The Deckhand newsletter · Issue 2 · Main Street

Engineering evolution with AI

In law, radiology, customer support and coding, AI takes the first-pass work and a named person still signs. On 15 public simulation problems, the best AI agent came within 10 % of the known answer on one.

Prepared: Analyst Deckhand (AI assistant) · Checked: Editor Deckhand · Approved: Vamsee Achanta · 3 November 2026

In FEABench, a test of AI agents on physics problems in a commercial simulation tool, 88 % of the best agent's commands to the solver ran, yet only one of 15 problems with known correct answers was solved within 10 %, and the authors doubt even that one (Google Research and Harvard, April 2025). NCEES, the US body that sets the engineering licensing exams, says the licensed engineer keeps responsibility and must check AI output independently (Position Statement 6.10, 2025). AI takes the drafting, searching and first-pass work. People keep the decision wherever a court, a regulator or a signature holds a named person to account.

Our hypothesis: AI has cooked coding, and it is now cooking many other areas, such as finance and legal. In each of them, AI has democratised the work, so that more people can now do it. Engineering will be democratised in the same way: with a basic understanding, anyone will be able to take on engineering problems. The future of the engineer then lies in engineering judgement. The question is whether engineering is pretty much cooked. We test it here, on the Deckhand journey, our series that takes real-world cases, builds the engineering behind them and gives it to the industry to reuse and build on.

On today's evidence, engineering work still needs the engineer. This article sets out that evidence, field by field, then two ways it could go for engineering. They are scenarios to argue about.

Law, radiology and customer support

In law, AI drafts and searches, but lawyers must review its output before relying on it (American Bar Association, Formal Opinion 512, July 2024). A public database lists 2,231 court decisions involving AI-invented content, 880 of them involving lawyers (Charlotin, 28 October 2026). The lawyer signs the filing.

In 2016, Geoffrey Hinton predicted that radiologists would be replaced. Today 76.2 % of the AI-enabled devices on the US Food and Drug Administration list are for radiology (Complizen count, September 2026), yet US radiologist numbers have grown about 10 % over the past ten years (Fortune, May 2026). The physician still signs the report.

In customer support, Klarna said in 2024 that its AI assistant did the work of 700 agents (Klarna, February 2024). In May 2025, Klarna's chief executive cited lower quality and the company began hiring people again (Bloomberg, via Entrepreneur, May 2025).

Coding: jobs, AI use, speed and code review

Coding is where AI arrived first and where the change is loudest. US tech companies cut more than 191,000 jobs in 2023, about 127,000 in 2025 and at least 94,046 from January to August 2026 (Crunchbase, October 2026). The reason given has changed. AI was cited in 33 % of tech layoff events in 2026, up from 1 % in 2024 (Layoffs.fyi, via Crunchbase News, September 2026).

Figure 1. US tech jobs cut each year, 2022 to August 2026

Crunchbase count of US-based tech companies (tracker updated 28 October 2026; 2026 to August from Crunchbase News, 25 September 2026). 2022 and 2023 are given as "more than", 2025 as "around". Data: fig1-us-tech-layoffs.csv.

Use is now near universal, and trust is not. In Google's DORA (DevOps Research and Assessment) survey of nearly 5,000 technology professionals, 90 % use AI at work and 30 % have little or no trust in the code it writes (Google Cloud, September 2025). In the Stack Overflow survey, 84 % of developers use or plan to use AI tools; 45.7 % distrust its accuracy and 32.7 % trust it (Stack Overflow, July 2025).

The speed gain depends on who measures it. On one simple lab task, developers with an AI assistant finished 55.8 % faster (Peng and others, February 2023). Across three company field experiments, developers completed 26 % more tasks, with the largest gains for less experienced staff (Cui and others, 2025). In a randomised trial, experienced open-source developers took 19 % longer with AI while believing they were about 20 % faster (METR, Model Evaluation and Threat Research, a non-profit research group, July 2025).

Figure 2. Measured and believed change in developer speed with AI, by study

Measures differ by study: time on one set task (lab), tasks completed (field), time on real tasks (randomised trial). The diamond is what the trial's developers believed. Sources: Peng and others 2023; Cui and others 2025; METR 2025. Data: fig2-ai-coding-speed.csv.

Review is still the bottleneck. In over 456,000 GitHub pull requests, a pull request being a proposed code change, people merged 76.8 % of human changes but 38 % to 65 % of those from five coding agents (Li and others, July 2025).

The jobs picture is upheaval, not collapse. US software postings rose almost 15 % from February 2025 to June 2026 while all postings fell about 7 %, yet they remain about 27.5 % below February 2020 (Indeed Hiring Lab, July 2026). The US Bureau of Labor Statistics (BLS) splits the outlook by task, not by field. It projects jobs for computer programmers to fall 7 % from 2025 to 2035, and jobs for software developers, quality assurance analysts and testers to grow 10 % (Figure 3).

Figure 3. US employment projections for 2025 to 2035 split by task, not by field

US Bureau of Labor Statistics, Occupational Outlook Handbook, pages modified 27 August 2026. The BLS cites AI automation for computer programmers. Data: fig3-us-job-projections.csv.

What AI takes and who still signs, by field

FieldWhat AI takesWho still signsEvidence
LawDrafting and searchThe lawyer, on the filing880 court decisions on lawyers' AI-invented content
RadiologyImage screeningThe physician, on the reportRadiologists up about 10 % in ten years
Customer supportFirst-line chatsA person, again, after a quality dropKlarna rehired in 2025
CodingWriting and changing codeThe reviewer who merges the changeAgent changes merged less often than human ones
EngineeringSet-up, runs, first draftsThe licensed engineer, on the sealSee below

Engineering: the evidence so far

On today's evidence, engineering is not cooked, for three reasons.

Our own team is one case. Its members include AI assistants, such as the Analyst Deckhand and the Editor Deckhand named at the top; they prepare and queue solver runs, check each other's work and draft reports. The analysts do the heavy specialist work, thus providing resource for the engineer to review; the engineer sets the method and the acceptance criteria, judges the results and signs. This article is one example. Vamsee Achanta set its question. The assistants found the sources, checked each figure against its source, built the charts and drafted the text. He reviewed the drafts, changed the title and approves the article.

That first pass is where young engineers have always learned. Across the wider economy, the first rung is narrowing. In payroll data, employment for workers aged 22 to 25 in the most AI-exposed jobs is 19 % below where it would be had it kept pace with less-exposed peers (Stanford, August 2026). Only 4.5 % of Indeed US software postings in the first quarter of 2026 were entry level (Indeed Hiring Lab, July 2026).

Two scenarios for engineering

The other professions show two paths, and engineering could take either. Each step below is drawn from something already seen in another field. The scenarios are ours, to argue about.

Good path: engineers who check
Bad path: output nobody can check
First-pass work

AI drafts, sets up and runs; engineers check more options per study.

Drawn from: Radiology: 76.2 % of AI devices, radiologists up about 10 %.

Firms cut the people who did the first pass, then quality drops.

Drawn from: Customer support: Klarna's 700 agents, then rehiring in 2025.

The entry rung

Juniors learn by checking AI first passes against hand calculations.

Drawn from: Field experiments: largest gains for less experienced staff.

Entry hiring dries up, so few people learn to check.

Drawn from: Software: 4.5 % of postings entry level; ages 22 to 25 down 19 %.

The seal

The named engineer checks and can reproduce the work.

Drawn from: NCEES 6.10 and the North Carolina board require independent checks.

Plans are sealed after a cursory look at AI output.

Drawn from: NSPE Case 24-2, where the Board found it unethical to seal AI-drafted plans after a cursory review; 880 court decisions on lawyers' AI-invented content.

Right answers

Problems with known answers become the routine check.

Drawn from: FEABench Gold: 15 public problems with target values.

Code that runs is taken as a right answer.

Drawn from: FEABench: 88 % of commands ran, 1 of 15 within 10 %.

Speed

Measured gains go into harder problems and more checking.

Drawn from: Field study: 26 % more tasks completed.

Felt speed replaces measured speed; delivery stability (how often changes fail and must be fixed) slips.

Drawn from: METR: 19 % slower while feeling 20 % faster; DORA 2025: lower delivery stability with AI use.

Figure 4. Two scenarios for engineering, each step drawn from something that has already happened in another profession

Our scenarios. Each "Drawn from" line names a source in the list at the end.

The difference between the two paths is who learns to check. In the good path, the first pass becomes training. In the bad path, it becomes a gap that shows only when the senior engineers retire.

The models change too. Each new model may do more of the engineering well, as advances in mathematics and physics are taken into the main model through further training. Engineering that once had to be written out for a model then becomes something it knows unaided. In our own work, the code once had to spell out the catenary equation for a hanging line; a frontier model (one of the most capable AI models of the day) now knows that equation without being given it. Figure 5 shows this direction in three stages. In the good path above, the check stays with the engineer at every stage.

Illustration of a direction, not measured data Stage 1: Written out in your own code. The engineer's files spell out the method and its equations.Stage 1Written out in your own codeStage 2: Given to the model in files. The model knows the method; the design code's edition and clauses, and the firm's own choices, come from files.Stage 2Given to the model in filesStage 3: Known to the model unaided. The model knows the method itself; the engineer still checks it against a known answer.Stage 3Known to the model unaided Next model generation: further training may take in more mathematics and physics.Next model generation: further training may take in more mathematics and physics. Each new model generation may take in more mathematics and physics through further training Example: the catenary equation for a hanging line. Our code once had to spell it out; a frontier model now knows it unaided (from our own work). once written out in our code now known to a frontier model Example: the catenary equation (hanging line) In the good path of Figure 4, the engineer checks the work against a known answer and signs it at every stage. Good path, at every stage: the engineer checks against a known answer, and signs
Figure 5. How the next model may take in engineering: an illustration of a direction, not measured data

Stages are in order only. The catenary example is from our own work. Hover a stage or the example for detail.

AI does the first-pass work that once trained engineers thus democratizing engineering. If so, engineers likely will evolve the judgement part.

Sources

Reusable prompts

From the sources listed above, compare who does the first pass and who still signs in law, radiology, customer support and coding; test the engineering case against FEABench Gold and the licensing boards' statements; and write a short article that states the hypothesis, with the charts and the two scenarios shown here.

Data: the public data set for this article is FEABench Gold. It holds 15 physics and engineering simulation problems, each with its target quantity, units, a plain-language model specification and the reference solution code for a commercial multiphysics solver. It is the set behind the "one of 15" result above. Source: Google Research, in the FEABench code and data release on GitHub (https://github.com/google/feabench, file data/benchmark.zip). Public status: public, Apache 2.0 licence. Date: last updated 7 May 2025; retrieved 29 October 2026. A reader can pose any of the 15 problems to the tool of their choice and compare its answer with the target value. The numbers behind Figures 1 to 3 are in the CSV files named under each figure.

Readers are welcome to share their own results with the FEABench Gold problems, or data that would improve this comparison, at support@aceengineer.com.

Questions: support@aceengineer.com · AceEngineer · © 2026 AceEngineer · 11511 Piping Rock Dr., Houston, TX 77077
Deckhand: Make Engineering Enjoyable Again