13 October 2026

What engineering knowledge is valuable for AI training (and what isn't)

If you run a UK engineering consultancy, your document management system is full of material that looks impressive on paper and turns out to be almost useless for AI training once you open it. A finished drawing pack without the reasoning behind it, a fee proposal, or a set of emails about a programme delay will rarely help a model learn how a chartered engineer actually thinks.

The reverse is also true. A modest options appraisal, a red-lined calculation pack, or a senior reviewer's mark-ups can be far more valuable than a glossy final report. This article sets out, in practical terms, what AI developers tend to want from engineering firms, what they usually do not, and how to judge your own archive before you speak to a data partner.

Why "reasoning" beats "outputs"

Most general-purpose models have already seen a great deal of public web text. What they still struggle with, especially in safety-critical domains, is the intermediate work that professionals do under real constraints: codes, ground conditions, budget limits, programme pressure, and the need to leave a defensible audit trail.

That gap shows up clearly in research. The EngTrace benchmark was built specifically to test whether large language models can follow physically grounded engineering reasoning, not just produce a final number. Its authors argue that engineering requires the synthesis of scientific principles, quantitative modelling and practical constraints, and that outcome-only tests miss flawed intermediate steps that would matter on a live project. Separate work on EngiBench similarly stresses open-ended engineering problem solving across systems, structural and chemical domains, rather than multiple-choice recall.

In commercial terms, that research direction matches what specialist data sellers are starting to offer. The US firm JWSE, for example, now offers licences to anonymised, rights-screened portions of more than 30 years of civil and environmental engineering work, including dated CAD revisions and the project work around them, on the premise that real project evolution is harder to scrape than textbook answers. For a UK consultancy, the practical point is that if your files show how a decision was reached, they are closer to what the market wants than files that only show what was finally issued.

Material that is usually valuable

Design rationale and options appraisals

Reports that set out why option B beat option A, what was discarded, and which risks drove the choice are high value. They contain trade-offs, assumptions and constraint handling in one place. Stage reports that merely restate the preferred scheme with little comparison are weaker.

Worked calculations with commentary

Numbers on their own say little, and calculations only become useful when they include load cases, partial factors, code clauses applied, sensitivity checks, and notes on where judgement overrode a default. A pack that a checker could follow without calling the author is usually a good candidate.

Inspection, condition and diagnostic reports

Observations tied to diagnosis and recommended remedial work teach models how engineers move from evidence to action. Photographs and sketches help only when they are accompanied by that narrative, and only after personal data and identifiable addresses have been stripped.

Checking comments, red-lines and design reviews

Senior mark-ups on a junior engineer's work are one of the richest records of expert correction in any firm. They encode "what a competent reviewer looks for" in a form that is hard to invent synthetically. Technical queries (RFIs) and responses fall into the same category: a real problem, a considered answer, and often a negotiated resolution.

Lessons-learned and post-project reviews

Where outcomes are known, archives become more powerful. A note that explains why a detail failed in service, or why a temporary works design was revised on site, links process to consequence, which is the kind of record that is rarely published on the open web.

Structured session capture from practising engineers

Document archives are not the only source. Light-touch interviews or walkthroughs in which an engineer explains how they would approach a class of problem can produce clean reasoning traces, provided they are recorded with consent, anonymised, and approved before licensing.

Material that is usually less valuable, or harder to use

Drawings and models with no explanation

CAD files and drawing sets matter when they are paired with revision history, design notes or review comments. On their own, they often lack the verbal reasoning that training pipelines need, and they raise harder rights and anonymisation questions. Some buyers do want CAD corpora; they typically want revision sequences and context, not a dump of orphaned sheets.

Purely commercial documents

Fee proposals, invoices, change-order correspondence and commercial negotiations rarely teach engineering judgement. They also tend to be confidential for reasons that have nothing to do with technical IP.

Content you cannot clear for IP or confidentiality

Even excellent technical material is unusable if the client owns the copyright, your appointment forbids reuse, or a non-disclosure agreement blocks licensing. Under the Copyright, Designs and Patents Act 1988, section 11(2), employee-created works usually vest in the employer, but client appointments frequently assign or restrict those rights. Check the paperwork before you invest time in preparation.

Anything that cannot be anonymised properly

Names, signatures, homeowner addresses, photos of people, and distinctive project details that can be re-identified from planning portals or press coverage are personal data or confidentiality risks. The Information Commissioner's Office's anonymisation guidance stresses effective anonymisation: the risk of identification must be sufficiently remote. If you cannot get there, leave the file out.

Low-signal boilerplate

Template method statements copied project to project, generic H&S text, and empty transmittal cover sheets add volume without teaching anything. Buyers prefer fewer, denser examples over a large weak corpus.

A simple quality test for your archive

Before you shortlist folders, ask five questions of each candidate set:

Does it show reasoning? Could a competent outsider follow how the conclusion was reached?
Is the judgement professional and checked? Prefer material that went through review, not first drafts never issued.
Can you clear the rights? Firm ownership, client permissions, and confidentiality all need a yes.
Can you anonymise it? Personal data and re-identification risk must be dealt with before anything leaves.
Would you be comfortable approving the final extract? Nothing should be licensed without firm sign-off.

If a set fails any of those, it is usually not ready. Starting with a small, clean sample is better than promising an entire DMS.

How this fits the UK licensing climate

Licensing professional material is not a fringe idea. In March 2026 the UK government published its Report on Copyright and Artificial Intelligence, confirming that a broad text-and-data-mining exception with opt-out is no longer its preferred approach, and that it will keep monitoring the licensing market rather than intervening in it for now. Existing copyright law therefore remains the frame: permission through a licence is the clean commercial route for most firm-owned material.

That backdrop favours firms that can offer rights-cleared, anonymised, reasoning-rich datasets rather than raw scrapes of public PDFs, because quality and clearance count for more than volume.

Practical next steps for engineering firms

Pick one discipline or project type with rich commentary, for example bridge inspections or drainage options appraisals.
Pull a short sample and strip personal data.
Confirm IP and confidentiality with whoever owns the appointments.
Ask a data partner what format and density they need, then prepare only to that standard.
Insist on written terms for permitted use, exclusivity, duration and payment, including whether you are paid by revenue share or a one-off fee.
Approve every dataset before release.

Talk to ID43 Labs

ID43 Labs works with UK engineering firms to capture professional reasoning from approved documents and expert sessions, structure and anonymise it, and license the resulting datasets to AI labs on a revenue-share basis. You stay in control of what is shared. If you want an honest view of whether your archive has the kinds of material described above, start a partnership conversation at id43labs.com.

This article is general information, not legal advice. Take independent legal advice on IP, confidentiality and data protection for your own appointments and archives.