Start with an honest picture
You probably already have an opinion on AI assistants. Maybe ChatGPT wrote you a tidy email, or maybe Copilot gave you a confident answer that was wrong. Both happen, and the reason is the same. Before you use these tools on real engineering work you need a clear idea of where they help and where they trip you up.
Think of a general-purpose AI assistant (ChatGPT, Gemini, Claude, Microsoft 365 Copilot) as a very well-read graduate who has never been on site. It is quick and articulate and has read a huge amount. It has no professional indemnity, it has never seen your drawings, and it will sometimes sound sure about things it has made up.
How fluent text can still be wrong
A large language model (LLM) is trained on vast amounts of text to predict the next likely word fragment (token), given everything before it. It does that one token at a time until the answer is finished. The aim is text that looks right. Nothing in that process checks the answer against a code of practice or a calculator.
That explains three failure modes you'll meet in engineering work:
- Made-up references. Clause numbers, standard titles and paper citations follow patterns the model has learned well, so it can produce a convincing "BS EN ... clause 6.2.4" that says something different, or doesn't exist.
- Arithmetic slips. The model is predicting digits as text, not running a calculation. It can get the method right and the number wrong.
- Missing context. It doesn't know your site, your client's brief or your firm's design basis unless you tell it, and even then it can quietly assume a default.
Worked example: a simple beam
Here is the kind of exchange that catches people out:
Prompt: Simply supported steel beam, 6 m span, UDL 10 kN/m (factored).
What is the maximum bending moment?
AI: For a simply supported beam with a uniformly distributed load,
M_max = wL^2 / 8.
M_max = 10 x 6^2 / 8 = 10 x 36 / 8 = 54 kNm.
The formula is right and the working looks clean, but the answer is wrong: 360 / 8 = 45 kNm. Nothing in the tone warns you. The fix is an independent check, done by hand, on a calculator or with a few lines of code you can read:
# Independent check of the AI's bending moment
w = 10.0 # kN/m, factored UDL
L = 6.0 # m, span
M_max = w * L**2 / 8 # standard result for a simply supported UDL
print(f"M_max = {M_max:.1f} kNm") # -> M_max = 45.0 kNm, not 54
Gotcha: the polish of the explanation tells you nothing about whether it's right. Check every number the AI gives you that matters, by a route that doesn't depend on the AI.
Sorting your own tasks: three buckets
A useful first exercise is to sort your week into three buckets:
| Bucket | Typical tasks | Why |
|---|---|---|
| AI can help | Rewording an email, structuring meeting notes, explaining an unfamiliar term, brainstorming questions for a design review | Low consequence if imperfect, and you will read it anyway |
| AI can draft, a human must check | First draft of a method statement section, a spreadsheet formula, a summary of a long report, a list of possible failure modes | Useful speed, but errors could carry into real work, so check against sources and your own judgement |
| AI must not be used | Deciding a design is safe, signing off a calculation or certificate, making the final call on a safety-critical change | Professional accountability rests with a competent person and can't be handed to a tool |
The line between buckets is about consequence and accountability, not about how clever the tool is. A strong model drafting a temporary works check still belongs in the middle bucket at best. Your firm's policy may move tasks between buckets too. That's the next level.
What the evidence says (with dates)
Keep claims about "everyone using AI" anchored to real surveys, and judge each source:
- IET Skills Survey 2025 (YouGov fieldwork with 1,316 respondents with managerial responsibility at UK engineering and technology employers, 10 February to 13 March 2025; published October 2025). It found 58% of engineering employers use AI, but only 18% use it regularly. Reliability: fairly strong, with a large sample and an independent polling firm. It measures employers, not individual engineers.
- Association of Nordic Engineers (fieldwork December 2024, published February 2025, Nordic STEM professionals). It found 59% use AI at work, yet only 8% feel adequately trained and 23% say their company offers extensive AI policies or guidance. Reliability: a useful view from the professionals themselves, but it covers Nordic countries, not the UK.
- Bentley Systems, Pinsent Masons, Mott MacDonald and Turner & Townsend (published October 2025, 130 infrastructure and building professionals). It found that over a third of firms lack adequate project controls for managing AI risk in design and construction. Reliability: small sample, and the authors include a software vendor with a commercial interest, so read it as an indication, not proof.
Across all three the pattern is the same: wide experimentation, shallow and irregular use, and training and controls lagging behind. Professional bodies are responding. The Engineering Council's AI Working Group reported in June 2025 that AI should be embraced as an opportunity while its risks, especially ethical ones, are considered.
Try it on your own work
List ten tasks from last week. Put each one in a bucket and write one sentence on why. Be strict: if a wrong answer could reach a drawing, a calculation or a site, it doesn't go in the first bucket.
Key takeaways
- LLMs predict likely text. Fluent output can still hold wrong numbers or made-up references.
- Check every number that matters by an independent route: hand, calculator or code you understand.
- Sort tasks by consequence and accountability: help, draft-and-check, or not at all.
- Quote surveys with their date, sample and author, and say how far each one can be trusted.
- The evidence shows wide but shallow use, with training and controls lagging. Being careful puts you ahead, not behind.