Last updated: 27 September 2026
Short answer: AI is changing engineering documentation mainly by speeding up the first 80 percent of the work: drafting work instructions, brainstorming failure modes for an FMEA, summarising nonconformance reports and finding the right clause in a long specification. It does not change who is accountable. ISO 9001 and AS9100 are technology-neutral, so AI-assisted documents are acceptable as long as they are reviewed, approved and controlled like any other document. The risks are invented values, outdated revisions and confidential data leaving the building.
Where is AI actually useful in engineering documentation?
In aerospace and manufacturing, where I have spent my career, a large share of engineering time goes into writing, reviewing and finding documents rather than engineering. That is exactly the kind of work current AI tools handle well, with clear limits.
| Task | What AI does well | What a person must still do |
|---|---|---|
| Work instructions | Turn rough notes or a video transcript into clear numbered steps | Verify every value, tool and safety step against the approved process |
| FMEA | Suggest structure, functions and candidate failure modes | Rate severity, occurrence and detection from real process knowledge |
| Nonconformance reports | Summarise long NCR histories and group them by likely cause | Decide root cause and disposition |
| Inspection reports | Turn measurement data and notes into a readable narrative | Confirm the data, judge acceptance, sign |
| Specifications and standards | Find the relevant clause in hundreds of pages | Read the clause itself and interpret it |
| Translation | Produce a workable first translation of instructions | Have a qualified reviewer check technical terms |
| Meeting notes | Extract decisions and action items | Confirm owners and dates |
The pattern is consistent: AI is strong at language and pattern-finding, and weak at anything that needs knowledge of your specific process, your data or your customer's intent.
Can AI write an FMEA?
It can draft a useful starting point, but not a finished one.
The AIAG-VDA FMEA handbook, the reference for many automotive and industrial teams since 2019, defines seven steps: planning and preparation, structure analysis, function analysis, failure analysis, risk analysis, optimisation and results documentation. AI is most helpful in the middle steps. Given a product description or a process flow, it can propose a system breakdown, list the functions of each element and suggest failure modes, effects and causes that a team might not think of in the first session.
Research backs this up, with caveats. A 2025 paper in Design Science, AI-driven FMEA: integration of large language models for faster and more accurate risk analysis, reported faster analysis when language models supported the process, while relying on domain experts to validate the technical correctness of the output. Other studies have found models occasionally propose failure modes that human analysts missed. None of them suggest removing the engineers.
Risk analysis is where AI should step back. Severity, occurrence and detection ratings, and the Action Priority that the AIAG-VDA method derives from them, depend on field data, process capability and the controls you actually have. A model that has never seen your scrap data or your inspection plan can only guess, and a confident guess in a risk rating is worse than a blank cell.
Do ISO 9001 and AS9100 allow AI-written documents?
Neither standard prohibits it. Both are technology-neutral: they care about the result and how it is controlled, not the tool used to draft it.
The relevant requirement is clause 7.5, documented information. Controlled documents must be suitable for use, reviewed and approved before release, identified, version-controlled and protected. An AI-drafted work instruction meets that requirement in exactly the same way a hand-written one does: a competent person reviews it, approves it and releases it through your document control system.
Two practical points follow:
- The approver owns the content. "The AI wrote it" is not an acceptable answer in an audit or a failure investigation. The signature means the same thing it always did.
- Keep the review visible. Your document history should show who reviewed and approved each revision. Some organisations also note when a draft was AI-assisted, which is useful for later process reviews.
What are the real risks?
Invented values. Language models can produce plausible torque values, tolerances or material grades that appear nowhere in your drawings. In a technical document that is the most dangerous failure, because it looks right. Always give the AI the source data rather than asking it to recall figures, and check every number against the controlled source.
Outdated information. A model's general knowledge may reflect an older revision of a standard or specification. Work from the current controlled copy, not from what the model remembers.
Confidential and controlled data. Drawings, customer specifications and export-controlled technical data may not be allowed on a third-party service at all. Check contracts and export rules before anything goes into a cloud tool. For sensitive material, running a model locally may be the only acceptable option; I compared the choices in local vs cloud AI for a small business.
Personal data. Inspection records, training files and customer correspondence often contain names and contact details, which brings data protection law into play. I summarised the rules that matter in what engineers should know about data privacy laws in 2026.
Standards licensing. Published standards are copyrighted, and licence terms vary. Check yours before pasting large sections of a standard into an online tool.
What does ISO/IEC 42001 have to do with this?
ISO/IEC 42001, published in December 2023, is the first international management system standard for artificial intelligence. It sets auditable requirements for how an organisation governs its use of AI, covering risk assessment, impact assessment, lifecycle management and oversight of suppliers. Organisations can be certified against it.
A small engineering firm does not need certification to use AI sensibly. But the standard's structure is a good checklist for anyone writing an internal AI policy: decide what AI may be used for, assess the risks, control the data, and review how it is working.
How would I introduce AI into a quality or engineering team?
- Start with low-risk documents. Meeting notes, training material and first drafts of internal procedures, not certification records or customer deliverables.
- Write down what data is allowed. A one-page rule on what may and may not go into which tool prevents most problems.
- Build prompt templates. A good template for "turn these notes into a work instruction in our format" saves time and makes output consistent. The approach in my practical guide to prompt writing applies directly.
- Make review non-negotiable. Every AI-assisted controlled document goes through normal review and approval, with no shortcuts.
- Measure it. Track time saved and errors caught in review for a few months. That tells you honestly whether it is working.
What changes for engineers?
Less time on formatting and first drafts, and more on judgement: checking, deciding and signing. The skill that grows in value is the ability to review an AI draft critically and quickly, spotting the confident error in an otherwise tidy document. That has always been part of a senior engineer's job. AI makes it a bigger part.
FAQ
Can I use ChatGPT or Claude to write work instructions?
Yes, as a drafting tool, provided you are allowed to share the source information with that service. The draft must then be checked against the approved process and released through your normal document control.
Will auditors accept AI-assisted documents?
Standards like ISO 9001 and AS9100 do not prohibit AI. Auditors look at whether documents are accurate, reviewed, approved and controlled. An AI-assisted document that passes normal review is treated like any other.
Can AI calculate FMEA risk ratings?
It can suggest ratings, but it does not know your field failures, process capability or detection controls. Treat any AI-suggested severity, occurrence or detection value as a prompt for discussion, not an answer.
Is it safe to paste drawings or specifications into an AI tool?
Only if your contracts, customer requirements and export rules allow it, and the tool's data terms are acceptable. When in doubt, keep controlled technical data out of public AI services.
