They're deploying autonomous operations. Nobody built autonomous verification.

Four press releases landed in my feed this week. Hitachi and NVIDIA announcing HMAX for "end-to-end autonomous operations through the integrated control of physical AI." Honeywell showcasing agentic workflows for autonomous asset optimisation. IntelliAM launching an agentic AI platform for manufacturing productivity. XMPro named a sample vendor in the Gartner Hype Cycle for Agentic Automation. Four vendors, four claims of autonomy. The word "verification" appears in none of them. That's not an omission. It's the whole problem.

What autonomy actually means

In a quality context, "autonomous" doesn't mean a system that acts without a human. It means a system that self-detects deviations, self-verifies its corrective actions, and self-corrects before the deviation reaches the next operation. Detect, verify, correct. All three. Closed loop. Every time.

The autonomous operations vendors picked up only the third leg. Their definition of autonomy is "the agent decides and acts." Decision authority without verification authority isn't autonomy. It's uncontrolled variation at machine speed.

I've spent twenty years building quality systems across automotive and aerospace. Every framework I've worked under — IATF 16949, AS9100, VDA 6.3 — rests on one non-negotiable principle: you prove the process does what it claims before you let it run. APQP exists because launching an unverified process is a liability. PPAP exists because customers need evidence, not enthusiasm. First Article Inspection exists because "the model thinks it's right" is not a quality record.

Your PPAP assumes a static process

At Airbus, we cut internal lead time by 97% through Routing Verification KPIs — a discipline where every process step has to prove itself before work moves forward. The part doesn't advance until verification is complete and recorded. That compressed the timeline. Not speed. Proof.

Now think about what an autonomous agent does on a manufacturing line. It ingests sensor data, makes a decision, and adjusts process parameters — feed rate, torque, temperature, sequencing — in real time. Every one of those adjustments is a process change. Your PPAP, which took weeks to validate and freeze, assumed those parameters were stable. Your PFMEA assessed risk against a defined process window. The agent steps outside that window because a model decided it should, and your quality system has no mechanism to evaluate that decision before it takes effect.

AS9100 requires control of changes. IATF 16949 requires control of changes. Neither standard has a clause for "an AI agent modified the process 340 times this shift and each decision needs independent validation." The methodology doesn't exist yet. I see BMC talking about governed AI agents in enterprise workflows, and SREs demanding that agents prove themselves before touching production. Those are IT governance models — access control, role-based permissions, audit logging. They govern who the agent is. They don't verify what the agent decided.

Not governance. Verification.

What autonomous verification looks like

I built MultiPS — a multi-model AI orchestration platform running 63 models in parallel — because I don't trust single-model decisions in isolation. One model says increase feed rate by 4%. A second model, independently prompted with the same data, says that adjustment will push tolerance limits at the next station. A third flags a tool-wear correlation the first two missed. Consensus synthesis resolves the tension. The decision either survives scrutiny or it doesn't. Nothing executes until it does.

That same architecture needs to sit between the agent and the physical process. Four components:

• Consensus gating — Agent decisions above a defined risk threshold must pass independent verification before they execute. A second model, a physics simulation, a rule engine. Something with no stake in the agent being right.
• Stop authority — Every agent decision logged, scored, and subject to stop-the-line authority. If the verification gate fails, the line stops. Not the agent. The line. I have shut down production for a single unverified dimension — I will not accept lower standards from a machine.
• Routing verification for AI outputs — The KPI discipline I used at Airbus to cut lead time by 97% needs to wrap every agent action. Agent proposes a parameter change. Verification confirms it sits within the validated process window. Only then does the change propagate downstream.
• Pattern drift detection — A real-time monitor that flags when agent decisions cluster outside historical norms. Not because each individual decision is wrong, but because patterns reveal what single decisions don't — model drift, sensor degradation, an adversarial condition.

None of this is exotic. It's what a competent quality engineer does instinctively. Question the result. Verify independently. Escalate when something doesn't add up. The technology industry has spent millions building agents that act and hasn't built the layer that gates them.

Autonomy without verification is just another word for a quality incident you haven't detected yet.

Deploy the agents. I'm not arguing against the technology — I build AI systems, I deploy multi-agent architectures, I believe in the trajectory. But understand what you're shipping when you ship autonomy without verification. You're pre-authorising a quality incident every shift and betting the model gets it right more often than it doesn't. In aerospace, in automotive, in any regulated manufacturing environment, "the agent decided it" is not a defence I want to bring into an EASA audit or a customer escalation. Build the verification layer first. Then let the agents run.