The record is automated. The liability isn't.

In 2019 I was hired to build quality from bare concrete – a greenfield automotive plant, nine hundred people, no forms, no signatures. The first system I designed was not a gauge room. It was a signature matrix: who may open a nonconformance, who may disposition it, who may close it. Every record that plant produced carried a name, and every name carried accountability. Rockwell has now wired FactoryTalk Analytics VisionAI verdicts straight into Plex QMS – detection, disposition and record close in one loop, no transcription, no queue. A genuine step for throughput. It is also the first time in twenty years of quality work that a record can be created, dispositioned and closed without anyone asking who the author is.

What the loop actually automates

Detection was never the bottleneck. Cameras have been ejecting parts on conveyor lines for thirty years. What this integration automates is everything downstream of the verdict: the disposition, and the record itself. The path runs image capture at the cell, inference at the edge, a confidence score against a threshold, a call – OK or NOK – then Plex opens the nonconformance, assigns the disposition code, triggers containment and writes the history. Minutes collapse into seconds. The queue at the quality terminal disappears.

Now look at what the QMS entry inherits, and what it does not. It inherits the result. It rarely inherits the model version behind the verdict, the threshold in force that morning, the date of the last retraining, the calibration state of camera and lighting, a hash of the image that triggered the call. Compare that with a CMM report: a single 0.02 mm dimension arrives with an instrument ID, a calibration certificate and a GR&R study behind it. A machine-authored disposition arrives with a result. Provenance drops out at birth.

When the record becomes evidence

Most days nobody notices. You notice on recalls, customer audits and courtrooms, and they ask the same question in different accents. A VDA 6.3 auditor phrases it politely: who inspected this part, by what method, and can you show the method was capable. IATF 16949 asks it through traceability and validation clauses. Opposing counsel skips the politeness – you claim the part was inspected on the twelfth: by whom, running which model version, at what threshold, retrained since when. The honest answer is a piece of software. Software cannot be cross-examined.

The product-liability lens from my LL.M. is blunt about this. Under strict liability the claimant must show the defect existed when the product left your plant; the QMS record is normally your best evidence that it did not. An entry without provenance is not evidence – it is a gap with a timestamp, and gaps in discovery get filled by whoever is motivated to fill them. Containment on a single escape can burn €100,000 in weeks. Litigation adds a zero. This is what plants miss: an unprovenanced QMS is not a defence, it is discovery material – handed over complete, every silent field mapped, every missing signature exhibit-numbered.

The record can also be attacked, not merely doubted. In the same news cycle, US agencies were warning that AI-generated attack scripts are being aimed at industrial control systems. Anyone who can reach a PLC can reach a confidence threshold – nudge it from 0.92 to 0.99 and a defect rate quietly evaporates. A habit carried over from the security side of my work: I do not trust a log that can be edited. If the quality record is not append-only, hash-chained and time-synchronised, it is not a record. It is a suggestion.

A record that cannot name its author cannot defend its owner.

Provenance is the new signature

Here is the part that costs almost nothing. Govern the vision layer exactly as you govern measurement equipment, because a neural net making accept–reject calls is a gauge. Treat it like one.

• MSA for the model. Attribute agreement studies, false-accept and false-reject rates tracked per version. A retrained model is a new instrument – it inherits nothing from its predecessor, least of all the capability study.
• Version control as calibration records. Model version, threshold, retraining date and dataset lineage written into the QMS entry itself, not parked in a data lake nobody audits.
• Tamper-evident logging. Append-only, hash-chained, synchronised to a trusted time source. A nonconformance timestamped against a drifting camera clock will not survive an auditor with a stopwatch, never mind cross-examination.
• Disposition authority stays human. The machine may propose scrap. Use-as-is is a judgement, not a classification – it needs a name from the authority matrix, the same matrix I drafted at that greenfield plant before the first inspection station was ordered.

Then put all of it in the control plan, not the IT budget. In the PFMEA the vision system is a control; controls get validated and versioned like any torque wrench, and the control plan should reference the model by name and revision. The old discipline still decides the rest: an 8D that leans on a camera verdict is exactly as credible as the QRQC that surrounded the failure.

Buy the integration. The throughput is real, and I have watched too many good engineers key nonconformances into terminals at shift end to mourn that queue. But audit the authorship before you scale it. A QMS is a system of record before it is a system of decisions, and the record is only as defensible as its weakest entry. The signature has moved from the hand to the system. Provenance is how the system signs – and in a plant running this loop without it, that is the only signature the part will ever have.