LOThe question
AI in skilled jobs: what it does now, and who still signs
Put a dozen skilled occupations side by side and the same shape appears. Machines have taken the parts of the work that produce material — the photographs, the drafts, the scripts, the routine changes — and they are fast at it. What they have not taken is the decision that the material is right, and in most of these jobs the law says that decision belongs to a named, qualified person who signs for it.
What moved: the photographing, the drafting and the routine change
Delta has written drone inspections into the maintenance manual for its mainline fleet; the drone takes the pictures, and the airline says technicians and inspectors decide on an aircraft's condition up to 82% faster. NASA engineers let generative-design software propose the shape of flight parts once they have set the constraints. Verizon says its closed-loop automation executed over 70 million network configuration changes autonomously in 2025, and Walgreens says its 12 micro-fulfillment centers fill prescriptions for more than 5,000 stores.
In each case the machine produces something a person used to produce by hand — an image, a geometry, a configuration, a filled bottle. That is real change in real work, and none of it is hypothetical.
- Drone inspections after lightning strikes were added to the maintenance manual for Delta's whole mainline fleet, with technicians and inspectors judging airworthiness from the images
- NASA Goddard uses AI generative design to produce mission hardware in as little as an hour or two, and says the algorithms need a human eye because they can make structures too thin
- Verizon's closed-loop automation executed over 70 million network configuration changes autonomously in 2025, its chief technology officer wrote
- Walgreens said its 12 micro-fulfillment centers support more than 5,000 stores and fill over 3.5 million prescriptions a week, as it opened another in Minnesota
Where it still fails: the number that decides whether the work is right
The benchmarks recorded on this site show a consistent gap between producing and getting it right. On FEABench, models produced executable calls to a commercial simulation tool 88% of the time but almost never computed a result within 10% of the correct answer. On BacktestBench the best of 23 models reached 67.41% overall accuracy, with volatility and Sharpe ratio 'disaster zones' for all of them. On a benchmark of network configuration repair, the best model fully resolved only 25% of scenarios.
The pattern is not that models are useless — they run the software, write the code, draft the geometry. It is that the step where a qualified person asks 'is this number credible?' is exactly the step the models keep missing.
- On FEABench, AI models produced executable calls to a commercial simulation tool 88% of the time but almost never computed a result within 10% of the correct answer
- On BacktestBench, the best of 23 language models reached 67.41% overall accuracy, and volatility and Sharpe ratio remained 'disaster zones' for all of them
- On Cornetto, a benchmark of 231 network configuration repair problems, the best of nine language models fully resolved only 25% of scenarios
What stayed: the signature, by rule
Under US aviation rules the signature on a maintenance record is itself the approval for return to service, and only certificated people may give it. In Singapore, the licensed electrical engineer is the one entitled to take full responsibility for an electrical installation. In the US, anyone primarily responsible for designing an algorithmic trading strategy must register as a Securities Trader; in the EU, a person designated by senior management must authorise each deployment of a trading algorithm.
None of these rules mentions software, and none needed to. They were written around the question every one of these jobs eventually faces — who answers for this? — and that question does not have a machine as its answer.
- The maintenance record must carry the signature, certificate number and kind of certificate of the person approving the work, and that signature is the approval for return to service
- In Singapore, a licensed electrical engineer, who must be a registered professional engineer, is entitled to design and take full responsibility for any electrical installation
- People primarily responsible for designing or significantly modifying an algorithmic trading strategy must register with FINRA as Securities Traders
- In the EU, a person designated by senior management must authorise each deployment or substantial update of a trading algorithm, and records must show who approved every change
When the automation errs, a person recovers
At Cloudflare, a routing change pushed through its policy automation leaked routes for 25 minutes; the post-mortem records that a network operator reverted it by hand and paused the automation. NTT DOCOMO's maintenance agent presents recommended actions to engineers, who act on them. Unilever says its people provide the judgement needed to make the final decisions on a forecast its engine produces every week.
- A routing change pushed through Cloudflare's policy automation leaked BGP routes for 25 minutes until a network operator reverted it by hand and paused the automation
- NTT DOCOMO put an agentic AI into network maintenance that presents recommended actions to engineers, cutting response time for complex failures by more than 50%
- Unilever says its forecast engine produces a 104-week forecast every week across more than 5 million product–customer combinations, while its people make the final decisions
The exception: where the decision itself is moving — and how rare it is
There is one place in this evidence where the decision, not just the material, is moving to the machine: routine replenishment at the largest retailers. JD.com measures the share of replenishment decisions that models output and that are adopted without change, and reports a 5.27-day cut in turnover from a field deployment; Alibaba has been exploring a system in which algorithmic recommendations are final.
The same body of evidence also records how far this is from normal practice: across the EU in 2025, only 6.08% of enterprises that used any AI applied it to logistics.
- In a field deployment at JD.com, an inventory model cut turnover by 5.27 days, raised in-stock rates by 2.29% and cut holding costs by 29.95%, on items chosen for high automation rates
- Alibaba has been exploring a replenishment system in which algorithmic recommendations are final, after finding its algorithms beat human buyers on stock-outs and inventory
- Only about one EU enterprise in eighty used AI for logistics in 2025 — 6.08% of those using any AI — Eurostat's enterprise survey shows
What the official projections say about these jobs
US projections expect mechanical engineering employment to grow 11 percent and industrial engineering 12 percent from 2025 to 2035, and in both cases name automation as a source of demand — someone has to design, integrate and justify it. The one projection here that points down, network and computer systems administrators at 4 percent, gives three reasons, and only one of them is automation; the other two are work moving to DevOps teams and to outsourced providers.
- Mechanical engineering employment will grow 11 percent from 2025 to 2035, much faster than average, as factories add automation, the US Bureau of Labor Statistics estimated
- Industrial engineering employment will grow 12 percent from 2025 to 2035, partly on demand for automation expertise, the US Bureau of Labor Statistics estimated
- Network and systems administrator jobs will decline 4 percent from 2025 to 2035 as tasks move to DevOps, outsourcing and automation, the US Bureau of Labor Statistics estimated
What this does not establish
This is a pattern across the occupations added to this site on one night, not a survey of all skilled work, and the deployments come mostly from large companies describing their own systems. The rules are from a handful of jurisdictions and say who must answer, not how the work is done. Benchmarks date quickly; a model that fails today's test may pass next year's. Nothing here measures how many people these jobs employ or how their hours have changed.
Everything cited here
Each one opens the full record: its source and source tier, the dates, the scope it applies to, who verified it and what it does not establish.
- Drone inspections after lightning strikes were added to the maintenance manual for Delta's whole mainline fleet, with technicians and inspectors judging airworthiness from the images — Delta News Hub — Industry first: FAA accepts Delta's plan to use drones for maintenance inspections (Oct 4, 2024)
- NASA Goddard uses AI generative design to produce mission hardware in as little as an hour or two, and says the algorithms need a human eye because they can make structures too thin — NASA — NASA Turns to AI to Design Mission Hardware, by Karl B. Hille (Goddard Space Flight Center, Feb 09, 2023)
- Verizon's closed-loop automation executed over 70 million network configuration changes autonomously in 2025, its chief technology officer wrote — Verizon — Architecting autonomy (Verizon News, published Jun 24, 2026, by Yago Tenorio, chief technology officer)
- Walgreens said its 12 micro-fulfillment centers support more than 5,000 stores and fill over 3.5 million prescriptions a week, as it opened another in Minnesota — Walgreens — press release, Walgreens expands micro-fulfillment network with new facility in Brooklyn Park, Minnesota (Deerfield, Ill., May 20, 2025)
- On FEABench, AI models produced executable calls to a commercial simulation tool 88% of the time but almost never computed a result within 10% of the correct answer — Mudur et al. (Google Research, Harvard) — FEABench: Evaluating Language Models on Multiphysics Reasoning Ability, arXiv 2504.06260 (submitted 8 Apr 2025)
- On BacktestBench, the best of 23 language models reached 67.41% overall accuracy, and volatility and Sharpe ratio remained 'disaster zones' for all of them — Wang, Yang, Wu et al. (Beijing Normal University) — BacktestBench: Benchmarking Large Language Models for Automated Quantitative Strategy Backtesting, arXiv 2605.17937 (submitted 18 May 2026)
- On Cornetto, a benchmark of 231 network configuration repair problems, the best of nine language models fully resolved only 25% of scenarios — Protogeros, Asadli, Hoffman, Vanbever (ETH Zürich) — Benchmarking LLM-Driven Network Configuration Repair, arXiv 2604.22513 (submitted 24 Apr 2026)
- The maintenance record must carry the signature, certificate number and kind of certificate of the person approving the work, and that signature is the approval for return to service — Federal Aviation Administration — 14 CFR 43.9, Content, form, and disposition of maintenance, preventive maintenance, rebuilding, and alteration records (29 FR 5451, Apr. 23, 1964, as amended)
- In Singapore, a licensed electrical engineer, who must be a registered professional engineer, is entitled to design and take full responsibility for any electrical installation — Singapore Statutes Online — Electricity (Electrical Workers) Regulations, regulations 4 and 5 (current version as at 30 Sep 2026; last amended by S 452/2025 with effect from 1 July 2025)
- People primarily responsible for designing or significantly modifying an algorithmic trading strategy must register with FINRA as Securities Traders — FINRA — Rule 1220(b)(4), Securities Trader (announced in Regulatory Notice 16-21, published June 06, 2016; effective January 30, 2017)
- In the EU, a person designated by senior management must authorise each deployment or substantial update of a trading algorithm, and records must show who approved every change — Commission Delegated Regulation (EU) 2017/589 of 19 July 2016 (RTS 6, organisational requirements for algorithmic trading; OJ L 87, 31.3.2017)
- A routing change pushed through Cloudflare's policy automation leaked BGP routes for 25 minutes until a network operator reverted it by hand and paused the automation — Cloudflare — Route leak incident on January 22, 2026 (Bryton Herdes, Tom Strickx; January 23, 2026)
- NTT DOCOMO put an agentic AI into network maintenance that presents recommended actions to engineers, cutting response time for complex failures by more than 50% — NTT DOCOMO — press release on agentic AI for network maintenance (February 25, 2026)
- Unilever says its forecast engine produces a 104-week forecast every week across more than 5 million product–customer combinations, while its people make the final decisions — Unilever — Building resilience, driving growth: inside Unilever's AI-powered supply chain (Published: 13 August 2026), interview with its Chief Product Supply Chain Officer for Personal Care
- In a field deployment at JD.com, an inventory model cut turnover by 5.27 days, raised in-stock rates by 2.29% and cut holding costs by 29.95%, on items chosen for high automation rates — Zhao, Yu, Qi et al. (Tsinghua University, JD.com Supply Chain Tech, University of Hong Kong) — ORPR: An OR-Guided Pretrain-then-Reinforce Learning Model for Inventory Management, arXiv 2512.19001 (submitted 22 Dec 2025; v2 6 Jan 2026)
- Alibaba has been exploring a replenishment system in which algorithmic recommendations are final, after finding its algorithms beat human buyers on stock-outs and inventory — Liu, Lin, Xin, Zhang — AI vs. Human Buyers: A Study of Alibaba's Inventory Replenishment System, INFORMS Journal on Applied Analytics 53(5) (Published Online: 26 Sep 2023)
- Only about one EU enterprise in eighty used AI for logistics in 2025 — 6.08% of those using any AI — Eurostat's enterprise survey shows — Eurostat — Artificial intelligence by size class of enterprise (isoc_eb_ai), EU27, reference year 2025 (dataset updated 15 June 2026)
- Mechanical engineering employment will grow 11 percent from 2025 to 2035, much faster than average, as factories add automation, the US Bureau of Labor Statistics estimated — U.S. Bureau of Labor Statistics — Occupational Outlook Handbook: Mechanical Engineers (Last modified date: August 27, 2026)
- Industrial engineering employment will grow 12 percent from 2025 to 2035, partly on demand for automation expertise, the US Bureau of Labor Statistics estimated — U.S. Bureau of Labor Statistics — Occupational Outlook Handbook: Industrial Engineers (Last modified date: August 27, 2026)
- Network and systems administrator jobs will decline 4 percent from 2025 to 2035 as tasks move to DevOps, outsourcing and automation, the US Bureau of Labor Statistics estimated — U.S. Bureau of Labor Statistics — Occupational Outlook Handbook: Network and Computer Systems Administrators (Last modified date: August 27, 2026)