AI governance

The model never signs.
A person always does.

AI models read, extract and propose. They do not write to a governed record, release a certificate or sign anything. That boundary is in the architecture, not in a policy document, and this page is where it is drawn.

What the model touches

It reads. It extracts. It proposes.

Four jobs, all of them upstream of a decision. Note what is missing from the list: nothing here writes, releases or signs.

It reads the drawing

Every dimension, tolerance, note and feature control frame on the sheet, found and returned as a ballooned drawing with a characteristic list beside it. The numbering itself is deterministic and done in code, not decided by the model.

It reads what surrounds the drawing

Specifications, operation sheets and supplier documents, parsed into structured requirements rather than summarised into prose.

It answers with the source attached

A question about a project document comes back with the exact passage it was drawn from, so the answer can be checked rather than trusted.

It watches the line and suggests

The assistant reads the same live data the operator terminal reads, and proposes where to look next. It has no write path to the machine.

The boundary

What crosses it, and what it leaves behind.

A proposal crosses from the model to a person. Nothing crosses the other way. Each crossing leaves something an auditor can follow.

MODEL PERSON GOVERNED RECORD proposal signed no write path from the model to a record, a certificate or a signature

Four artefacts, and where each one comes from in the drawing above. Two are produced by the model, one by the person, and one is the record itself.

  • From the model

    Extracted value

    Carries the drawing sheet and zone it came from, so a figure challenged in a review is traced back to the millimetre.

  • From the model

    Answer

    Carries the passage it was drawn from. An answer with no source is not an answer.

  • From the person

    Signature

    A person re-authenticating at the moment of signing. It lands in the signature manifest, together with what was signed.

  • The governed record

    Quality record

    Append-only and hash-chained. Records are superseded, never edited or deleted, prior versions stay readable, and a sealed record cannot be removed by an administrator.

When it goes wrong

The three questions worth asking a vendor.

A model that is right most of the time is a liability if nothing catches the rest. These are the answers, and they are the reason the boundary above is drawn where it is.

What if the AI model is wrong?

It proposes; a person confirms, corrects and signs. A wrong proposal is corrected before anything is recorded, and the correction is part of the record.

What if the AI model is unavailable?

The platform still works. Ballooning, inspection, the shop floor and the record do not depend on a model being reachable.

What if the conditions are not met?

Interlocks refuse. An operation out of sequence, a gauge out of calibration, a certificate with a non-conformance open: the system returns early rather than warning and continuing.

Questions

Asked plainly, answered plainly.

Does the AI make decisions?
No. It produces a proposal and a person disposes of it. Every governed action, a write to a quality record, a release, a signature, is taken by a named person who re-authenticates to take it.
Can we see why it proposed something?
Yes. Extracted values carry the sheet and zone they came from, and retrieval answers carry the passage they came from. Provenance is attached at the point of extraction rather than reconstructed afterwards.
What happens to a wrong proposal?
The engineer corrects it before it becomes a record, and the correction is itself part of the record. Nothing the model produces reaches a governed record without a person in between.
Does the platform stop working without the model?
No. The model does a first pass on documents. Everything else, the shop floor, inspection, genealogy and the record, runs whether or not a model is reachable.
Is our data used to train models?
Your platform runs in its own environment, separated at the cloud tenant and repository boundary, sharing no infrastructure, credentials or data with anyone else. Data residency is in an EU region.
How is this different from adding a chatbot to an MES?
A chatbot sits beside the system and answers about it. Here the boundary between proposing and recording is built into the platform, so the governance holds whether or not anyone is watching the assistant.

Next step

Bring us the drawing you would not trust a model with.

The fastest way to judge this is an awkward part and a sceptical engineer. Give us both, and compare what comes back against what your engineer would have produced by hand.