The workflow you already run#
When a batch drifts out of spec, quality has to write a deviation report — root cause, impact, disposition. It's slow, so a model drafts it from the batch record and the deviation log. The report comes back well-structured and the reviewer signs. Cycle time drops. Everyone's pleased.
The risk you can't yet see#
Then the audit arrives. An investigator asks the question audits always ask: how was this conclusion reached, and can you show the trail? The model read the batch record, weighed some numbers, and wrote a confident paragraph — but which readings did it use, were the units consistent, and did anything in the source violate the SOP it was supposed to apply? None of that was captured. A fast report with no provenance is, to a regulator, a liability wearing a nice format.
Why nobody misconfigured anything#
The model was asked to draft a deviation report from a batch record, and it did. It read the values, weighed them, and wrote a conclusion in the house style. The conclusion was, as far as anyone could tell, correct.
The problem is that "as far as anyone could tell" is not a standard that survives an audit. In a regulated environment a conclusion is only finished when the path to it can be shown. That is not bureaucracy. It is what makes a conclusion checkable by someone who was not there.
A model that reasons well and records nothing produces work that has to be redone by hand before it can be used. That is worse than slower, because it looks finished.
What a wrong one costs here#
Deviation reports feed decisions about batches, and batches have a disposition. So a wrong report is not a document problem. It is a product problem, and it is expensive in a direction most AI business cases never model: not rework, but a hold.
There is also a subtler cost. If the trail is missing, an auditor cannot distinguish a wrong conclusion from an unprovable one. Both get treated the same way, which means good work is penalised alongside bad.
That is the argument for recording before automating. The record is not overhead here. It is the thing that lets a right answer count as one.
The smallest version that produces a number#
One product line, one deviation category, six weeks. The model keeps drafting as it does now. Watch mode records the inputs it read, the values it weighed, and the conclusion it reached.
At the end you can put two things side by side: the report as drafted, and the trail behind it. Then ask a quality reviewer the only question that matters — is this enough to stand behind? Their answer is worth more than any accuracy figure, because they are the person who will have to.
If the answer is yes, you have a defensible workflow and evidence for it. If the answer is no, you have learned it for the price of six weeks rather than during an inspection.
What an Inspector caught#
Started in Watch mode, an Inspector sat in front of the same drafting call and recorded what it would have done:
On record#
Units it would have normalized before the model reasoned over them. A reading outside spec that the source log never flagged, surfaced and recorded. And every figure in the draft traced back to the line it came from — the provenance an auditor actually wants. Signed, on the manufacturer's own stack, nothing run. That's Watch: prove the trail before you trust the draft.
The drafting step was untouched, and the model kept its speed advantage. Nothing was handed any authority it did not already have.
What the site gained is a conclusion that can be checked by someone who was not in the room. In a regulated environment that is not a nice-to-have on top of the answer — it is the difference between an answer and a finding. A fast draft nobody can verify still has to be redone before it counts.
What to check in your own stack#
Four questions. None of them needs a tool to answer, and all four are worth asking before you buy one.
- What is in the payload, not what is in the prompt? The prompt is written by a person and reviewed. The payload is assembled by a query and never read by anyone.
- Who would notice a wrong one, and when? Name the person and the moment. If you cannot, the review path does not exist yet.
- Where does the model run, and what leaves the tenant? "In our cloud" and "in our tenant" are different sentences.
- What record survives the conversation? If the answer is application logs, the answer is no.
What this does not tell you#
An Inspector in Watch mode reports what it would have caught. It does not tell you whether the workflow was worth automating, and it does not find a rule you have not written. Your policy is the ceiling.
It also will not fix data it cannot reach. If the access is not there, Watch mode says so in week one — an uncomfortable finding rather than a deliverable, and better in week one than month nine.
Related