to decline 3% from 2024 to 2034, while about 23,200 openings a year are projected, all of them arising from workers transferring out of the occupation or leaving the labour force 3. Demand for the occupation is shrinking and the replacement requirement is nonetheless large. What walks out with those workers is undocumented site knowledge: which cabinet the spare fibre is coiled in, which gate code works, which pole is the one that was replaced after the 2019 storm. That is a knowledge-capture problem with a machine-readable answer, and on the evidence assembled here it is a firmer basis for an asset digital twin programme than any wage forecast, because the 23,200 figure holds whichever side of the scarcity argument is right.

Safety is the other standing argument for removing people from towers, and the rate is contested. OSHA's own count, which has not been updated past 2016, records 13 communication tower fatalities in 2013, 12 in 2014, 3 in 2015 and 6 in 2016 11. The rate dispute is sharper. CWA derived 19.3 deaths per 100,000 workers from a five-year average of 5.6 deaths over a 29,000-worker denominator; safety experts using a workforce denominator of about 37,000 and a fatality database put 2021 at 10.8 per 100,000 on 4 deaths and 2022 at 8.4 on 3 deaths 12. Both rates are carried here as published: three deaths against 37,000 is 8.1 per 100,000, so the published 8.4 implies a denominator nearer 35,700, and this review did not locate year-specific denominators in that source 12,42. Decomposing the 1.8x gap, about 1.4x of it comes from the numerator, a five-year average against a single year, and about 1.28x from the denominator, 37,000 against 29,000 12,42. Both parties are interested; CWA's survey had 103 respondents and Wireless Estimator's database is proprietary and internally inconsistent on its own date range. This review did not locate a BLS occupational code isolating tower technicians, so no neutral rate is available from the sources compiled here. NATE's framing is the more actionable one: technicians typically ascend and descend six to ten times a day, and eliminating two or three of those climbs is argued to reduce fatigue-driven risk 13. That is a claim about removing some climbs from a day rather than removing the crew, and it matches the position on the trajectory that the deployed record supports.

6. What does a fiber mile cost?

The Fiber Broadband Association's third annual deployment cost report, produced by Cartesian and published in January 2026, is the best public per-unit benchmark in this area and the denominator any automation claim under hypothesis 7 has to beat. Median reported cost to deploy fiber in the US in 2025 was $18.00 per foot underground and $8.00 per foot aerial, roughly $95,000 and $42,000 per mile 1,42. Underground rose 3% year on year and aerial 14%. The sample covers projects in 38 states, gathered by phone interview and online survey in September and October 2025, self-reported by operators and contractors, with sample size undisclosed in the executive summary and each year's panel independent, so the year-on-year comparison is not like-for-like.

Cartesian reports labour at 72% of underground deployment cost and 64% of aerial, with median labour of $12.23/ft and $4.50/ft against median materials of $4.65/ft and $2.50/ft 1. Those shares reproduce exactly as labour divided by labour plus materials: 72.5% and 64.3% 42. Against the median total cost figures they are 67.9% and 56.2% 42. The gap means about $1.12/ft underground and $1.00/ft aerial, or 6.2% and 12.5% of the reported total, sits outside the labour-and-materials frame in permitting, make-ready, engineering and surface restoration 42. Medians of different quantities are not additive, so this decomposition is indicative rather than exact, and the exact match of both stated shares to the labour-plus-materials denominator makes the inference reasonably firm. A labour-productivity saving modelled against total project cost should use the lower share.

The most useful calibration point in the report is the internal-versus-contractor comparison. Using internal crews rather than contractors cut labour cost 13% underground, $11.13/ft against $12.73/ft, and 11% aerial, $4.00/ft against $4.50/ft 1. Cartesian notes that internal and external labour figures may carry different context by respondent, and some internal numbers may exclude overhead that contractor invoices include. Taken at face value, that delta is worth $8,448 per underground mile and $2,640 per aerial mile, which is 8.9% and 6.2% of total deployment cost 42. Vertical integration is the best-documented organisational lever available and it moves total cost by single digits, which sets the bar for hypothesis 7: an autonomous system claiming materially more than that is claiming something structurally different from better scheduling, and the report of it should say what.

Method selection is the larger lever and the one most amenable to an automated planner working from terrain and subsurface data. Trenching carries the highest cost of the common methods, 60% above plowing and 6% above directional boring on median costs, and direct burial runs about 40% cheaper than laying in new conduit 1. Moving a mile from trenching to plowing where ground conditions permit is worth about 37.5% of that method's cost 42, which is several times any software-driven productivity claim in the field-service literature. Cartesian notes that direct burial offers less flexibility and potentially higher maintenance cost, so the 40% is a first-cost figure and this review did not locate a lifecycle counterpart for it.

Deutsche Telekom's own figures make the same point at a different scale: about 85 euro per metre for traditional underground deployment, about 65 euro for reduced-depth less-invasive methods, and about 10 euro per metre overground 36. Going aerial is roughly an eight-fold saving, larger than anything a robot has demonstrated in the record reviewed here. These are the operator's own numbers with no methodology stated alongside them, and German civils costs are dominated by local reinstatement and permitting rules that do not transfer. For orientation, the US median underground figure converts to about $59 per metre 42, the same order once currency is set aside.

The cost curve is moving upward and the respondents name the reason. 92% reported deployment cost increases in 2025 and 88% expect further increases in 2026, attributing them primarily to labour rates, then materials, permitting and make-ready 1. The expectation data has one year of calibration behind it: 80% expected a 2025 increase and 92% then reported one, so this panel under-forecasts rather than over-forecasts, on a small self-selected sample. A rising labour line raises the value of every point of productivity that hypothesis 7 is about.

Table 2. The 2025 US fiber deployment cost stack, with author's derivations marked. Source: FBA/Cartesian 1; derivations 42.

7. What has to be done again and again?

WIA and iGR put 2025 US cellular network operating expenditure at about $54.7bn against $10.2bn of capacity-and-coverage capex 4, so recurring operations run at 5.4 times expansion capital on those two figures 42. The $10.2bn is capacity and coverage only and is not comparable to CTIA's roughly $30bn total wireless capex for 2024, which uses a wider definition 6. Even allowing for that, the ordering is the fellow's thesis in the industry's own numbers, and it is the premise hypothesis 5 tests.

Per-site denominators are slippery. Dividing $54.7bn by WIA's 254,850 macrocells gives about $214,600 per macro site per year, and the same opex also covers 198,100 outdoor small cells and 830,350 indoor nodes, so spreading it across all 1,283,300 nodes gives about $42,600 4,42. The bracket for cost per site is therefore $43,000 to $215,000 depending entirely on which sites the opex belongs to, and this review did not locate a published split.

At global scale the two most-cited houses differ on the largest line. MTN Consulting's 2Q25 tracker covers 140 operators and puts annualised capex at $292.9bn, down 2.6% year on year and the lowest twelve-month total since 2011, with labour at $258.1bn across headcount of 4.36 million, 21.7% of opex against 22.2% a year earlier, and EBITDA margin at 33.8% 19. Omdia puts labour opex at about $300bn, falling 2.2%, within total global opex of $1.63 trillion against revenue of $2.01 trillion 20. The two are 16% apart 42 and do not cover the same universe, so an unknown part of the gap is population coverage rather than disagreement about the same quantity. MTN's revenue base of about $1.8 trillion 19 sits between Deloitte's $1.55tn 21 and Omdia's $2.01tn 20. Omdia attributes part of the 2024 decline to divestments rather than efficiency and notes that disclosures of quantifiable AI savings remain scarce. A later MTN tracker is published; 2Q25 is used throughout for consistency with the labour split.

MTN's figures yield an average fully loaded labour cost of about $59,200 per telecom employee per year 19,42, the most useful single anchor available. This review did not locate a source isolating field-operations labour from total telecom labour: the 4.36 million includes retail, network engineering, IT, customer care and back office. The field-crew subset is the denominator every value claim in this subject is measured against, and it is the first item in section 13.

Capital intensity is similarly contested. Dell'Oro's 2026 forecast, covering about fifty providers representing roughly 80% of global capex, has capex flat in 2025, declining 2% in 2026, then growing at about 1% CAGR to 2030, with capex-to-revenue approaching 14% by 2029 and wireless intensity around 11%, down seven points from the 5G peak 18. Its own September 2025 release gave 15% and 12-13% for the same horizon. Deloitte puts capex intensity around 18% against global revenue of roughly $1.55 trillion citing IDC 21. Two named houses span 14% to 18% on capital intensity, against revenue bases that differ by 1.3x across the three houses publishing one, $1.55tn at Deloitte, about $1.8tn at MTN and $2.01tn at Omdia 42. Dell'Oro's own year-on-year revision is comparable to the effect a successful field-automation programme would have on capital intensity, which sets the measurement requirement: an automation effect is detectable in matched-cohort instrumentation and is smaller than the spread between the houses measuring the aggregate. Carrier guidance shows divergence rather than a trend, with AT&T at roughly $22bn in 2025 and $23-24bn in 2026, Verizon at $17bn falling to $16-16.5bn, and T-Mobile at $10bn in both years 22.

Field-service benchmarking gives the realistic headroom for hypothesis 5. TSIA puts the industry median at 87% of on-site incidents resolved in a single visit, with engineer utilisation at 83% industry-wide and 90.6% among top performers 15. Roughly one visit in eight needs a second, and about a sixth of technician time is already non-productive. Closing the utilisation gap to top-performer level is worth 9.2% more productive hours from the same headcount with no new technology at all 42, and that 9.2% is the non-technology baseline against which an autonomous scheduler is measured.

The truck roll is the unit most of this field is denominated in, and the published figure for it carries the weakest sourcing of any number in this review: $150-500 direct, rising to roughly $1,000 fully loaded, with the $1,000 attributed to TSIA 14. No primary TSIA publication stating $1,000 could be located, and every citation runs through a vendor selling truck-roll avoidance. The range spans nearly sevenfold and mixes residential installs with enterprise service. An operator-audited figure per fault class replaces it.

The operator-attributed results are consistent with each other and with that headroom. Vodafone reduced technician dispatch rate by 10%, raised first-contact resolution five points and raised NPS fifteen points using remote visual assistance 16. Vodafone's Field Technician Assist, built by Celfocus on Google Cloud, delivers top-three recommended actions to field engineers and is credited with a 28% reduction in repeated field visits 32; applied to TSIA's 13% repeat rate that removes about 3.6% of on-site incidents, or about 3.2% of total visits once the repeat visits are counted into the denominator 15,32,42. Comcast reports a machine learning system predicting with about 90% accuracy whether a technician needs to be dispatched, credited with saving tens of millions of dollars in avoided truck rolls 17,34. No denominator is applied to that here: the truck rolls sit in Connectivity & Platforms, and dividing a segment-level saving by total corporate revenue, which includes NBCUniversal and Sky, would measure it against the wrong base. Accuracy without a precision and recall split does not determine the operating point for an imbalanced dispatch decision, where a false negative leaves a customer broken and a false positive costs a van, so the constraint on the next increment is AI algorithms, in the specific form of decision calibration under class imbalance rather than raw accuracy. Elsewhere in the same vendor literature, claims of 60% truck-roll reduction appear with no operator attribution attached.

TSIA's 87% first-visit resolution does not decompose into diagnosis failure, access failure and parts failure 15, and that decomposition is where hypothesis 10 lives. The public record reviewed here contains no telecom figure for field inventory carrying cost, stockout rate or contribution to failed first visits, so inventory is the one capability in the fellowship scope with no evidenced position on the trajectory.

8. What is deployed today?

The largest asset-twin programme in the public record is Verizon's, and it is the primary evidence for hypothesis 2. Verizon states that it has drone-derived twins across tens of thousands of cell sites and, from early 2025, automatic capture and digitisation of every new build and modification, with drone capture cutting some disaster-response site assessments from several hours to about fifteen minutes 25. The software partner is vHive, which raised a $25M Series B in April 2022 led by PSG and reports generating tens of thousands of twins in a year across telecom, construction and cranes, with Verizon and Deutsche Telekom named as customers 26; Deutsche Telekom's Telekom Innovation Pool participated in that round, so DT is an investor in the supplier as well as a customer. Consilience Analytics states it has processed over 35,000 twins since 2019 30, a count that is global, spans sectors and is cumulative across repeat captures, so it cannot be set against the US tower estate as a coverage figure and no such ratio is computed here. This review did not locate independent confirmation of scale, cadence or accuracy for either programme.

The mechanism inside Verizon's programme is worth separating from the hardware. Capture on every change is a process commitment that keeps a plant model from drifting, and the drone is the instrument that makes it affordable. Site count is given as a range rather than a figure, and this review did not locate an accuracy metric, a cost per twin or a refresh cadence for the legacy estate. American Tower and Crown Castle both build drone-derived twins, and American Tower reports being rated industry leader in quality of asset information in its own 2025 customer survey 10. This review did not locate a scan count or an accuracy figure from either towerco.

On accuracy the fullest public disclosure is Bentley's sponsored validation of OpenTower iQ, which found automated equipment detection and dimension measurement 95% accurate and tilt and azimuth accurate within a 5% error margin, with tower height on one example reading 72 ft 3 in against 73 ft in the design documents, a 1.03% difference 28,42. The sample was four towers drawn from a group of fifty, and an accuracy rate from four observations carries a binomial interval spanning most of the usable range, so the 95% is recorded here as the vendor's characterisation and not as a measurement. The comparison baseline was best-existing data, meaning RF data sheets, old structural analyses and CAD drawings. That is the crux for hypothesis 2: the study measures agreement with the records, and the reliability of the records is the quantity in question.

Bentley's economics are a model rather than a measurement. The company estimates manual tower operations and maintenance at about $5,700 per tower against $2,200 with OpenTower iQ, a $3,500 annual saving, alongside 50% reporting time savings, a 15% annual opex reduction and 42% lower ten-year TCO per tower 29. The per-tower figures imply a 61% cut in the stated O&M line, which the same document does not reconcile with its own 15% and 42% claims; the three percentages are computed over different denominators that are never named 42. This review did not locate a matched-cohort before-and-after published by any operator. The model excludes the cost of maintaining twin currency, re-capture after each site modification, and pilot and mobilisation cost, all of which fall on the operator. Applied to WIA's 158,500 purpose-built towers, the $3,500 would come to about $555M a year, roughly 1% of WIA's $54.7bn cellular opex, though the two are different denominators and the comparison is only for scale 4,29,42.

Aerial structural inspection is the most widely offered application in this domain, and every published price comparison located here is graded self-published. Providers quote $350-600 per tower against $800-1,500 for a climbing crew, with drone crews completing four to six towers a day against one to two for a two-person climbing crew, and 30-50% overall cost savings 8. This review located no independent, operator-audited or peer-reviewed comparison, and the comparison is not scope-matched, since a drone pass performs no torque checks, connector work, grounding continuity testing or hands-on remediation. On the quoted midpoints the implied saving is 59%, and the full spread of the quoted ranges runs from 25% to 77% 42, so the quoted 30-50% is a narrower band than the 25-77% implied by the providers' own price ranges.

ANSI/TIA-222 recommends major structural inspection every three years for guyed masts and every five years for self-supporting structures, plus after severe weather, and TIA's TR-14 committee states explicitly that this is a recommendation rather than a mandate 7. Applied to 158,500 purpose-built towers that implies roughly 32,000 to 53,000 structural inspections a year depending on the guyed and self-support mix, which this review did not locate published 7,42. At the quoted midpoint saving of $675 per tower, the present national addressable value of drone structural inspection is $21-36M a year, 0.04-0.07% of US cellular opex 4,8,42. That is the position of hypothesis 1 on the value axis today, and section 10 supplies the physical conditions that currently hold it there. Revision I of ANSI/TIA-222, effective 1 January 2024, admits unmanned aircraft as a method for collecting data on existing structures, which is the standards step that makes the practice routine rather than exceptional 60. AT&T announced drone tower inspections in 2016 across roughly 65,000 cell sites where it has equipment and stated that the programme would break even during the initial ramp-up period; an AT&T official declined to quantify the saving beyond that 9. This review did not locate a later cost figure from AT&T, and that absence is not itself evidence of the size of the effect.

Physical work on the plant sits earlier on the trajectory, at trial and early deployment rather than at scale, and it is the evidence for hypothesis 7. Openreach is trialling MicroBORE, a trenchless robotic drilling system developed over five years by BT Group, Synthotech and the University of Surrey from lunar and Martian drilling research, to bore narrow channels under pavements and driveways for FTTP drops; early testing at BT's Suffolk research centre confirmed operational viability and broader deployment is under evaluation 35. This review did not locate a published bore rate, cost per drop, failure rate or comparison against manual moling. Openreach's current programme, 15bn pounds to reach 25 million premises by December 2026, runs on conventional methods at about 600 pounds per premises passed, which is the cost baseline a robotic drop has to beat 35,42. Built Robotics retrofits standard excavators for autonomous trenching in about a day with contractors reporting 20-30% savings, and its volume business is autonomous solar pile driving, including a nationwide agreement with Blattner announced in September 2025 37. The best-funded autonomous trenching company found its volume business in open, repetitive, greenfield geometry; telecom conduit is short-run, congested with existing utilities and constrained by surface restoration, which is a harder geometry for current perception and planning and which names the constraint on hypothesis 7 as AI algorithms plus engineering implementation rather than economics.

Ground robots inside telecom are deployed mainly in controlled indoor environments and as carrier-sold services, which is the evidence for hypothesis 9. AT&T fields Ghost Robotics quadrupeds over FirstNet using its Geocast service for beyond-line-of-sight command and control, with documented use in public safety and defence including Air Force flight-line and perimeter patrol at Tyndall AFB, rather than inspection of AT&T's own plant 38; that is a connectivity product sold to others. NTT DATA began rolling out data centre facility inspection robots across its Japanese sites from April 2023 and moved to commercialise the offering 39, the clearest at-scale ground-robot inspection programme inside telecom-owned facilities, in controlled indoor environments rather than central offices or outside plant. The operator-led embodied deployments are happening in Asia: at MWC Shanghai in June 2026 China Telecom announced a Robotics-as-a-Service model with AgiBot for aviation and logistics, China Unicom described robot deployment for hazard inspection in chemical plants, and State Grid announced plans for roughly 8,500 robots, about 5,000 quadrupeds, 3,000 dual-arm wheeled units and 500 humanoids, across grid inspection and maintenance 40. Those are announced programme figures from state-linked entities, and this review did not locate independent verification of units in service, uptime or task success. If the model works it is a services line for carriers, a different business thesis from internal cost reduction, which is why hypothesis 9 is stated separately. Peer-reviewed work on robots that physically service telecom towers remains at feasibility stage 41.

Table 3. Deployment ledger: what is documented, and at what verification grade.

9. What does the regulator allow for uncrewed inspection?

The US DOT Office of Inspector General audited the FAA's beyond-visual-line-of-sight efforts and published Report AV2025034 on 30 June 2025. It is the only independent audit of BVLOS operational practice located by this review, and it measures the distance between approvals granted and crewless flights actually flown. That distance is hypothesis 3.

The number most often cited is that FAA approvals for BVLOS operations rose from 1,229 in 2020 to 26,870 in 2023, across small-UAS rule waivers, air carrier operating certificates and regulatory exemptions 23. That is an approvals count rather than a measure of autonomous flying. An approval can cover a single flight or thousands, the categories are not comparable, and the count carries no information about mission type or autonomy level.

The operational number is different. Of more than 44,000 BVLOS flights flown under FAA's flagship BEYOND programme through October 2024, fewer than 763, about 2%, were flown without a visual observer. In the parallel PSP and IPA programmes, no operator flew a single flight without a visual observer 23. The economic case for drone tower inspection assumes the person is removed from the site. In the FAA's own flagship programme 98% of BVLOS flights carried a visual observer, which places the crewless configuration early on its deployment curve rather than at scale.

Mission mix is equally instructive. Out of more than 66,000 BEYOND operational flights, 97% were package delivery, about 1% were infrastructure inspections and under 1% were public safety 23. Infrastructure inspection is one of BEYOND's three named mission types and it produced roughly 660 flights across four years 23,42. The regulatory learning curve that will govern telecom tower BVLOS was built almost entirely on parcel delivery, whose flight profile of long straight legs and low hover with no close approach to a radiating structure has little in common with a tower orbit. An FAA manager conceded to the OIG that the programme's KPI design was more tailored to package delivery. Building the inspection-specific evidence base is therefore itself one of the things that moves this boundary, and section 10 sets out which physical quantities that evidence would have to contain.

The waiver base is thin and concentrated. FAA had issued 190 BVLOS waivers to 134 distinct operators as of October 2024, up from 6 in 2020 and 122 in 2023. Two BEYOND lead participants accounted for 95% of all BVLOS operations, and one participant alone flew over 60% of the no-observer flights 23. Waiver counts are not broken out by industry, so the number of the 134 that are telecom or tower-sector operators is not available in the audit. The waiver route has cleared roughly 134 organisations in five years, which is the present throughput of the approval mechanism and the quantity a national programme is sized against.

Two caveats cut in the operator's favour. The OIG's figures cover only the partnership programmes, so commercial waiver holders outside them are not counted, and the OIG separately found FAA's data validation process vulnerable to errors, having mis-categorised at least one operator's observer status. More importantly, telecom operators largely inspect under ordinary Part 107 within visual line of sight, so their flight volume never enters this dataset at all. The dataset measures one thing: the specific configuration the largest savings depend on, flying a site with no crew present, is at an early stage across every mission type it covers. Telecom inspection within visual line of sight under Part 107 is routine and does not enter it.

The gate is Part 108. FAA's BVLOS rule was proposed on 7 August 2025, drew roughly 3,000-3,100 comments, and the final rule reached OIRA review on 10 July 2026, with publication expected late 2026 or early 2027 and a six to twelve month transition after that 24. None of the trade sources reporting this carries an FAA commitment date, and OIRA timelines slip routinely. The rule's detect-and-avoid and operator-certification requirements are not in the public docket, so the compliance cost is the largest unquantified term in the crewless-inspection case and could be large enough to change the per-site economics on its own. Until it publishes, per-site waivers and posted visual observers are the operating configuration, so the labour saving available today is the climb removed from the technician's day rather than the crew removed from the site 13. Two dated quantities move this boundary: the publication date of Part 108 with its transition period, and the content of its detect-and-avoid and certification requirements.

10. What do the physical tests show on hypotheses 1, 3 and 4?

Hypotheses 1 and 3 turn on a sortie count, and hypothesis 4 on what a sortie can measure. A provider quoting $350 to $600 per tower and four to six towers a day is quoting an aircraft that flies a certain number of times per site, for a certain number of minutes, at a certain distance from the steel, with a sensor of a certain resolving power 8. Those quantities are set by battery energy, the radiofrequency field around a live antenna, and geometrical optics, and all are computable from manufacturer datasheets, federal regulations and standard textbook results. This review did not locate them in any of the vendor comparisons it collected. Every number below is a datasheet value, a federal standard or government series, a standard textbook result, a named industry benchmarking figure, arithmetic on those, or an assumption stated and labelled as such at the point of use, and the arithmetic is graded modelled throughout.

Energy per sortie. A DJI TB65 pack holds 263.2 Wh in about 1.35 kg, 195 Wh/kg, and the Matrice 4 series pack 99.5 Wh in 401 g, 248 Wh/kg 44,45, so a two-battery Matrice 350 sortie leaves the ground with 526.4 Wh, or 1.90 MJ, and a Matrice 4E sortie with 0.36 MJ. DJI publishes 55 minutes for the Matrice 350 without payload at about 8 m/s in still air to a flat battery, and for the Matrice 4E a 49-minute cruise figure at about 9 m/s alongside a 42-minute hover figure 44,45. Taking that 42 to 49 ratio as the hover penalty gives 47.1 minutes, and scaling by mass through the three-halves power law that momentum theory gives for induced power 57, a Matrice 350 carrying the 800 g Zenmuse P1 has about 40 minutes of hover endurance and about 30 minutes once a 25% reserve is held back, that reserve being an operating convention adopted here rather than a figure from a source. The chain is modelled and has three weak links: transferring a ratio measured on a 1.2 kg airframe to one five times heavier; applying a 9 m/s-referenced ratio to an 8 m/s-referenced figure, so the two cruise conditions differ in airspeed; and the treatment of power terms, since profile power on a fixed-pitch rotor scales on close to the same three-halves exponent, leaving only avionics and payload draw genuinely thrust-independent, so the mass exponent overstates the payload penalty by that term alone. Momentum theory further assumes a constant figure of merit, which this review did not locate published for either rotor 57, and this review did not locate a hover endurance figure for the Matrice 350 RTK in DJI's published specifications. Two published wind figures are not the same quantity: the Matrice 350 RTK gives 12 m/s unqualified while the Matrice 4 series gives 12 m/s explicitly as resistance during takeoff and landing, and this review did not locate an in-flight wind figure for the latter 44,45.

The exposure field is the one term in this chain whose reference level is fixed by federal regulation. A macro site MPE study filed by AT&T's consultant for a Massachusetts site lists eight technology channels on a single sector, radiation centre 75 ft, and in the study's own column order they are UMTS850 at 59.04 dBm ERP, LTE700DE at 59.55, LTE1900 at 64.63, LTE700BC at 63.35, LTE850 at 61.25, LTE WCS at 64.55, 5G850 at 60.00 and LTE700DE KMW at 58.75, with per-channel gains of 9.75 to 15.55 dBd 50. Those sum to 12.70 kW ERP and, at the 1.64 dipole reference factor, 20.84 kW EIRP, with per-band subtotals of 6255 W at 700 MHz, 5142 W at 850 MHz, 4763 W at 1900 MHz and 4676 W in WCS. FCC exposure limits are frequency-dependent below 1500 MHz, at f/1500 mW/cm2 for general population and f/300 mW/cm2 for occupational exposure, giving 0.501 and 2.503 mW/cm2 at 751 MHz and 0.588 and 2.938 at 881.5 MHz, against the flat 1.0 and 5.0 that apply to the 1900 MHz and WCS channels 48. OET-65 requires each emitter to be assessed against its own applicable limit with the percentage contributions then added 49. Summing the fractions and applying the far-field formula in which power density equals radiated power times gain over four pi radius squared 49, the sector reaches the composite general population limit at 15.6 m on the main beam and the composite occupational limit at 7.0 m, the composite general population limit being 0.679 mW/cm2. Assessing all eight channels against a single 1.0 mW/cm2 limit instead puts the contour at 12.9 m, where the correctly summed fraction is 1.47, or 47% over the limit; since power density falls with the square of range that 12.9 m figure is 18% closer than the summing rule gives, and equivalently the summing rule places the contour 21% farther out. The result is insensitive to the exact 700 MHz channel centre, moving only between 15.8 and 15.6 m across 717 to 768 MHz. This is one site and one carrier and is not representative of the estate.

Both distances rest on the far-field equation, which OET-65 states over-predicts power density in the near field, offering Tell's cylindrical model for sector antennas as the near-field alternative with a crossover depending on aperture and gain 49. Setting the two equal gives a crossover radius of gain times beamwidth in degrees times aperture height divided by 720; on the study's tabulated gains, with a 65 degree beamwidth and 1.8 m aperture assumed here rather than taken from the study, crossover falls between 2.5 and 9.6 m and scales linearly with each assumption. The 15.6 m figure sits past crossover, so the far-field equation is the appropriate model there. The 7.0 m figure is still inside crossover for the higher-gain channels, where the cylindrical model predicts a smaller power density than the far-field equation, so the far-field 7.0 m contour is the conservative one; applying the cylindrical model where it governs gives a summed fraction of 0.96 at 7.0 m and moves the contour in to about 6.8 m, so that value is model-dependent, and only the WCS channel at 9.6 m and marginally the 1900 MHz channel at 6.5 m are affected. Both distances lie inside the 21.03 m minimum slant range the filed study itself evaluates, so both extrapolate its method. Neither contour is a flight rule: 47 CFR 1.1310 sets exposure limits for people and OET-65 evaluates compliance for human exposure 48,49, neither restricts where an uncrewed aircraft may fly, and this review did not survey Title 47 as a whole for one that does. The filed study reports at 6 ft above ground with the antenna vertical pattern taken into account; the calculation here deliberately excludes that pattern, because an aircraft at radiation-centre height directly in front of the antenna is the one geometry it does not attenuate, and it is the geometry a close pass requires. The filed powers are also not a ceiling, since 47 CFR 24.232 permits 1640 W EIRP at an emission bandwidth of 1 MHz or less and 1640 W/MHz above that, up to 300 m HAAT 51, leaving the site's 4.76 kW 1900 MHz channel inside the ceiling for any bandwidth of 2.9 MHz or more; this review did not locate a carrier bandwidth for that channel, and the 5 MHz carrier assumed here would allow 8.2 kW against 2.30 kW at 1.4 MHz.

The consequence for positioning is a dynamic-range problem, and it is the physical form of the state-estimation requirement in hypothesis 4. The GPS L1 C/A signal is specified to arrive at a minimum of -158.5 dBW, which is -128.5 dBm, into a 3 dBi linearly polarised antenna above five degrees elevation 47. An isotropic aperture at 1575.42 MHz has an effective area of 2.88 x 10^-3 m2, about 29 cm2, which at 15.6 m intercepts +12.9 dBm of the site's own transmissions and at 7.0 m, +19.9 dBm. Three idealisations apply and only one runs toward an upper bound: the frequency term, 20 log10 of the ratio, moves by -3.5 dB at 2350 MHz but by +5.0 dB at 881.5 MHz and +6.4 dB at 751 MHz, and the 700 and 850 MHz channels carry 11.4 kW of the 20.8 kW EIRP, 55% of the total, so evaluating each channel's aperture at its own frequency at 0 dBi raises the 15.6 m figure to +16.7 dBm; a 3 dB reference mismatch sits inside the subtraction, since the GPS minimum is defined into a 3 dBi antenna; and the realized gain of an L1-tuned element at cellular frequencies is far below 0 dBi and is the only term carrying the upper-bound label, which this review did not locate any antenna manufacturer publishing. The separation between the ambient field and the wanted signal is therefore of order 140 dB, 141 dB at 15.6 m and 148 dB at 7.0 m. The US DOT adjacent-band assessment measured high-precision receivers, the category RTK belongs to, losing 1 dB of carrier-to-noise density at received interference power as low as about -90 to -95 dBm with a 10 MHz LTE interferer 52, so reaching that from +12.9 dBm requires 103 to 108 dB of rejection ahead of the low-noise amplifier. Those masks were measured between 1475 and 1675 MHz and a GNSS front end rejects 850 and 1900 MHz far harder, so the requirement is overstated by an unknown margin. What the calculation establishes is the scale of the linearity problem near a live sector, and receiver desense is the operative constraint on how close the aircraft can work. vHive's statement that electromagnetic interference around towers disrupts RTK badly enough to force manual flights and repeat site visits 27 is the operational form of the same effect. This review did not locate a published measurement of carrier-to-noise density, RTK fix rate or position error against distance and bearing from an operating antenna, which section 13 identifies as the most useful piece of open work in this domain.

Optical resolution sets what a pass can claim to have looked for, and the coverage rate sets how long it takes. Ground sample distance is pixel pitch times standoff divided by focal length. The Zenmuse P1 has a 35.9 by 24 mm sensor at 4.4 micrometre pitch, and with the 35 mm lens DJI publishes ground sample distance as standoff divided by 80 in centimetres 46; the sensor geometry reproduces that to within 1%, giving standoff divided by 79.5 on the stated pitch and 79.9 on the 35.9 mm width across 8192 columns, so at 10 m that is 1.26 mm per pixel. The nearest published generic anchor this review located for what a visual inspection must resolve is the ASME visual examination requirement, which uses a fine line of 1/32 inch, or 0.8 mm, to demonstrate a procedure and requires remote visual systems to be shown equivalent to direct visual observation 53; the inspection-standard literature was not surveyed exhaustively, so this is the nearest anchor located rather than the only one. Putting that line across a single pixel needs a 6.4 m standoff, and across the three pixels conventionally required for reliable detection, 2.1 m, the three-pixel convention being a machine-vision heuristic rather than a standard. Diffraction bounds the small-pixel sensors these aircraft carry. This review did not locate a published pixel size for any Matrice 4 series camera, so the pitch is derived from the published 1/1.5-inch format, 10.67 mm diagonal on the 16 mm per inch convention, 4:3 giving an 8.53 by 6.40 mm active area, and 48 MP across that width is about 1.06 micrometres at 8064 columns, a derivation graded modelled. At f/2.8 and 550 nm the Airy disc is 3.76 micrometres across and the Rayleigh resolvable separation is its radius, 1.88 micrometres, spanning 1.8 of those pixels 58. The telephoto's nominal 0.268 mm per pixel at 10 m follows from the 36 mm width of the 35 mm reference frame across 8000 columns at the published 168 mm equivalent focal length, and the diagonal-equivalence route through the published format gives 0.258 mm, a 4% difference from the aspect-ratio convention; diffraction degrades that nominal figure to about 0.5 mm. Stopping the P1 to f/8 for depth of field puts its Rayleigh limit at 5.37 micrometres, 1.22 of its 4.4 micrometre pixels, so its 1.26 mm per pixel at 10 m degrades to about 1.53 mm. On coverage, at 80% forward and 60% side overlap, both assumed here, and at the P1's published 0.7 s minimum photo interval, a 15.6 m standoff yields about 19.6 square metres per second of new coverage, a 10 m standoff about 8.0, and a 2.1 m standoff about 0.36 46, which in 30 minutes of capture is roughly 35,000 square metres at 15.6 m, 14,000 at 10 m and 650 at 2.1 m. This review did not locate a published developed surface area for a macro lattice tower, so the result is an area threshold rather than a sortie count: any structure whose developed surface exceeds 650 square metres needs more than one sortie at 2.1 m, and one under 14,000 square metres fits a single battery pair at 10 m. What determines the actual sortie count is the number of distinct aspects each lattice member must be seen from, which this review did not locate in the publications of the service providers, aircraft manufacturers or operators surveyed. ANSI/TIA-222 has admitted unmanned aircraft as a data-collection method for existing structures since Revision I took effect on 1 January 2024 60, and the standard is not an open document, so this review cannot confirm in either direction whether it carries a resolution or standoff requirement.

The energy comparison bears on where the saving from a flight lands. The DOE fuel properties table gives its combined gasoline and E10 row a lower heating value of 112,114 to 116,090 Btu per gallon, 118.3 to 122.5 MJ per gallon, the low end corresponding to E10 54; the mileage conversions use the 116,090 endpoint because the fuel economy series is reported in gasoline gallon equivalents, and that choice moves the per-mile figures by 3.5%. The comparison is kept on a volume basis because this review did not locate a gasoline density in the AFDC table or in the EIA thermal conversion and metric conversion documents checked. One physical gallon of gasoline holds the energy of 62 to 65 full Matrice 350 sorties, or 125 to 129 TB65 packs. FHWA's 2024 Highway Statistics puts light trucks and vans at 18.5 mpg and single-unit delivery trucks at 7.19 mpg in gasoline gallon equivalents 55, which is 6.6 and 17.0 MJ per mile, so a full sortie is the fuel a light service van burns in about 460 m of driving or a single-unit truck in about 180 m, and one mile of driving is 3.5 sorties of stored energy in a light van and 9.0 in a single-unit truck. The vehicle that attends a tower is often a bucket truck, and Utilimarc's utility fleet benchmarking, reported by Government Fleet in September 2015, gives heavy-duty bucket trucks at 5.18 mpg averaged over 2010 to 2014, high 5.47 in 2011 and low 5.0 in 2010, across more than 50 utility and municipal fleets and more than 23,600 vehicles between 33,000 and 55,000 lb GVWR running diesel or alternative fuel 61. Priced on the AFDC diesel lower heating value of 128,488 Btu per gallon, 135.6 MJ per gallon, 5.18 mpg is 26.2 MJ per mile, so one mile of bucket-truck driving is 13.8 sorties of stored energy and one sortie about 117 m of it; assigning diesel to the whole sample is an assumption since the benchmarking mixes diesel and alternative fuel, and the 5.0 to 5.47 mpg spread moves the figure between 13.1 and 14.3 sorties. Vocational vehicles of this class carry no federally published miles-per-gallon rating, being above 8,500 lb GVWR and regulated in gallons per 1,000 ton-mile 56, which establishes the absence of a regulatory rating only. At an assumed 40-mile round trip, a distance this review did not locate a source for, the drive is 265 to 682 MJ for the two light classes and 1047 MJ for a bucket truck, 140 to 360 and 552 times the sortie, every ratio scaling directly with the assumed distance. DOE puts only 12% to 30% of vehicle fuel energy on the road 59, which leaves the assumed drive between about 17 and 166 times the sortie; that figure changes basis, setting vehicle propulsive energy against aircraft stored energy, and since this review did not locate a published drivetrain efficiency for either aircraft it is conservative against the flight. The energy of a site visit sits almost entirely in the vehicle that carries the crew there, so substituting a flight for a climb acts on labour hours and leaves the vehicle term unchanged. The vehicle term falls only when the trip itself is removed, which requires at minimum that no visual observer be posted, and fewer than 763 of the more than 44,000 BVLOS flights in the FAA's BEYOND programme through October 2024, about 2%, were flown without one 23. Two further quantities were searched for and not located: the charging efficiency of field battery stations, and the average distance driven per telecom truck roll. NREL's Fleet DNA carries a bucket-truck vocation of 20 vehicles and 283 operating days from utility and telecommunications fleets and is the most likely public source for a measured duty cycle 62; nrel.gov was not reachable during compilation, so no figure was taken from it.

Read against the hypotheses, these quantities state the standoff conflict in numbers, and that tension is the finding for hypothesis 1. The composite general population reference level sits at 15.6 m from the sector and the composite occupational level at 7.0 m 42,48,49,50, while resolving a 0.8 mm feature across three pixels with a Zenmuse P1 and the 35 mm lens requires 2.1 m 42,46,53. Exposure fields push standoff out and defect resolution pulls it in, coverage per battery pair falls by a factor of 22 between the 10 m and 2.1 m standoffs, and receiver desense rises as the aircraft closes. At 10 m and beyond the imagery cannot resolve a 0.8 mm feature; at 2.1 m the aircraft is inside the occupational contour and inside the crossover radius of every channel at this site. Three measured changes move that boundary and they are separable. The first is optical: putting a 0.8 mm feature across three pixels at a 10 m standoff requires 0.267 mm per pixel, which the 168 mm equivalent telephoto already samples at 0.268 mm per pixel, and requires the optics to resolve at that scale as well, so under the Rayleigh criterion at 550 nm it needs an entrance pupil of about 25 mm against the roughly 15 mm that telephoto has at f/2.8 on a 1/1.5-inch format, a factor of 1.7 in aperture diameter 42,45,58. That is an engineering implementation threshold rather than an algorithmic one. The second is procedural: a scheduled sector power reduction or downtilt during a pass moves the contour directly, and because that contour is a human-exposure compliance boundary rather than a flight restriction the change sits under the operator's own control; this review did not locate a published operator procedure of that kind. The third is state estimation: the constraint at a live site is holding position under partial GNSS denial, which points to onboard visual-inertial estimation referenced to the structure itself, and which section 11 develops as a requirement on hypothesis 4. None of the three is a question of stored energy, which is negligible against the vehicle term at 1.90 MJ for a full sortie against 26.2 MJ for a single mile of bucket-truck driving 42,44,54,61. Battery endurance is a fourth term and it does bind: the same 30 minutes of usable hover covers about 14,000 square metres at a 10 m standoff and 650 at 2.1 m, a factor of 22, so any structure whose developed surface exceeds 650 square metres needs more than one sortie at close standoff, and the sortie count is what sets the price per tower 42,44,46.

11. What does an actionable plant model require?

The fellow's project statement identifies the scientific problem as giving embodied systems a reliable model of the physical plant and the judgment to act on it when conditions get messy 43. That is hypothesis 4, and the evidence assembled here decomposes it into six requirements, each partly met in the deployed record and each measurable.

Ground truth. The Bentley validation compared the twin against best-existing data, meaning the RF data sheets, structural analyses and CAD drawings on file 28, and the twin agreed with the records 95% of the time on equipment detection, a figure drawn from four towers and carried here, as in section 8, as the vendor's characterisation rather than as a measurement 28,42. What the study cannot say is whether the twin or the record was right in the 5% where they disagreed, and the one case it details, a RAD centre documented at 180 ft and measured at 175 ft, is presented without saying which was subsequently confirmed. A validation whose reference is the registry measures agreement with the registry, so it does not resolve registry error, which is the quantity a twin is commissioned to reduce. The measurement that resolves it is a climb-verified or total-station-verified sample of thirty or more structures, stratified by tower type and age, with per-parameter agreement reported and disagreements adjudicated by physical measurement. This review did not locate such a study in public.

Navigation at a radiating structure. vHive states that electromagnetic interference around telecom towers disrupts RTK positioning badly enough to force manual flights and repeat site visits, and announced in November 2025 that it had restored autonomous capture on the DJI Matrice 4E in high-interference zones with a roughly five-minute automated 360-degree survey 27, with no positioning-error figures or degradation curve accompanying the claim. The physical situation is quantified in section 10: the wanted GNSS signal arrives at -128.5 dBm 47 while an idealised L1 aperture at the general population contour intercepts about +12.9 dBm of the site's own transmissions, a separation of order 140 dB implying 103 to 108 dB of front-end rejection to hold a high-precision receiver within 1 dB of carrier-to-noise density 42,47,50,52. The binding constraint at a cell site is therefore actual engineering implementation, in the specific form of state estimation under partial GNSS denial rather than perception, which points toward onboard visual-inertial estimation with the structure itself as the reference frame. That work belongs on the device, because the backhaul at a site is least reliable during exactly the storm conditions when the inspection is most needed.

Accuracy budgets tied to tasks. A twin quoted as 95% accurate with tilt and azimuth within a 5% error margin 28 does not resolve to a usable tolerance, since five percent of a 360-degree bearing is roughly 18 degrees, which may clear a structural-engineering tolerance and be far outside an RF sector planning tolerance 42. The usable unit is a per-parameter accuracy budget stated against the tolerance of the task that consumes it: mount height for structural loading, azimuth and downtilt for RF, cable and jumper routing for maintenance planning, ground-lead continuity for safety. A single whole-model accuracy figure does not resolve to any of those tolerances. Accuracy per parameter, and the reference standard against which it was measured, are absent from every disclosure this review examined.

Acting when the model is known to be wrong. The DIRT dataset is the canonical case: 40,400 damages a year where the buried-asset model was wrong or unfindable, and a 38% chance across eight 811 centres of being unable to start on time because locate responses were incomplete 31,42. A system that plans work from a registry inherits every error in it. The Charlot Lab position, applied here, is that each attribute in the plant model should carry a value, an uncertainty, a provenance and an age, and that every autonomous work order should carry an explicit predicate over those attributes that must hold for the plan to remain valid. When the predicate fails in the field the correct output is abstention and escalation rather than a best guess, which makes refusal a first-class measurable output and supplies a second metric alongside accuracy. A vendor reporting 95% accuracy on the 60% of sites its system found tractable is not comparable to one reporting 90% across all of them, and no disclosure this review examined reports the denominator.

Contact. Everything deployed at scale in telecom is non-contact sensing: drones photograph, twins measure, dispatch systems decide. This review located no fielded system that torques a bolt, dresses a connector or tests grounding continuity, and peer-reviewed work on robots that physically service towers remains at feasibility stage 41. Built Robotics' trajectory is the useful signal, its volume business being solar pile driving rather than telecom conduit 37, because open repetitive greenfield geometry is what current autonomy handles and congested urban civils with surface-restoration constraints is a harder case. NATE's framing of the near-term role, removing two or three climbs from a day of six to ten 13, matches what the deployed record supports over the next several years.

Currency. The mechanism carrying the most value in the deployed record is a process commitment: Verizon's automatic capture and digitisation of every new build and modification from early 2025 25, with the drone as the instrument that makes it affordable. A plant model decays continuously, so a one-time capture produces an artefact that is wrong again within a maintenance cycle. Bentley's cost model explicitly excludes the cost of maintaining currency and of re-capture after each site modification 29, so the published economics omit the line item that determines whether the model stays useful. Recapture after each modification is a recurring cost and the initial sweep is a one-time cost; the published economics include only the second.

12. Where does the value sit?

Ranking by the size of the addressable pool, the quality of the evidence and what is technically and legally available now produces an ordering whose first rank, avoided and better-targeted dispatch, requires no embodied hardware, and whose second rank splits into a planning layer available now and a machine layer still in trial. Structural tower inspection ranks seventh on present addressable value, and the two constraints holding it there are quantified in sections 9 and 10.

The table below sets out the ranking, the size of each addressable pool, the verification grade of the figures in it and the evidence behind it. Sizes are drawn from the cited sources and, where marked, from the author's arithmetic on them. The final column gives position on the trajectory rather than a buy or hold judgment.

Table 4. Physical AI and logistics opportunities in telecom, ranked by evidenced value and current availability, with the verification grade of every figure in the size column.

13. Which questions are open, and what would settle them?

The following are the questions this review could not close. Each is stated with what would settle it, because several are answerable from data an operator already holds, and each names the hypothesis it bears on.

Field-operations labour as a distinct line, bearing on hypotheses 5 and 6. MTN reports $258.1bn of telecom labour across 4.36 million employees 19 and Omdia about $300bn 20, and this review did not locate a source isolating field crews from retail, care, IT and back office. The figure is derivable from payroll and cost-centre data held internally by any operator. This review did not locate it in the public record, which is why no independent estimate of the automatable share of telecom field labour is possible here.

The real cost of a truck roll, bearing on hypothesis 5. The $150-500 and $1,000 figures 14 have no traceable methodology and reach telecom through vendors selling avoidance. What is needed is a fully loaded figure per fault class, since a fibre splice, a small cell reboot and a residential install are different events with different costs.

Twin accuracy against physical truth, bearing on hypotheses 2 and 4. This review located no study comparing an asset digital twin against climb-verified measurement rather than against existing records 28. Until one is located, twin accuracy claims are claims about agreement with the artefact whose reliability is the reason the twin was commissioned.

The guyed to self-supporting mix of the US tower estate, bearing on hypothesis 1. TIA-222 recommends three-year inspection for guyed masts and five-year for self-supporting structures 7, and national inspection volume varies by two-thirds depending on the mix, which this review did not locate published 42. That single ratio sets the width of the $21-36M band.

Whether telecom field labour is scarce, bearing on hypothesis 6. Continuum Capital projects a shortfall of 58,000 workers now and 119,000 over ten years 2; WIA reports the wireless workforce falling 7.2% in a single year 4; Wireless Estimator reports hiring at a twenty-year low 5; BLS projects the technician occupation declining 3% 3; CWA calls the shortage narrative misinformation 2. These cannot all describe the same labour market, and this review located no neutral source that adjudicates.

The tower fatality rate, bearing on hypothesis 1. This review did not locate a BLS occupational code isolating tower technicians, and published rates differ by 1.8x according to which denominator and period is chosen 12. The safety argument for removing people from towers cannot be quantified on the sources compiled here until a neutral series is located.

Cost of BVLOS compliance under Part 108, bearing on hypothesis 3. The final rule's detect-and-avoid and operator-certification requirements are not in the public docket 24, so the cost side of the crewless-inspection case is the largest unquantified term in it.

Positioning degradation near radiating structures, bearing on hypotheses 1, 3 and 4. vHive's claim that EMI disrupts RTK enough to force manual flights 27 comes with no error figures and no degradation curve. Section 10 bounds the problem from first principles at a separation of order 140 dB and an implied 103 to 108 dB of front-end rejection 42,47,50,52. The measurement is directly obtainable with a survey-grade receiver and a spectrum analyser at a live site, and it is the piece of open work this evidence base most clearly calls for.

Field inventory and spares economics, bearing on hypothesis 10. This is one of the four capabilities in the fellowship's own project statement 43 and this review located no public figure for telecom field inventory carrying cost, stockout rate or contribution to failed first visits. TSIA's 87% first-visit resolution 15 does not decompose into diagnosis, access and parts failure, and that decomposition is where the inventory case lives.

Whether any of it shows in the aggregates, bearing on every hypothesis. Omdia notes that disclosures of quantifiable AI savings remain scarce 20, Omdia's global telecom labour opex line fell 2.2% in 2024 partly through divestment rather than efficiency 20, and Dell'Oro revised its own capital-intensity forecast by about a point within twelve months 18. Any automation effect at current scale is smaller than the spread between the houses measuring the aggregate, which makes instrumented matched-cohort measurement the only route to detecting it.

The developed surface area of a macro lattice tower, bearing on hypothesis 1. Section 10 states the coverage result as an area threshold rather than a sortie count because this review did not locate a published figure 42,46. That single number converts the threshold into a sortie count and therefore into a price.

14. Risk register

Regulatory schedule risk on BVLOS, high likelihood and high impact. The crewless-inspection case depends on Part 108, which reached OIRA on 10 July 2026 with publication expected late 2026 or early 2027 and a six to twelve month transition, on no FAA commitment date 24. A twelve-month slip pushes every associated payback out by the same amount. The compliance cost is unquantified because the detect-and-avoid requirements are not in the public docket.

Labour-market reversal, moderate likelihood and high impact on the business case. Automation displaces expensive scarce labour more readily than cheap available labour. WIA reports a 7.2% workforce decline 4 and Wireless Estimator reports tower construction hiring at a twenty-year low with jobseeker registrations up more than 300% 5. If that slack persists, the wage pressure that funds automation weakens even as the technology matures, which is the downside branch on hypothesis 6.

Vendor-sourced economics, high likelihood and moderate impact. The headline savings in this field are self-published or modelled. Bentley's $3,500 per tower is a model with three unreconciled percentages 29,42; the $350-600 versus $800-1,500 inspection comparison is not scope-matched 8; Comcast's tens of millions has no baseline, period or definition of an avoided roll 17,34; Vodafone's numbers were published by its suppliers 16,32. Every figure carried forward from those four sources carries its grade with it, and none has been reproduced against an operator-audited baseline located by this review.

Registry error compounding into autonomous action, moderate likelihood and high impact. A system that plans work from records containing a five-foot RAD centre error 28 or that digs where 20.5% of reported damages already trace to a facility unmarked or marked in the wrong place 31,42 will act confidently on bad inputs. This is the failure mode that turns an efficiency programme into an outage, and the mitigation is the uncertainty-and-abstention discipline in section 11 rather than a larger model.

Scope substitution in inspection, high likelihood and moderate impact. A drone pass and a climbing crew do different work. Torque checks, connector servicing, grounding continuity and remediation have no uncrewed equivalent in the fielded record 8,41. Programmes that book the price difference as a saving without booking the residual crew visits will not realise the number.

Standoff and resolution mismatch, moderate likelihood and moderate impact. Section 10 shows that an inspection pass flown at a standoff comfortable for exposure and coverage cannot resolve a 0.8 mm feature, and that a pass flown close enough to resolve it sits inside the occupational contour and inside the GNSS desense regime 42,46,48,52,53. A programme that specifies neither standoff nor per-parameter resolution in its acceptance criteria will accept imagery that does not support the finding it is used for.

Single-vendor dependency in the asset-model layer, moderate likelihood and moderate impact. vHive supplies the twin software behind at least two of the largest operator programmes, including Verizon and Deutsche Telekom 26, and Consilience Analytics has processed 35,000 twins into the Bentley stack 30. The plant model is becoming a system of record, and the quantity that would size the concentration is the share of the US tower estate whose twin is produced by a single supplier, which this review did not locate published. That is hypothesis 2 read as a risk.

Model-currency debt, high likelihood and moderate impact. Published twin economics exclude the cost of recapture and currency maintenance 29. A programme funded on the initial sweep and not on capture-on-change produces an asset that decays at the rate the estate is modified, and that modification rate, sites touched per year as a share of sites held, is the quantity to measure, because it sets the refresh cadence the funding has to cover.

Capital tightening, high likelihood and moderate impact. Global capex was flat in 2025 and is forecast to fall 2% in 2026 18, and MTN records the lowest twelve-month capex total since 2011 19. Verizon is guiding down while AT&T guides up 22. Programmes that need multi-year capital commitments compete in a tightening envelope, and which side of the capital-and-operating line a programme is funded from therefore bears on whether it proceeds.

Measurement invisibility, moderate likelihood and moderate impact. If a programme's effect is smaller than the 16% difference between Omdia and MTN on telecom labour cost 19,20,42, aggregate financials will not show it, and instrumented matched-cohort measurement has to be designed in from the start.

15. Scenarios to 2032

Scenario A, decision support carries the value. Part 108 publishes in late 2026 or early 2027 24 and detect-and-avoid and certification costs keep per-site crewless economics marginal outside dense clusters of towers. Asset twins become standard practice through capture-on-change policies of the Verizon type 25. Embodied field systems in the US stay at limited scale. Savings appear as dispatch avoidance and first-visit resolution, in the range the operator-attributed evidence already supports: 10% fewer dispatches 16, 28% fewer repeat visits 32, a few points of technician utilisation 15. This is the base case on current evidence and it corresponds to hypothesis 5 holding and hypothesis 1 advancing slowly. Discriminating indicator: no US carrier publishes an audited before-and-after on robotic field work by the end of 2028.

Scenario B, the twin becomes the system of record. Operators and towercos mandate capture on every modification, and the twin displaces the RF data sheet and the structural drawing as the authority for what is on a structure. Value accrues through reduced rework, faster co-location approval and fewer failed site visits rather than through removed climbs. This is hypothesis 2 resolving upward, and it is the scenario in which the fellowship's work has the largest effect, because it makes the plant model a contractual object with accuracy obligations attached. Discriminating indicators: a carrier or towerco publishes per-parameter accuracy against a climb-verified baseline; lease and co-location contracts begin referencing the twin rather than the drawing; a vendor publishes an abstention rate alongside an accuracy figure.

Scenario C, robotics becomes a carrier revenue line rather than a cost line. The Chinese operator model generalises: China Telecom's Robotics-as-a-Service with AgiBot, China Unicom's hazard inspection deployments and State Grid's 8,500-robot programme 40 establish connectivity-plus-robotics as a sellable bundle, and Western carriers follow the pattern already visible in AT&T's Ghost Robotics work, which is a FirstNet product sold to public safety and defence rather than an internal tool 38. This is hypothesis 9 resolving upward, and under it the operative question for an operator is what its network can charge other people's robots. Discriminating indicators: a US or European carrier reports a robotics services P&L line; FirstNet-style bundles are sold to utilities and industrial customers at volume.

A downside branch runs across all three. If the labour slack visible in WIA's 7.2% workforce decline 4 and Wireless Estimator's twenty-year hiring low 5 persists through 2028, the wage pressure funding automation weakens, and programmes justified on scarcity get repriced against cheap available crews. The indicators to watch are WIA's annual FTE figure and the postings-to-registrations ratio at 5, both published yearly and neither requiring access to proprietary research.

Across every scenario, one structural fact holds. WIA's own numbers put recurring operations at 5.4 times expansion capital in US wireless 4,42, and Dell'Oro has capex flat to slightly declining through 2030 18. The estate keeps growing while the capital to build it does not, which pushes the problem onto operations regardless of which scenario obtains.

16. Hypothesis resolution

Each hypothesis resolves to three things: where it stands today, stated in degrees rather than as a verdict; the constraint currently binding it, named from material science, actual engineering implementation, energy, computation efficiency and AI algorithms, or named as regulatory or workforce where that is what binds, with one row recording that no class is yet identifiable and naming the measurement that would identify one; and the measured change that would move the boundary, stated as a threshold, a factor, a date or a rule. The constraint named is the one governing the next increment of capability rather than the only one in play, and where a second is close behind the table says so.

Three patterns run across the table. First, energy per sortie is negligible against the vehicle term: a full two-battery sortie stores 1.90 MJ against 26.2 MJ for a single mile of bucket-truck driving, and at an assumed 40-mile round trip the drive is 140 to 552 times the sortie on a stored-energy basis and 17 to 166 times on the mixed propulsive basis 42,44,54,55,59,61. Battery endurance is a separate term and it does bind: the same 30 minutes of usable hover covers about 14,000 square metres at a 10 m standoff and 650 at 2.1 m, a factor of 22 42,44,46, which is why endurance sits in hypothesis 1's constraint cell beside optics and regulation. Second, the constraint most often binding is actual engineering implementation. Third, the two hypotheses furthest along, 5 and 2, are both about the record rather than the machine, and both advance on calibrated reporting of what the system does not know, which for hypothesis 2 is per-parameter accuracy against a physical reference and for hypothesis 5 is precision and recall at the chosen operating point.

Table 5. Resolution of the ten hypotheses: position on the trajectory, binding constraint, the measured change that moves it, and the verification grade of the figures carried in the position cell. Constraint classes: material science, actual engineering implementation, energy, computation efficiency, AI algorithms, regulatory, workforce.

17. Conclusions, and the measurements that would close the largest gaps

The economic opening this work is aimed at is real and the sources agree on its shape. Labour dominates buildout cost 1, recurring operations dominate wireless spend by more than five to one 4,42, costs are rising and 88% of respondents expect them to keep rising 1, and the plant model carries measurable error at every scale from the national site count 4,6 down to a RAD centre height 28. Ten hypotheses were tested against that background and each resolves to a position on the trajectory rather than to a verdict.

Two applications are already carrying load at national scale, and both are software running on an asset record. Dispatch prediction, remote visual assistance and technician decision support are documented in three programmes at two large operators 16,17,32,34, and the field-service benchmarks show 9.2% more productive hours available from closing a utilisation gap with no new technology at all 15,42. The constraint on the next increment there is AI algorithms, in the specific form of decision calibration under class imbalance, since accuracy without a precision and recall split does not fix an operating point for an imbalanced dispatch decision.

Aerial structural inspection is deployed at carrier scale and, on industry characterisations of the paywalled ANSI/TIA-222 Revision I that this review did not read directly, admitted by the tower standard since January 2024 9,60, and its present national addressable value at vendor-quoted prices is $21-36M a year 42. Two constraints hold that figure where it is and both are quantified. The regulatory one is Part 108: under 2% of BVLOS flights in the FAA's flagship BEYOND programme were flown without a visual observer through October 2024 23, and the rule reached OIRA on 10 July 2026 with publication expected late 2026 or early 2027 24. The physical one is the standoff conflict computed in section 10: the composite general-population exposure reference level sits at 15.6 m from a filed macro sector and the occupational level at 7.0 m, while putting a 0.8 mm feature across three pixels needs 2.1 m, with coverage per battery pair falling by a factor of 22 between a 10 m and a 2.1 m standoff 42,46,48,49,50,53. Neither contour is a flight rule, since 47 CFR 1.1310 and OET-65 govern human exposure only 48,49, so the 15.6 m figure enters as an operator-policy standoff derived from a human-exposure compliance boundary, which makes a scheduled sector power reduction during a pass a procedural change under the operator's own control rather than a regulatory negotiation. The optical change that relaxes the resolution term is an entrance pupil of about 25 mm against the roughly 15 mm a 168 mm equivalent telephoto has at f/2.8, a factor of 1.7 42,45,58. Energy per sortie is negligible against the vehicle term, 1.90 MJ stored against 26.2 MJ for a single mile of bucket-truck driving 42,44,54,61, and battery endurance is the separate term inside the same figure: the 30 minutes of usable hover that covers about 14,000 square metres at 10 m covers 650 at 2.1 m, and the developed surface area of a macro lattice tower, which this review did not locate published, converts that area threshold into a sortie count and therefore into a price 42,44,46.

Every application in this report consumes the same asset record, which is why hypothesis 4 sits underneath the other nine. A plant model good enough to plan physical work against needs accuracy stated per parameter against the tolerance of the task consuming it, and an abstention rate stated beside it so that a system's silence is as reportable as its answers. Two metrics are absent from every vendor disclosure this review examined: accuracy stated per parameter against a named reference standard, and abstention rate, meaning the fraction of sites or assets on which the system declined to produce a measurement, with accuracy reported on the remainder.

Six measurements would close the largest evidence gaps identified here. Five are producible from data an operator already holds or from a field study of modest scale; the sixth waits on publication of the final rule. Each moves a named hypothesis. First, the field-operations labour line, derivable from payroll and cost-centre data and not located by this review in any public source 19,20; hypotheses 5 and 6 are both denominated in it. Second, a truck-roll cost per fault class rather than a single average, since the published $150-1,000 range 14 averages across events with little in common; decomposing failed first visits into diagnosis, access and parts failure is the same exercise and answers hypothesis 10 15,43. Third, a climb-verified twin accuracy audit on a stratified sample of at least thirty structures, reporting agreement per parameter and adjudicating disagreements by physical measurement; this review located no such study 28, it is inexpensive relative to a twin programme, and it settles hypotheses 2 and 4 together.

Fourth, the five-year cost of keeping a twin current, since capture on every new build and modification is the mechanism that keeps a model true 25 and is the line item the published economics omit 29. Fifth, a measured positioning-degradation curve near a live sector: section 10 bounds the problem at a separation of order 140 dB and an implied 103 to 108 dB of front-end rejection 42,47,50,52, and vHive's operational claim carries no error figures at all 27. A survey-grade receiver, a spectrum analyser and access to live and dark sites would produce a degradation curve against transmit power, frequency, distance and bearing; it sits within the Charlot Lab's on-device state-estimation remit and bears on hypotheses 1, 3 and 4. Sixth, the compliance cost of Part 108 once the final rule is readable, since its detect-and-avoid and operator-certification requirements are the largest unquantified term in the crewless-inspection case 24; the waiver route has cleared roughly 134 organisations in five years 23, which is the throughput the interim configuration supports.

The graded record shows a field partway along several distinct trajectories. The software-on-a-record applications are at national scale. The asset twin is close behind and advancing on process commitment rather than on invention. Aerial inspection is deployed and admitted by the standard, held at its present value by one regulatory date and one aperture factor. Contact work on live plant is at feasibility in the peer-reviewed record. Each of those positions is measurable, each has a named constraint, and each has a threshold attached, and the six measurements above are what would move them.

References

1. 1 Fiber Broadband Association / Cartesian, "Fiber Deployment Cost Annual Report 2025", 3rd edition, January 2026. Operator and contractor self-reports, September-October 2025, projects in 38 states; each year's sample independent, so year-on-year comparison is not like-for-like; sample size not disclosed in the executive summary. reported
2. 2 Continuum Capital broadband workforce study for the Fiber Broadband Association and the Power & Communication Contractors Association, Fiber Connect, July 2024. A forecast commissioned by trade bodies whose members benefit from workforce funding; predates the BEAD restructuring and the 2025 wireless construction contraction. The Communications Workers of America publicly disputes US telecom workforce shortage claims. modelled
3. 3 US Bureau of Labor Statistics, Occupational Outlook Handbook, telecommunications equipment installers and repairers except line installers, 2024-34 projections. A model rather than a measurement; bls.gov blocked automated retrieval, so these figures come from search-index extracts rather than a direct read. Excludes line installers, who are projected to grow. modelled
4. 4 Wireless Infrastructure Association, "By the Numbers" 2025, research by iGR, 18 March 2026. Association-commissioned and modelled by iGR from operator and contractor inputs; this review did not locate a published methodology or sample; figures are full-time equivalents rather than headcount. modelled (iGR estimate for WIA)
5. 5 Wireless Estimator, tower construction employment against its own twenty-year record, 5 February 2026. Proprietary single-source data from one industry job board, which is also the publisher's commercial base; postings volume is a noisy proxy for demand. reported
6. 6 CTIA, Annual Wireless Industry Survey, 2025 edition, reporting year-end 2024. A survey of CTIA's own members, operator-supplied and unaudited; CTIA's site definition differs from WIA's and this review did not locate published definitions from either body that reconcile the two counts. reported (member-supplied, trade association)
7. 7 ANSI/TIA-222 (Telecommunications Industry Association), TR-14 committee FAQ on structural inspection intervals. TIA states explicitly that the interval is a recommendation rather than a mandate and that owners may lengthen or shorten it per structure. verified
8. 8 Drone tower inspection service providers (AeroDeploy, Drone Harmony, DroniTech, Voliro, ABJ Drones), 2025-2026 marketing material. Every figure traces to a company selling the service; this review located no independent, operator-audited or peer-reviewed comparison; the drone and climbing-crew comparison is not scope-matched. self-published
9. 9 AT&T drone tower inspection programme, via Computerworld and RCR Wireless, 2016-2017. Nine to ten years old; this review did not locate an updated AT&T disclosure; an AT&T official declined to state the cost. The 65,000 figure counts cell sites where AT&T had equipment at the time and predates subsequent tower sales. self-published
10. 10 American Tower and Crown Castle communications on drone-derived asset digital twins, 2025. American Tower's industry-leader rating comes from its own customer survey with no published methodology, sample or comparison set; this review did not locate structures scanned, refresh cadence or model accuracy from either company. self-published
11. 11 US Occupational Safety and Health Administration, Communication Towers topic page: 13 fatalities in 2013, 12 in 2014, 3 in 2015, 6 in 2016. Not updated past 2016 as of August 2026; counts only, with no denominator published alongside them, and this review did not locate a current official federal series for this work. verified
12. 12 Communications Workers of America survey (103 respondents) against the Wireless Estimator fatality database and cited safety experts, 2023. Both parties are interested; the database is proprietary and internally inconsistent on its date range; this review did not locate a BLS occupational code isolating tower technicians, so no neutral rate is available from these sources. reported
13. 13 National Association of Tower Erectors, via Commercial UAV News, on climb frequency and fatigue. A qualitative characterisation; this review located no underlying measurement study and no data linking climb count to incident probability. reported
14. 14 Technology & Services Industry Association truck-roll cost figure as cited across field-service vendor literature (PTC, SightCall, Blitzz, VSight, Smarty), 2024-2026. No primary TSIA publication stating the $1,000 figure could be located; every citation runs through a vendor selling truck-roll avoidance; the $150-500 band has no traceable methodology and mixes residential installs with enterprise service. reported
15. 15 Technology & Services Industry Association, field services benchmarking and "The State of Field Services 2025". Member-supplied data behind a membership gate, drawn from technology services generally rather than telecom field operations specifically; comparison companies self-select into the benchmark. reported
16. 16 Vodafone remote visual assistance results, published by the vendor TechSee, undated case study. Published by the supplier rather than by Vodafone, with no date, market, baseline or measurement period stated. self-published
17. 17 Comcast dispatch-prediction machine learning system, via StreamTV Insider and Rethink Research. Comcast's own claim, with no baseline dispatch volume, no time period and no definition of an avoided roll. self-published
18. 18 Dell'Oro Group worldwide telecom capex forecast, 2026, covering about fifty service providers representing roughly 80% of global capex. Dell'Oro's own September 2025 release gave capex/revenue approaching 15% by 2029 and wireless intensity of 12-13%, so the house revised itself materially within a year; this review did not locate an absolute dollar total. reported
19. 19 MTN Consulting, Telco Market Tracker 2Q25, December 2025, covering 140 operators. Varying disclosure quality across the sample; the labour figure aggregates all employees rather than field crews; figures are annualised trailing rather than calendar-year actuals. reported
20. 20 Omdia, Global Telecoms Opex Tracker. Omdia's roughly $300bn labour figure and MTN's $258.1bn describe the same thing on different samples and definitions; Omdia attributes part of the 2024 decline to divestments rather than efficiency and notes that disclosures of quantifiable AI savings remain scarce. reported
21. 21 Deloitte, 2026 Telecommunications Industry Outlook, citing IDC for global telecom revenue. Deloitte's capex intensity of about 18% is four points above Dell'Oro's projection and its revenue base differs from Omdia's and MTN's; the outlook contains no field-operations or maintenance figures. reported
22. 22 AT&T, Verizon and T-Mobile capital investment guidance, 2025-2026. Guidance rather than outturn; definitions of capital investment differ between the three carriers. self-published
23. 23 US Department of Transportation, Office of Inspector General, "FAA Has Made Progress in Advancing BVLOS Drone Operations but Can Do More To Achieve Program Goals and Improve Data Analysis", Report AV2025034, 30 June 2025. As of October 2024 BEYOND participants had flown over 66,000 flights including over 44,000 BVLOS flights: "of the over 44,000 BVLOS flights in BEYOND, operators flew less than 763 (approximately 2 percent) of these flights without a visual observer. In the PSP/IPA programs, no operators have conducted flights without a visual observer." Covers the partnership programmes only, so commercial waiver holders outside them are not counted; the OIG also found FAA's data validation vulnerable to errors and mis-categorised at least one operator's observer status. verified
24. 24 FAA Part 108 BVLOS rulemaking and OIRA docket status, via trade press (Airdata, UAV HQ, Pilot Institute, DLA Piper), 2026. No FAA commitment date exists in any of the trade sources; OIRA timelines slip routinely; the final rule's detect-and-avoid and operator-certification requirements are not in the public docket at the time of writing. reported
25. 25 Verizon asset digital twin programme (David Wolff, VP Central Engineering; Candice DeMonteiro, Associate Director), via Inside Towers, 2025. Site count given as a range rather than a figure; this review did not locate an accuracy metric, cost per twin or refresh cadence for the legacy estate; the 15-minute disaster-assessment claim is not independently verified. self-published
26. 26 vHive Series B funding and twin production volume, via VentureBeat and company materials, 2025. Counts of twins generated conflate first captures, re-captures and non-telecom verticals; this review did not locate a breakdown by industry or customer. self-published
27. 27 vHive announcement on electromagnetic interference, RTK degradation and autonomous capture on the DJI Matrice 4E, 25 November 2025, via DroneLife. No positioning-error figures, no degradation curve against transmit power or frequency, no mechanism description and no third-party validation; the sole-provider claim is not independently checkable. self-published
28. 28 Bentley Systems / Consilience Analytics validation whitepaper on OpenTower iQ, by Apurba Tribedi (Bentley), 2024. Vendor-sponsored; four towers drawn from a group of fifty; the comparison baseline was best-existing data (RF data sheets, structural analyses and CAD drawings from under a year to several years old) rather than climb-verified ground truth; the paper does not state whether the 5% margin is of reading or of full scale. self-published
29. 29 Bentley Systems operations and maintenance cost model in the same 2024 whitepaper: about $5,700 per tower manually against $2,200 with OpenTower iQ, plus 50% reporting time savings, 15% annual opex reduction and 42% lower ten-year TCO. Not a measurement; this review did not locate a matched-cohort before-and-after published by any operator; the model excludes twin currency maintenance, re-capture after site modification, and pilot and mobilisation cost. modelled
30. 30 Consilience Analytics twin production volume via the Bentley whitepaper, 2024: over 35,000 digital twins processed since 2019. Cumulative processing includes repeat captures of the same asset, so it is an upper bound on unique sites and cannot be read as estate coverage. self-published
31. 31 Common Ground Alliance, 2024 DIRT Report, 28 August 2025, analysing 196,977 unique damage reports. A voluntary dataset CGA itself treats as an undercount, with a reporting population that changes year to year; root cause is assigned by the reporting party, who has an interest in the assignment; telecom's share is inflated by telecom construction being a major cause of damage to other utilities; the five root causes quoted here sum to 67.9% and are a subset of the top ten to which CGA attributes about 85% of reported damages, so the two figures are consistent; this review did not locate the remaining five causes broken out individually. reported
32. 32 Vodafone and Celfocus, Field Technician Assist on Google Cloud, 2025. Operator's own figure with no baseline period, market or fault mix disclosed and no independent audit; "repeated visits" is not defined in public materials. self-published
33. 33 Comcast alarm-clustering and tooling dispatch system for mass outages, trialled in the 2024 hurricane season and nationwide from September 2025. "Effectiveness" is not defined and the 50% claim has no stated denominator; this review located no third-party verification and no per-storm breakdown. self-published
34. 34 Comcast dispatch-prediction accuracy and savings, via Rethink Research and an H2O.ai case study. Accuracy without a precision and recall split does not fix the operating point for an imbalanced dispatch decision; the savings figure is unaudited and the timeframe is not stated. self-published
35. 35 Openreach / BT Group MicroBORE trenchless robotic drilling with Synthotech and the University of Surrey, November 2025, via ISPreview, Telecompaper and IT Brief. This review did not locate a published bore rate, cost per drop, failure rate, trial volume or comparison against manual moling; Openreach's 15bn pound programme to 25 million premises by December 2026 is being met without it. self-published (operator, via trade press)
36. 36 Deutsche Telekom deployment cost per metre (about 85 euro traditional underground, 65 euro reduced-depth, 10 euro overground), via teltarif.de. Operator's own figures with no methodology published alongside them; German civils costs are dominated by local reinstatement and permitting rules that do not transfer to the US or UK. self-published
37. 37 Built Robotics Exosystem excavator retrofit, 2025, and the Blattner nationwide solar piling agreement reported by ENR, September 2025. The 20-30% saving is a vendor-reported contractor figure with no named contractor or measured baseline; this review did not locate a publicly documented telecom fibre-build deployment of the trenching product. self-published
38. 38 AT&T and Ghost Robotics quadrupedal ground vehicles over FirstNet using AT&T Geocast, 2023 onward, via Light Reading and Urgent Communications. This review did not locate a fleet size, deployment count or telecom-facility use case; announcements date from 2023 with limited public updates since. self-published
39. 39 NTT DATA data centre facility inspection robots across Japanese sites from April 2023, via DataCenterDynamics and Digital Infrastructure Lab. 2023 vintage; this review did not locate a current fleet count, a measured labour saving or evidence of extension to central offices or exchanges, and the programme is not independently verified. self-published
40. 40 GSMA MWC Shanghai 2026 coverage of China Telecom's Robotics-as-a-Service with AgiBot and China Unicom's hazard-inspection deployments; State Grid Corporation of China robot programme announcements, 2025-2026. Announced figures from state-linked entities, and this review did not locate independent verification of units in service, uptime or task success; targets of this kind routinely diverge from delivery. self-published (operator, via event coverage)
41. 41 Shah and Dave, peer-reviewed review of robotic systems for communication tower inspection and maintenance, cited only for the qualitative finding that such systems remain at feasibility stage. The full citation was truncated in the compiled evidence record and must be completed before publication; no quantitative claim rests on it. reported (literature, citation incomplete)
42. 42 Author's calculations. Every figure carrying this index is computed from the cited primaries rather than measured, and is presented as modelled. Economic derivations: per-mile conversion of per-foot fiber costs at 5,280 ft/mi; labour share against total median cost and against labour plus materials; the residual between the sum of medians and the median total; per-mile value of the internal-versus-contractor labour delta and of a 10% labour productivity gain; the trenching-to-plowing saving implied by a 60% method spread; opex-to-capex ratio and per-site opex on WIA denominators; TIA-222 implied inspection volume across 158,500 towers at three- and five-year intervals; national value of drone structural inspection at quoted midpoints; extrapolation of Bentley's per-tower saving across the US tower estate; DIRT locate-error damage counts and root-cause sum; labour cost per employee from MTN aggregates; the Omdia-MTN labour gap; TSIA repeat-visit and utilisation headroom arithmetic; and decomposition of the tower fatality rate dispute into numerator and denominator contributions. Physical derivations in section 10: battery specific energy and stored sortie energy; momentum-theory endurance scaling; ERP and EIRP sums and per-band subtotals from the filed MPE study; composite exposure contours under the OET-65 far-field formula with the FCC percentage-summing rule, and the single-limit error case; the far-field to cylindrical-model crossover radius and the cylindrical-model contour inside it; effective aperture, intercepted power, signal separation and implied front-end rejection at L1; ground sample distance, standoff for a 0.8 mm feature at one and three pixels, derived Matrice 4 series pixel pitch, Airy and Rayleigh limits, effective resolution, and the entrance-pupil diameter that would resolve 0.267 mm at 10 m; capture rate and the area threshold it sets; and fuel energy per gallon, energy per vehicle mile, sortie-equivalent driving distances and the assumed-distance ratios. Working was executed and checked programmatically on 5 and 6 August 2026. modelled
43. 43 Ronnie Gomez, project statement, Industrial Research Fellowship, The Charlot Lab, Institute for Physical AI @ John Bailey Institute, 2026. The fellow's own framing of scope and intended outcomes; used to define the four capabilities the review tests and not as evidence for any quantitative claim. self-published
44. 44 DJI, "Matrice 350 RTK Specs", enterprise.dji.com/matrice-350-rtk/specs, accessed 6 August 2026. Max takeoff weight 9.2 kg; weight with two TB65 batteries approx. 6.47 kg; max flight time 55 min at approximately 8 m/s without payloads in a windless environment to 0% battery; max wind speed resistance 12 m/s, published without a takeoff-and-landing qualifier; 2110s propellers; TB65 battery 263.2 Wh, approx. 1.35 kg. Manufacturer figures under manufacturer-chosen conditions. This review did not locate a hover endurance figure, a drag polar, a power-against-wind curve or a rotor figure of merit for this aircraft in DJI's published specifications. self-published
45. 45 DJI, "Matrice 4 Series Specs", enterprise.dji.com/matrice-4-series/specs, accessed 6 August 2026. Matrice 4E takeoff weight 1219 g; max flight time 49 min at approximately 9 m/s windless to 0%; max hover time 42 min windless at sea level from 100% to 0%; max wind speed resistance 12 m/s stated explicitly as during takeoff and landing, with no in-flight wind figure located by this review; max pitch angle 35 degrees; battery 99.5 Wh, 401 g; telephoto 1/1.5-inch CMOS, 48 MP, 168 mm equivalent, f/2.8. This review did not locate a published pixel size for any camera in this series. self-published
46. 46 DJI, "Zenmuse P1 Specs", enterprise.dji.com/zenmuse-p1/specs, accessed 6 August 2026. Sensor 35.9 x 24 mm, 45 MP effective, 4.4 micrometre pixel pitch, 8192 x 5460 photo size; DL 35mm F2.8 lens; published ground sample distance for the 35 mm lens equals flying height divided by 80 in centimetres per pixel; minimum photo interval 0.7 s; weight approx. 800 g. Recomputation from the stated sensor geometry gives flying height divided by 79.5 on the stated pitch and 79.9 on the 35.9 mm width across 8192 columns, so the published formula reproduces to within 1%. self-published
47. 47 US Department of Defense, "Global Positioning System Standard Positioning Service Performance Standard", 5th edition, April 2020, section 2.2.1.1. L1 signals are transmitted with enough power to ensure a minimum received signal power of -158.5 dBW for the L1 C/A-code under IS-GPS-200 conditions, namely a 3 dBi linearly polarised receiving antenna near the Earth's surface with the satellite above five degrees elevation. The 3 dBi reference is material: any comparison against a 0 dBi aperture carries a 3 dB offset. verified
48. 48 US Federal Communications Commission, 47 CFR 1.1310, "Radiofrequency radiation exposure limits", Table 1 to paragraph (e)(1). For 300 to 1500 MHz the maximum permissible exposure is f/1500 mW/cm2 over 30 minutes for general population and uncontrolled exposure and f/300 mW/cm2 over 6 minutes for occupational and controlled exposure, f in MHz; for 1500 to 100,000 MHz the limits are 1.0 and 5.0 mW/cm2. The rule sets exposure limits for people and does not restrict where an uncrewed aircraft may fly. verified
49. 49 US Federal Communications Commission, Office of Engineering and Technology, Bulletin 65, Edition 97-01, August 1997. Equation (3), page 19: power density equals radiated power times antenna gain divided by four pi radius squared. Equation (20), page 32, Tell's cylindrical model for sector antennas: S = (180/BW) Pnet / (pi R h). Page 31 states that far-field equations may significantly overpredict the RF environment very close to an antenna while the cylindrical model becomes overly conservative further out, with the crossover depending on aperture dimension and gain. Page 33 states that at multiple-transmitter sites each contribution is assessed against the limit applicable to its particular transmitter and the percentage contributions are then added. The 1.64 dipole reference factor is given in the bulletin itself. verified
50. 50 SAI Group for AT&T, "Maximum Permissible Exposure Study, Theoretical Report, Site MA1306 Ashland Cedar Street", by Sanket Joshi, 4 March 2019, filed with the Town of Ashland, Massachusetts. Twelve panel antennas, four per sector, radiation centre 75 ft. Eight technology channels on one AT&T sector, not eight operators. In the study's own table order, channel, sector ERP in dBm and gain in dBd: UMTS850 59.04 / 11.26; LTE700DE 59.55 / 10.55; LTE1900 64.63 / 13.85; LTE700BC 63.35 / 11.35; LTE850 61.25 / 12.25; LTE WCS 64.55 / 15.55; 5G850 60.00 / 12.25; LTE700DE KMW 58.75 / 9.75. Per-band EIRP at the 1.64 dipole factor: 700 MHz 6255.3 W, 850 MHz 5141.7 W, 1900 MHz 4762.6 W, WCS 4675.7 W, total 20,835.3 W from 12,704.4 W ERP; the study's own summary gives a Cellular general population figure of 0.59 mW/cm2. A single-site, single-carrier theoretical study by the carrier's consultant with no field survey, reporting at 6 ft above ground with the vertical pattern taken into account, closest evaluated slant range about 21.03 m. The 65 degree beamwidth and 1.8 m aperture used in this report's crossover calculation are assumed and are not values from this study. Not representative of the estate. reported
51. 51 US Federal Communications Commission, 47 CFR 24.232(a). Base stations with an emission bandwidth of 1 MHz or less are limited to 1640 watts EIRP at antenna heights up to 300 metres HAAT, and above 1 MHz to 1640 watts/MHz on the same condition. This governs the 1900 MHz PCS channel at the cited site; at 1640 W/MHz the site's 4762.6 W channel is within the ceiling for any emission bandwidth of 2.904 MHz or more. The AWS ceiling in 47 CFR 27.50 governs none of that site's channels, because it carries no AWS equipment. verified
52. 52 US Department of Transportation, "GPS Adjacent Band Compatibility Assessment, Final Report", 26 April 2018, with preliminary interference tolerance mask results presented by K. Van Dyke, S. Mackey and H. Wassaf at the 56th CGSIC meeting, ION GNSS+ 2016, reported by D. A. Divis, Inside GNSS, 16 October 2016. Eighty receivers across six categories tested with a 10 MHz LTE interferer against a criterion of 1 dB degradation in carrier-to-noise density; high-precision receivers were the most susceptible, degrading at received interference power as low as about -90 to -95 dBm near 1550 MHz. A US government primary, but the specific numbers reach this review through trade-press reporting of DOT briefings, and transportation.gov blocked automated retrieval of the final report. verified assessment, figures reported
53. 53 ASME Boiler and Pressure Vessel Code, Section V, Article 9 (Visual Examination), T-953. A fine line of 1/32 in. (0.8 mm) or less in width may be used to demonstrate a visual examination procedure, and remote visual examination systems shall be demonstrated to have resolution capability and lighting intensity at least equivalent to direct visual observation. The code is paywalled and this wording is quoted from secondary reproductions rather than a direct read. Used as the nearest published anchor this review located for the size of feature a visual inspection is expected to resolve; the inspection-standard literature was not surveyed exhaustively, and this is not a telecommunications structure standard. verified standard, quoted secondarily
54. 54 US Department of Energy, Alternative Fuels Data Center, "Fuel Properties Comparison", accessed 6 August 2026. Gasoline and E10 lower heating value 112,114 to 116,090 Btu/gal (118.29 to 122.48 MJ/gal); low sulfur diesel 128,488 Btu/gal (135.56 MJ/gal). The gasoline row is stated as a range covering gasoline and E10, and this report quotes the range and names the endpoint it uses. This review did not locate a density, specific gravity or Btu-per-pound figure for gasoline in it, nor in EIA's thermal conversion factor documentation or metric conversion appendix, which give motor gasoline energy per barrel. verified
55. 55 US Department of Energy, Alternative Fuels Data Center, "Average Fuel Economy by Major Vehicle Category", data year 2024, citing Federal Highway Administration, Highway Statistics 2024, Table VM-1. Car 25.6 mpg; light truck or van 18.5 mpg; delivery truck 7.19 mpg; Class 8 truck 6.30 mpg; refuse truck 5.24 mpg. Gasoline gallon equivalents and fleet averages across all operating conditions, not a rating for any particular service vehicle. verified
56. 56 US Environmental Protection Agency and National Highway Traffic Safety Administration heavy-duty greenhouse gas and fuel efficiency programme, 49 CFR Part 535 and 40 CFR Part 1037. Vehicles above 8,500 lb GVWR are outside light-duty fuel economy labelling, and vocational vehicles, a class that expressly includes utility bucket trucks, are regulated in gallons per 1,000 ton-mile and carry no federally published miles-per-gallon rating. This establishes the absence of a regulatory per-vehicle rating only; fleet measurements are published, at [61] and [62]. verified
57. 57 J. G. Leishman, "Principles of Helicopter Aerodynamics", 2nd edition, Cambridge University Press, 2006, chapter 2. Momentum theory for a rotor in hover: induced velocity equals the square root of thrust divided by twice air density times disc area, so induced power scales with the three-halves power of thrust. Applied here at ISA sea-level density with a constant figure of merit assumed. Profile power on a fixed-pitch rotor scales on close to the same exponent, leaving avionics and payload draw as the terms genuinely independent of thrust. verified (standard text)
58. 58 E. Hecht, "Optics", 5th edition, Pearson, 2017, chapter 10. The Airy disc diameter equals 2.44 times wavelength times f-number and the minimum resolvable separation under the Rayleigh criterion is its radius, 1.22 times wavelength times f-number, both evaluated here at 550 nm. The angular form, 1.22 times wavelength divided by entrance pupil diameter, is the basis for the entrance-pupil threshold in section 10. verified (standard text)
59. 59 US Department of Energy and US Environmental Protection Agency, "Where the Energy Goes: Gasoline Vehicles", fueleconomy.gov. Only about 12% to 30% of the energy from the fuel put in a conventional vehicle is used to move it down the road, based on Oak Ridge National Laboratory analysis of over 100 vehicles. Applying this share to the vehicle while leaving the aircraft on stored energy is a mixed basis, labelled as such where used. verified
60. 60 ANSI/TIA-222 Revision I, effective 1 January 2024, adds unmanned aerial systems as an acceptable method for collecting data on existing structures, as described by American Tower and by the Telecommunications Industry Foundation planning advisory notice for ANSI/TIA-222-I, 2024. The standard itself is paywalled and this review did not read it, so this review relies on industry characterisations and does not establish, in either direction, whether the revision carries an image quality, resolution or standoff requirement. reported (paywalled standard, characterised by industry sources)
61. 61 Utilimarc heavy-duty bucket truck benchmarking, reported as "Heavy-Duty Bucket Truck Operating Costs Increasing", Government Fleet, 28 September 2015, retrieved and read 6 August 2026. Average fuel economy 5.18 mpg across 2010 to 2014, high 5.47 in 2011 and low 5.0 in 2010, from more than 50 utility and municipal fleet benchmarking clients and more than 23,600 vehicles between 33,000 and 55,000 lb GVWR, all running diesel or alternative fuel. Commercial benchmarking reported through the trade press: not a government series and not per-vehicle certified. The Government Fleet report gives aggregate figures only; this review did not locate the underlying vehicle-level data, and Utilimarc's benchmarking service is commercial. The mixed fuel composition makes the diesel heating value used in section 10 an assumption. reported
62. 62 US Department of Energy, National Renewable Energy Laboratory, Fleet DNA Project. The project publishes second-by-second GPS and CAN duty-cycle data for medium- and heavy-duty vehicles by vocation, and the bucket truck vocation covers 20 vehicles and 283 operating days, class 3 to class 7, in utility and telecommunications applications. Characterised from NREL programme descriptions retrieved 6 August 2026; nrel.gov and docs.nrel.gov did not resolve from this review's network during compilation, so no value has been taken from the dataset itself. Named because it is the most likely public source for a measured bucket-truck duty cycle. reported

Market and economic review, not investment advice. Figures carry the verification grade under each reference: verified (peer-reviewed or independently replicated), reported (a named source stated it, not independently confirmed), self-published (the organisation's own figure) and modelled (computed here, not measured). A modelled figure is never presented as a measurement. Corrections are welcome and will be recorded.