AutoOps AI

Every quality method I have used in a plant assumes the evidence is available. 5-Why assumes you can see what changed. FMEA assumes you know the historical failure modes. CAPA effectiveness checks assume you can find out whether the defect recurred. In practice that evidence is scattered across an ERP system, a MES, a dozen spreadsheets, inspection PDFs, supplier certificates and shift notes.

The consequence is not that engineers use bad methods. It is that they spend the majority of an investigation assembling context, and then run a good method on a thin slice of the evidence, because that was the slice they could reach before the meeting. The method is fine. The retrieval is the bottleneck.

Manufacturing teams do not primarily need another dashboard. They need a reasoning layer over their own operational data — one that can assemble the context an RCA, an FMEA or a CAPA actually requires, and cite where every claim came from.

The citation requirement is not a nicety. A quality conclusion that cannot be traced to a record is not usable in a quality system, so anything generated has to be attributable back to source data or it fails the only test that matters.

Three findings changed the product:

Nobody wanted a replacement quality system. They had one, usually mandated. What they wanted was something that made the one they had usable — which moved the product from “system of record” to “layer on top of the records.”

Output without provenance was worthless. An unsourced answer, no matter how correct, created work instead of removing it, because someone still had to go verify it. Citation moved from a feature to a constraint on the whole design.

The ERP integration was the hard requirement for organizations, and a blocker for individuals. That split produced the two-track structure: a complete quality toolkit that works on a team’s own uploaded data, and an organizational tier that adds the ERP layer.

A 20-user pilot at a manufacturing company. The pilot was scoped deliberately around the quality workflows rather than the chat surface, because the workflows are where a wrong answer has a consequence and therefore where the design had to be proven.

What held up: structured records with provenance, and the escalation sweep — the least glamorous feature and the one that changed behaviour, because it made aging items visible without anyone having to remember to look.

What did not: my first version assumed people would ask open questions. They mostly did not. They wanted the system to produce the artifact they already owed someone — a report, a CAPA, an FMEA row — which is a much narrower and much more useful product.

Manufacturing quality depends as much on information architecture as on analytical technique. The methods are well established and mostly taught correctly. The reason investigations are shallow is that the evidence is expensive to assemble, and that is an architecture problem.

Traceability is a product requirement, not a compliance checkbox. An answer that cannot be sourced cannot be used in a quality decision, so citation had to be designed into the data model rather than added to the output.

The unglamorous feature won. A scheduled sweep that makes aging items visible changed more behaviour than anything generative, which is consistent with everything I have seen on a shop floor: the control that runs without being remembered is the one that works.

Building the product changed how I do the day job. Having to model nonconformance, CAPA and FMEA as data forced precision about what those records actually require — which is now how I think about closure evidence in a plant.