Your IT and OT disagree on every shift. Your quality system trusts both.

I have stood in enough war rooms to recognise the pattern. A customer SCAR lands Monday morning, citing dimensional drift on parts received Thursday. The quality team pulls the QMS records — everything green, Cpk above 1.33, sign-offs clean. Then someone walks to the maintenance office and pulls the PLC log. At 14:32 on the shift in question, the injection barrel temperature on station 3 drifted eleven degrees above the upper control limit for twenty-three minutes before the heater band corrected. Nobody flagged it. The ERP recorded the batch as conforming at 15:10. The thermal excursion never crossed into the quality record because the system that caught it and the system that signed off the parts had never been introduced to each other.

This is not a technology problem. It is a governance problem disguised as an architecture problem. And every manufacturer now rushing to deploy physical AI — autonomous inspection, self-adjusting process controls, agent-driven scheduling — is about to hit it at speed.

What dies at the seam

The IT/OT boundary in most plants I have audited is not a clean interface. It is a graveyard. Process parameters, equipment state transitions, torque curves, thermal profiles, vibration signatures, cycle-time deviations — all of it lives in the OT layer, captured at millisecond resolution by PLCs and SCADA systems that were never designed to handshake with an ERP running batch reconciliations on a six-minute polling cycle. The data that matters most for understanding why a part conformed or didn't is generated on one side of the wall and consumed on the other. Everything in between gets stripped, summarised, or silently dropped.

The current wave of physical AI deployments — autonomous quality assurance, agentic scheduling, self-correcting process controls — all assume a data continuity that most plants simply do not have. You can deploy the most capable vision system money can buy. If the OT context that explains a defect never reaches the IT system that routes the corrective action, your autonomous loop is ornamental.

Physical AI needs one clock, not two stories

Autonomous manufacturing systems cannot tolerate the ambiguity that human supervisors quietly absorb every shift. When an operator sees a temperature warning and a subsequent dimensional check pass, they make a judgement call — often correctly, often from experience the system does not possess. An autonomous agent faces a harder problem. It receives the IT signal that the part passed and the OT signal that the process window was violated. Without reconciliation rules, it either halts the line for every anomaly or proceeds on whichever signal arrived last. Neither outcome is acceptable in a production environment.

If your IT and OT systems disagree about what happened on your line, your autonomous agent will inherit the worst instincts of both.

The vendors selling physical AI platforms are not solving this. They are solving the compute problem, the vision problem, the inference latency problem. The reconciliation problem — deciding which system holds the truth when the PLC logged a drift but the QMS logged a pass — is a governance decision. It belongs to the quality function, not the integration contractor.

Reconciliation is governance, not middleware

At Airbus, the routing verification KPI system I implemented reduced internal lead time by 97%. Not because we bought a better integration platform. We didn't. We forced every critical process parameter captured in the OT layer to reconcile against the quality record in real time. When the two disagreed, the system escalated to a named owner with the authority to hold the batch. That sounds straightforward. It required defining, for every critical-to-quality characteristic, which system was authoritative and under what conditions the OT signal overrode the IT record. That mapping did not exist before we built it. It does not exist in most plants today.

I know this because I have asked. The typical answer: IT owns the quality record, OT owns the process data, and they meet once a month in a review meeting to argue about discrepancies. That is a ritual, not a control point.

When I built the quality organisation at SNOP — a 900-employee greenfield plant — I architected the IT/OT data contract before the first machine was commissioned. Every critical process parameter had an owner, a reconciliation rule, and an escalation trigger before it produced a single part. Retrofitting this into an existing plant costs three to five times more and takes twice as long, because by then every department has a vested interest in the old ambiguity. Greenfield is cheaper because nobody has defended the legacy yet.

The seam as a critical control point

Treat the IT/OT boundary the way you treat a special characteristic on a PFMEA. Identify the failure mode. Define the detection method. Assign severity. Name an owner. When the systems disagree, someone with quality authority gets paged — not emailed. The reconciliation rule cannot be "we'll investigate during the next shift and circle back." That is how a thermal drift at 14:32 becomes a customer SCAR the following week.

Physical AI does not create the IT/OT divide. It exposes it at a speed that human workarounds cannot absorb. The manufacturers pretending the gap doesn't exist are building autonomous systems on top of two networks that cannot agree on what just happened on their own line — and handing those systems the authority to act on the disagreement. The quality system that trusts both has already failed. The only question is whether you define the reconciliation rule now, or wait for the SCAR to define it for you.