The most dangerous document in your plant today is a chat window. An engineer pastes a customer spec into a chatbot, gets back a fluent paraphrase, makes a disposition on that paraphrase, and moves on. No revision number. No approval. No record. When the part ships wrong, the trail ends at a session ID nobody can reproduce. Quality management already has a name for this object. We have been writing it up in audits for thirty years: the uncontrolled copy.

This week's news cycle delivered a fresh wave of AI-governance launches — enterprise control planes, security partnerships, shadow-AI warnings, banks appointing AI chiefs. Every one frames the problem as new. It isn't. Two decades in automotive and aerospace quality taught me that this failure mode predates the interface by a generation. Only the interface changed.

The uncontrolled copy grew a voice

When I built the greenfield QA/QC department at SNOP — 900+ employees, quality systems written from a blank page — I wrote the document control procedure before we ran a single serial part. Within weeks of starting production I found the first shadow spreadsheet: a tribal tweak-sheet taped inside a machine cabinet, maintained by a setter who trusted it more than the controlled work instruction. Honestly, in places it was better maintained. That taped Excel printout is the ancestral form of today's chatbot summary.

ISO 9001 clause 7.5 and IATF 16949 document control do not exist because auditors love red stamp pads. They exist because organisations discovered, painfully and repeatedly, that copies of truth drift away from truth. An uncontrolled copy is a nonconforming information flow. A chatbot paraphrase of a spec is the same nonconforming information flow — it just arrives dressed as authority instead of looking like a taped-up spreadsheet.

The spreadsheet was at least honest enough to look unofficial; the chatbot arrives wearing a suit.

I build multi-model AI systems myself — a platform that runs dozens of models in parallel and synthesises their outputs. I know exactly how confidently a model will paraphrase a specification it has half-read. The fluency of the sentence and the fidelity of the content are independent variables. Anyone who claims otherwise has never watched a model summarise a clause it truncated at the token limit.

What a chatbot summary does to traceability

Follow a concrete case. A quality engineer asks a chatbot whether a cosmetic deviation falls inside the customer's acceptance criteria. The bot answers in confident prose. The engineer dispositions the part as accepted. Six months later the customer escalates. You open an 8D, and at D4 — root cause — the team hits a wall: source of the requirement: a chat session. No revision number to check against. No approval trail. No evidence of which model version, which prompt, which document was actually pasted. Dead end.

Auditors have written up uncontrolled documents since the first ISO 9001 certificate hit a wall. They simply haven't seen them in prompt form yet. Give it a season. The finding will read exactly like every uncontrolled-copy finding I have ever issued or received in IATF and AS9100 audits: a decision input that cannot be verified against a controlled source.

There is a second blade on this. As a Certified Ethical Hacker, my instinct on any uncontrolled channel is that it is also a channel out. An engineer pasting a customer spec into a public chatbot has transmitted controlled technical information to a third party with no DPA, no retention agreement, no idea where the text lands. Shadow AI is an exfiltration risk wearing the costume of a productivity risk. The quality risk and the security risk are the same risk: information flowing outside controlled paths.

Govern the copy, not the tool

The vendors this week are selling governance in a box. Buy what you need for the data-plane problem — the security half is real. But do not confuse a purchased platform with document control. You already own the instruments. Extend them.

  • Bring AI outputs under document control where they serve as decision inputs. If a summary informed a disposition, that summary is a quality record. Capture it, version it, reference it.
  • Run AI adoption through management of change. A tool that changes how requirements are interpreted is a change to the process. It goes through the same MOC gate as a new tooling insert.
  • Put it in the control plan. If AI-assisted decisions touch product conformity, the control plan must state where the controlled source of truth lives and how the AI path is verified against it.
  • Audit prompts like documents. Sampling prompts during internal audits sounds absurd until you recall we audited taped-up spreadsheets for decades with a straight face.

Make the controlled path the easy path. We lost the spreadsheet wars wherever the controlled document was harder to reach than the shadow copy. Poka-yoke applies to information as much as to assembly: if the engineer can get a verified answer from a controlled AI source faster than from a public chatbot, they will use it. If the controlled source is slower, no policy will save you.

Key takeaways

  • Shadow AI is the uncontrolled copy with a confident interface — treat it as a document control failure, not a technology problem.
  • Any AI output used as a decision input is a quality record: capture it, version it, and make it auditable.
  • Every uncontrolled channel is also an exfiltration channel — pasting specs into public chatbots is a security incident waiting for a classification.
  • The fix is not a platform purchase but extending ISO 9001 7.5 discipline, management of change and control plans to AI-assisted decisions.
  • Make the controlled path the fastest path, or the shadow path wins — poka-yoke for information.

Quality functions have run this control since before the first chatbot could hallucinate a torque value. The taped spreadsheet embarrassed us for decades but never pretended to be official. The chatbot does. This is the decade where document control — the least glamorous clause in the standard — earns its keep twice. So walk your floor this week, find the last AI-assisted decision anyone made, and ask the question an auditor would recognise instantly: show me its revision number.