Walk any MES or PLM vendor’s booth right now and you’ll see the same demo. An engineering change order lands, the system diffs the new CAD model against the old one, pulls in the redlines, and within a minute or two spits out a formatted, step-by-step work instruction with images pulled from the model and text that reads like a human wrote it. It’s a genuinely impressive piece of software engineering. It’s also, if you publish it straight to the floor, a controlled document that nobody actually reviewed.
That’s the part of this trend that isn’t getting enough attention. The authoring problem — turning an ECO and a CAD delta into readable prose — is substantially solved, or close enough that it’ll be commodity functionality across MES and PLM platforms within a product cycle or two. The governance problem — deciding what a large language model is allowed to publish without a human in the loop — is not solved, and most plants adopting these tools haven’t thought it through yet. They’re inheriting a vendor’s default workflow, which usually means “human clicks approve” as the entire control strategy. That’s not a control strategy. That’s a rubber stamp with extra steps.
What the AI is actually doing
Strip away the demo polish and these tools are doing three things. First, they’re comparing CAD revisions and identifying geometric or BOM deltas — a new fastener, a relocated bracket, a changed tolerance. Second, they’re pattern-matching against a library of prior work instructions and batch records to infer format, step granularity, and phrasing. Third, they’re generating natural-language text and pulling exploded views or annotated images to match.
All three of those are inference tasks. None of them are engineering judgment. The model is very good at noticing that a bolt pattern changed from four holes to six. It has no reliable way of knowing whether the torque spec that applied to the old four-bolt pattern is still correct for the new one, whether the assembly sequence needs to change because a fastener now interferes with a harness routing step, or whether a safety callout that used to be adequate now needs to move earlier in the sequence because the new part introduces a pinch point. Those are exactly the categories of error that look completely plausible in generated text — the sentence reads fine, the number is in a reasonable range, the format matches every other instruction in the library — and are exactly the categories that hurt someone or scrap a lot of product if they’re wrong.
The three things that must never auto-publish
If you’re standing up one of these tools, or a vendor is telling you it’s coming in your next upgrade, draw a hard line around three categories before you draw it anywhere else.
- Torque and fastener specs. These are frequently carried forward from legacy instructions by pattern-matching rather than re-derived from the new engineering data. A model that infers “this looks like the same bolt as before” is not the same as an engineer confirming the joint design, material, and clamp load actually support that value.
- Sequencing changes. Reordering or inserting steps is where a GenAI draft is most likely to silently break an interlock, a cure time, or an ergonomic constraint that isn’t visible in the CAD model at all — it lives in process knowledge that was never in the training data.
- Safety callouts, PPE requirements, and lockout/tagout steps. These are the items most likely to get dropped, softened, or mis-positioned because the model is optimizing for fluent prose, not for regulatory or EHS completeness. A generated instruction that reads perfectly well and quietly omits a LOTO step is worse than no draft at all, because it looks finished.
Anything touching those three categories should route to a qualified engineer for explicit sign-off, every time, with no volume-based or confidence-score exception. That’s a policy decision, not a technology limitation — the model doesn’t know when it’s wrong, so you have to assume it might always be wrong in these categories.
What’s reasonable to fast-track
Not everything needs the same scrutiny, and treating every AI draft as equally risky is how review queues die and engineers start rubber-stamping out of fatigue. Cosmetic and structural changes — updated reference images, renumbered steps that don’t change order or content, reformatted text to match a template, pulled-in part numbers and descriptions from the BOM — are lower risk and can reasonably go through a lighter review: a single qualified reviewer, faster turnaround, still logged, still versioned, but not requiring the same sign-off chain as a torque spec change. The goal is a tiered gate, not a single choke point.
A workable three-tier gate
- Auto-publish with logging. Formatting, image refresh, non-substantive text cleanup. No content change to a step’s meaning. Logged for audit, no separate approval loop.
- Single-engineer review. New steps, reworded instructions, added detail, non-critical parameter changes. Requires a named reviewer and a documented reason, but doesn’t need a full change board.
- Full change-control review. Anything touching torque, force, sequence order, safety, or regulatory content. Routes through the same approval chain your quality system already requires for manually authored changes — because under most quality systems, it already has to.
Treat the draft as a draft, not as evidence
The most important framing shift for a plant adopting this technology: the AI output is a draft, not a validated document, and your electronic document control system needs to represent it that way. If your MES or EDMS lets a GenAI-generated instruction inherit the revision status of the document it’s replacing — effective immediately, no distinct pending state — you’ve built a system where a hallucinated torque value can reach the floor with the same authority as an engineer’s signature. Whatever platform you’re using needs a distinct “AI-drafted, pending validation” state that is visually and functionally different from “released,” and your operators need to know they will never see the former on the floor.
This also means audit trail requirements don’t relax because a machine wrote the first draft. If anything they get more important: you want a record of what the model generated, what a human changed, and why, so that when something goes wrong six months later you can tell whether the error originated in the ECO, in the model’s inference, or in a reviewer who missed it. Vendors are still catching up on this — many of the current GenAI authoring features log the final approved version but don’t cleanly preserve the diff between AI draft and human edit, which is precisely the record you’ll want if a nonconformance ever traces back to a work instruction.
What to actually do this year
If your MES or PLM roadmap includes GenAI-assisted authoring, don’t wait for the vendor to hand you a governance model — build the tiering above into your change-management procedure now, before the feature ships, so it’s a documented control rather than an improvised reaction the first time someone asks “did an engineer actually check this.” Pilot it on a low-risk document class first — cosmetic revisions, non-critical assembly steps — and deliberately withhold it from torque-critical and safety-critical instructions until you’ve seen a few cycles of how the model performs on your own CAD and ECO data, not the vendor’s demo dataset. The productivity gain here is real: cutting the time to draft a compliant instruction after an ECO is a legitimate win for engineering throughput. But the win only holds up if the humans who used to write those instructions from scratch are still the ones who decide the torque spec, the sequence, and the safety step — just later in the process, reviewing instead of authoring.
This article was written with the assistance of artificial intelligence. While we aim for accuracy, the information may be incomplete, out of date, or incorrect, and should be independently verified before you rely on it for any decision. It is provided for general information only and does not constitute professional advice.
