The future of engineering work
For a thought experiment, we took an engineer's week apart into tasks and used public data to answer three questions: what can AI speed up, by current research? What is already in use today? And what do we do with the time we gain?
Some of our tasks stay with people, some get help from AI, and some it takes over entirely. New tasks are added.
The way of looking at it is inspired by Anthropic's view of our economic future, which takes a job apart into tasks in just this way.1 Every dot below is a quarter of an hour of an engineer's week.
Forty hours, Monday to Friday
Each dot is fifteen minutes. An engineer's week has 160 of them.
Where saved time went before
Picturing the future is always hard. Sometimes looking back helps. Four times a technology took over a large share of engineering or office work, and the time it saved went into more and better work.
Welded ships, 1940s
Welding prefabricated sections instead of riveting on the slipway cut the build time six-fold by mid-1943. The yards did not stand idle: at the peak, three ships were launched a day, about 2,700 in all.13
Faster meant more.
Crash simulation, 1990s–2000s
A physical prototype loop cost over $300,000, a simulation under $5,000. Engineers did not bank the time: virtual crash tests now outnumber physical ones by orders of magnitude, and models grew tenfold in detail. For a recent 7 Series BMW ran around 90,000 simulations.14,15
Cheaper iterations meant more of them, and better cars.
Digital design, Boeing 777
The 777 was the first airliner designed entirely on screen. Digital pre-assembly removed 65% of change errors and rework, 15 points more than the target.16
The time went into getting it right the first time.
Spreadsheets, since 1980
Software took over most of the arithmetic. Routine bookkeeping shrank, analytical accounting grew: in 1978 there were almost two clerks per accountant, today there are more accountants than clerks.17
Routine shrank, judgement grew.
Not every story ends like this: bank tellers survived the ATM, then lost ground to mobile banking. But engineering and project work still hold so much untapped potential, in breaks between systems, duplicated work and waiting, that time gained there will be needed for a long while yet.
The trend continues
The ratings above describe what AI could do in 2023. METR, an independent research group, measures how long a task AI agents can finish on their own, counted in the time a skilled person would need. In early 2023 it was about four minutes, at the end of 2025 about five hours, and in spring 2026 the measurement passed sixteen hours, where METR's task suite stops measuring reliably. Over the whole period the length doubled about every seven months, for models since 2024 about every three.10
METR measures software and research tasks, and the agents succeed half the time, not always. Nobody measures this for engineering yet, CAD work included. Our view: if the trend holds even roughly, most of what an engineer does at a screen comes within reach of AI in the next few years, from checking a drawing to modelling a part. The physical world, judgement and responsibility do not.
Length of task, in the time a skilled person needs, that frontier AI agents complete half the time, on a log scale. Dots are METR's measurements; the line is the trend through them, dotted where it is carried forward. The engineering tasks illustrate the development.10
What research measures
of the work in engineering occupations is covered by AI use today by Anthropic's measure, about 46% would be possible. For programmers it is already 75%.5
Anthropic, March 2026time for individuals with AI in product development at Procter & Gamble, who matched the results of teams without AI.6
Dell'Acqua et al., 2025of expert tasks, mechanical and industrial engineering among them, where the best model matched or beat experienced professionals in blind comparison (as of October 2025).11
OpenAI GDPval, 2025What happens to the time gained
Hours saved are the first effect you can measure, and every use case has to answer for them. The more interesting question is where the hours go.
Covering for missing staff
As the large post-war cohorts retire, engineering teams lose people faster than they can hire. Time gained fills part of a gap that would otherwise stay open.
Higher quality
Check twenty variants instead of two. Read every requirement, not only the ones that look risky: an error found in operation costs 29 to more than 1,500 times as much to fix as one found in the requirements.12
More projects
The same team can handle more enquiries, price more quotations and run more orders, without every additional project needing a new hire.
More done in-house
Calculations, documentation and studies that went to service providers for lack of capacity can come back, and the knowledge with them.
Where to start
Not with what is technically possible, but where a lot of time is lost today. Take one role, list its tasks and mark honestly what stays, where AI reliably helps today and what it can already take over. Start where much time sits and the technology already carries the load, and follow with the rest as it matures.
Method: an engineer here is the average of all 34 engineering occupations in O*NET 29.3, weighted equally; their tasks are grouped by O*NET's generalized work activities. The time share per task is Anthropic's estimate; the rating (no meaningful speed-up, AI model alone, with software) is the human rating in Eloundou et al.; "in use today" is Anthropic's observation. The hours gained apply the lowest and highest time savings measured in controlled studies; new work is illustrated as a quarter of them. The occupational data is American.
Sources
- Anthropic: Scenarios for our economic future (Econ Scenario Explorer), September 2026
- O*NET 29.3 database, US Department of Labor: task statements, Tasks to DWAs, DWA and content model reference
- Anthropic: What work can robots do? (30 September 2026); per-task time shares in robot_exposure_tasks.csv
- Eloundou, Manning, Mishkin, Rock: GPTs are GPTs. Science 384, 2024; task ratings in full_labelset.tsv
- Massenkoff, McCrory (Anthropic): Labor market impacts of AI, 5 March 2026; task_penetration.csv and job_exposure.csv (engineering average: our calculation)
- Dell'Acqua et al.: The Cybernetic Teammate. NBER Working Paper 33641, 2025
- Noy, Zhang: Experimental evidence on the productivity effects of generative artificial intelligence. Science 381, 2023
- Humlum, Vestergaard: Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026
- Anthropic: Estimating AI productivity gains from Claude conversations, November 2025
- METR: Time Horizon 1.1 (29 January 2026) and time-horizons raw data, read 13 November 2026
- Patwardhan et al. (OpenAI): GDPval, arXiv 2510.04374, October 2025
- Stecklein et al. (NASA): Error Cost Escalation Through the Project Life Cycle, INCOSE 2004
- US National Park Service: SS Scotts Bluff, a forgotten contributor
- Spethmann, Thomke, Herstatt: The impact of crash simulation on productivity and problem-solving in automotive R&D. TU Hamburg-Harburg Working Paper 43, 2006
- BMW Group: Hybrid homologation, 6 February 2026
- IEEE Spectrum: Boeing's seventh wonder, October 1995
- US Bureau of Labor Statistics: Occupational Outlook Handbook 1980-81 and current edition (bookkeeping clerks; accountants and auditors)
Which challenge is keeping
your team busy right now?
Tell us, and in the demo we show you how Engera makes it easier.

