Here is an uncomfortable observation: the AI governance industry has just discovered the quality gate and is selling it back to us as innovation. In the past fortnight, OneTrust announced a platform overhaul aimed at AI governance at enterprise scale, Tokio Marine brought in Monitaur to operationalise its programme, and the gateway camp is pitching "one gateway to govern them all" – a single checkpoint every model call must clear before it runs. Their phrase for it is governance at the moment before execution. We had a name for that on the shop floor: andon. Before that, jidoka. End-of-line inspection was discredited in Western manufacturing a generation ago – sorting defects after the money was spent. The AI industry has replayed the whole arc, logs reviewed after the action, dashboards of yesterday's mistakes, and has now arrived at interception.
What a gate actually is: an interlock, not a document
A gate is not a meeting. Not a sign-off, not a checklist, not a compliance register. A gate is an interlock: a mechanism that stops flow when a defined condition goes red, and stops it without asking anyone's permission. Everything else is commentary.
I learned this the expensive way at Witte Automotive. Our Q-Wall began life as a board – red and green squares, hourly review, escalation paths on paper. It informed; it did not govern. Line leaders treated it like a weather forecast: noted, rarely acted on. What changed was giving the wall teeth. A red condition stopped flow to the next operation until a named owner picked it up, and QRQC discipline attached a 24-hour clock to every trip. Same wall, same squares – different physics. Failure costs came down substantially, and they came down in the areas the wall could stop, not the ones it merely reported on.
Years later I built a greenfield QA/QC department at SNOP – 900-plus people from an empty shell – and we drew the gates before we hired headcount: where flow would stop, what red looked like, who owned the first hour of every trip. The 70% defect-cost reduction and 98% customer satisfaction that followed get cited as a quality result. They were a gate-design result.
The inspection trap, replayed
Most of what is sold as AI governance today is end-of-pipe inspection. The agent acts, the logs go to a review board, someone produces a monthly report on incidents that already happened. In a plant that is final inspection with better typography. The parameter has changed, the order has shipped, the supplier has been paid. You are not controlling the process; you are cataloguing its exhaust.
The counter-move the vendors are now discovering is default-deny before execution. In security – I hold a CEH among thirty-odd certifications – this is the oldest interlock in the book. Nobody serious audits packets after they have left the building; you deny at the boundary. A firewall is a poka-yoke for packets. Inspect at the edge, refuse what fails policy, fail closed. Extending that to agents is not a new discipline; it is an old one pointed at a new pipe. The physics of interlocks does not care whether the flow is steel or tokens.
Meanwhile the agents keep arriving on the floor. Siemens and TSMC run them inside semiconductor design checks. Regal Rexnord has gone all-in on an agent platform. MongoDB and Palo Alto ship the plumbing. Each of those agents will eventually hold hands with systems of record – releasing lots, writing to the ERP, changing parameters. The governance question is not whether to watch them. It is where the interlock sits.
Where AI gates fail first
AI gates fail at two predictable points. The first: they intercept the model call, not the tool action. Gateways parse prompts and responses for policy violations, but the damage rarely lives in text. It lives in the function call – the write to the ERP, the lot release, the parameter change, the message to a supplier. A gate on the model call is a checkpoint on an operator's mouth while his hands go unwatched. Put the interlock on the action, at the boundary where the agent touches a system of record, and make it fail closed.
The second failure has no vendor, because it cannot be licensed: what happens when the gate trips. The gate blocks a lot release at 02:00 – and before it ever trips, you must be able to state what stops, who is woken, who owns containment and by when. IATF 16949 has required a documented reaction plan for every control in the control plan for years. The standards writers understood that a control without a reaction is a half-built control. At Witte, the quarter in which every Q-Wall trip carried a named owner and a 24-hour clock was the quarter we recorded zero critical customer escalations. I will not credit the wall for that. The wall tripped; the reaction plan held. That is what de-escalation actually looks like – a gate that closes and a team that moves.
A gate with no reaction plan is not a control – it is an alarm clock in an empty house.
Key takeaways
- Put the interlock on the tool action – the ERP write, the lot release, the parameter change – not only on the model call. Text is not where the damage lives.
- Before you buy a gateway, define the trip: named owner, containment scope, 24-hour clock. If you cannot name the owner before the gate trips, you do not have a gate.
- Borrow firewall discipline wholesale: default-deny, fail closed, no bypass without escalation. A poka-yoke for packets is a poka-yoke for agents.
- Drill the reaction, not the dashboard. Trip a gate deliberately every quarter and measure time-to-owner and time-to-containment, the way you run a fire drill.
The gateway is the cheap part – a licence and an architecture diagram. The expensive assets are stop authority, containment discipline and a feedback loop that turns every trip into a countermeasure: QRQC and 8D with newer plumbing. Plants that learned stop-the-line a generation ago already own all three. They hold a two-decade head start on agentic AI. Plants that never did will buy the gateways anyway, and the gateways will trip into a void – technically compliant, beautifully logged, stopping nothing that anyone is accountable for.