AI in the engineering office: speed without delegating the decision

An engineering office spends a considerable share of its time not designing, but retrieving: the right standard, the drawing from the last similar job, the calculation note that already dealt with this case. AI promises to shorten this digging and first drafting. The promise is real, but it stops at a clear boundary: AI prepares the material, the engineer decides. Confusing the two exposes a project to a design error that nobody has signed off.

Retrieving technical information: the safest gain

This is the most mature use and the least risky. An engineering office relies on a vast documentary base: standards and codes, supplier catalogues, drawings from past jobs, calculation notes, test reports, lessons learned. This material is scattered across servers, mailboxes and digitised archives, and retrieving it often comes down to the memory of a long-serving colleague.

An AI connected to these documents changes the nature of the search. Instead of looking for a file by its name, you ask a question in plain language: what minimum thickness was chosen for this type of vessel, which material was ruled out on that job. The mechanism that makes this reliable is RAG: the system first looks for the relevant passages in your documents, then writes an answer grounded in those sources, with their origin. Its principle, limits and governance are set out in RAG in industry.

Two caveats apply. An answer only has value if you can trace it back to the source document: a claim without a source cannot be verified and should not be taken as is. And quality depends on what you give the system to read: a base of scanned PDFs and poorly recognised tables produces mediocre search. The prerequisite is therefore to turn those PDF reports into usable data, and to tell apart what OCR or AI really does on technical documents.

Helping to draft, without signing in the engineer's place

Technical writing is repetitive in its form and demanding on substance: AI is useful on form, cautious on substance.

For a specification, it quickly produces a full outline from a few elements: the expected sections, reworded requirements, commonly forgotten points brought back in. The writer starts from a structured version rather than a blank page, then corrects and arbitrates. The gain is measured in formatting time, not in the relevance of the requirements, which stays an engineering act.

For a calculation note, the boundary is stricter. AI recalls the method, brings back a formula, structures the reasoning or spots an inconsistency in units. But it must not serve as a calculation engine: a language model produces a wrong numerical result with perfect confidence. Any value that feeds into a sizing must be recomputed with a verified tool and reread by a competent person. The gain concerns the method, never the responsibility for the figure.

Reusing what exists rather than starting from scratch

Many jobs resemble past ones, and an engineering office's expertise rests largely on its ability to reuse what has already been designed, tested and corrected. The problem is that this capital is poorly indexed: the right reference job exists, but no one remembers its number.

AI makes the connection. Describe the current need, and it surfaces comparable jobs, the changes decided along the way and the reasons recorded for them. It turns a dormant archive into an active memory, and acts as a guardrail against repeating errors already encountered.

Reuse does not remove the need for judgement. Two similar jobs always differ by a detail that matters: regulatory context, site constraint, material. AI brings them together, the engineer checks that the comparison is legitimate before transposing.

Assisting the standards compliance review

Checking that a design complies with the applicable standards and codes is meticulous, time-consuming and sensitive to fatigue. AI helps to prepare it: matching a code requirement to a point in the design, drawing up a checklist, flagging that a parameter seems to fall outside an admissible range.

The right word is prepare. AI does the rough work and draws attention to where you need to look, which counts when a reference framework runs to hundreds of clauses. But it does not pronounce compliance: it can misread a clause, miss an exception, or rely on an outdated version of a text. Compliance stays established and endorsed by an engineer, who commits the company. The tool reduces the load of the review, it does not replace it.

Generative design: real, but to be handled with care

This is the subject that stirs the imagination and calls for the most measured view. Generative design refers to tools that propose optimised geometries from fixed constraints: envelope, load to carry, mass to minimise. On well-bounded mechanical parts, the technology is real, sometimes striking.

The caveats, though, are heavy.

• The result is only optimal with respect to the constraints you managed to formalise. A forgotten criterion, an accidental load not declared, and the proposal is elegant but unusable.
• A generated geometry still has to be validated by the usual means: calculation, testing, manufacturability, maintainability, real cost.
• Integration into processes and manufacturing standards often limits the value: a shape impossible to inspect or repair has no place on an industrial site.

In practice, it is justified on a narrow, high-value scope, not by default: it produces leads to investigate, never solutions ready to manufacture. At the scale of an installation, the same simulation logic is found more usefully in the digital twin, which tests overall behaviour before touching the real thing.

The honest limit: the hand stays with the engineering office

It has to be said plainly: none of these uses shifts responsibility. AI proposes, retrieves, drafts, draws attention; it does not decide, does not sign, commits to nothing. The calculation note, the drawing, the declaration of conformity stay engineering acts, and the engineer answers for those choices to the company, the client and the regulator.

This split is not a stylistic precaution: it is the principle of human oversight that the European AI regulation sets down in black and white. A model that decided a sizing on its own would leave a technical decision that no one really made or checked, the worst situation in the event of an incident.

Confidentiality and intellectual property: not exposing your expertise

An engineering office's documents are its most valuable asset: drawings, methods, expertise, client data often under a confidentiality agreement. Sending them to a public AI service amounts to handing them to a third party, with non-trivial questions about their storage and reuse.

The answer is not to give up on AI, but to choose where it runs. For a sensitive base, a controlled deployment, on infrastructure the company controls, makes it possible to use the models without the documents leaving the perimeter. This is a trade-off to settle from the scoping stage, and the overall frame for these choices is set out in the guide to industrial AI.

• Treating AI as a calculation engine.A language model can produce a wrong figure with confidence. Every sizing value must be recomputed with a verified tool.
• Confusing a proposal with a design.Generative design produces leads to investigate, not solutions ready to manufacture.
• Transposing a past job without judgement.Two similar projects differ by a detail that matters: AI brings them together, the engineer validates the transposition.
• Sending confidential documents to a public service.Expertise and client data call for a controlled deployment, decided from the scoping stage.
• Believing AI establishes compliance.It prepares the review, it does not pronounce it: compliance stays endorsed by an engineer.