Here is an uncomfortable observation — the cheapest audit in the world just got cheaper. The regulator no longer needs to understand your AI. They ask for the policy and read you through it. PharmTech's pulse check this week reported that AI policies have become regulators' inspection yardstick, and when one regulator publishes a habit, every other regulator inherits it for free. The same week, AWS put out an engineering guide on context layers for production agents — four layers, in working code — confirming what anyone actually shipping agents already knew: the prompt and context stack is the operating layer. The auditor reads the paper. The engineer lives in the repository. Those two have never met, and one of them is about to be graded on the other.

What the auditor actually reads — and what a template confesses

Two decades of surveillance audits under IATF 16949 and AS9100/EN 9100 taught me one mechanical truth: auditors do not audit technology. They audit governance artifacts — ownership, traceability, change control. When I built the QA/QC function for a 900+ employee greenfield plant, the quality manual was approved and under revision control before the first part shipped. Not because auditors love binders. Because the manual is the index they use to decide which other drawers to open.

A template AI policy inverts that logic. Every clause is a promise the auditor can convert into evidence on demand. The policy says prompts are reviewed and approved — show me the review records. It says a human validates high-risk outputs — show me the log, the names, the training. A €3,000 policy copied from a vendor template, or drafted by procurement to pass a supplier questionnaire, is a confession with a cover page. It lists controls you do not have, signed by someone who will not be in the room when the auditor asks to see them.

The prompt is a control plan nobody controls

The AWS guide walks four context layers: instructions, retrieved knowledge, tool definitions, memory. Translate that into plant language. The system prompt decides what the agent sees, what it escalates, what it silently ignores — the same judgment a PFMEA encodes and a control plan operationalises. Tool definitions are work instructions. The retrieval layer is the agent's spec library, with all the obsolete-document risk that implies.

Now check where that artifact lives. A repository the process engineer treats as a scratchpad, or a chat thread. No revision number, no approver, no validation protocol, no drift check. We would stop a line for running an uncontrolled control plan. We let an agent that reads our work instructions, our 8D history and our customer complaints run on one every day.

My security work taught me the same lesson from the opposite direction. Certified Ethical Hacker, thirty-plus security certifications, responsible-disclosure work that put me on T-Mobile's public bug-bounty Hall of Fame — that trade is documenting how a system behaves under adversarial conditions, which is all an AI policy really is. Attackers do not read your policy. They read your system. A prompt is attack surface, and every context layer is another injection point: a poisoned document in retrieval, a tool description that talks the agent out of its guardrails. When I deploy multi-agent systems — the orchestration platform I built runs 63+ models in parallel in private beta and synthesises consensus — I design the prompt and context stack like a control plan: what the agent may never act on alone, which disagreements between models put an answer in quarantine. Consensus catches drift the way a quality wall catches a suspect lot. None of that discipline appears in the policy, because the policy was written before anyone had deployed anything.

The policy is the promise you made the auditor. The prompt is the promise you made the line. Only one of them touches product.

Put the agent in the document tree

The fix is boring, which is how you know it will hold. Prompts and context configurations go under the same change control as control plans: revision number, named approver, validation record, effectivity date. Escalation is defined the way andon is defined on a line — trigger, responder, timeout, and the class of decisions the agent may never take alone. A configuration audit pulls the live system prompt from the deployed agent and diffs it against the approved revision, exactly as you would audit a control plan against what is actually clamped on the machine. Divergence is a QRQC, not a backlog item: same-day owner, containment, countermeasure, closed loop.

The pressure is arriving from both ends. Agentic AI is moving past demos into operations — SAP's Asia-Pacific partners are wiring agents into end-to-end workflows, Hitachi is publishing on making agents behave in the physical world — so prompt logic is about to touch product, and AS9100's configuration-management clauses already reach software that affects product conformity. Meanwhile IATF, AS9100 and VDA 6.3 auditors travel the same conference circuit as the pharma inspectors. An inspection technique that works in Amsterdam or Silver Spring lands in a Detroit or Stuttgart audit room within two cycles. I have watched softer methods cross industries faster than that. A policy written by someone who has never deployed an agent is a novel with a signature block — and the auditor who finds the divergence will quote it back to you in the finding.

Key takeaways

  • Treat every deployed system prompt and context configuration as a controlled document — revision, approver, validation record, effectivity date — with no exception for "a demo that grew".
  • Define agent escalation like andon: trigger, named responder, timeout, and the decisions the agent must never take alone.
  • Run a configuration audit that diffs the deployed prompt stack against the approved revision — quarterly, and after every model or vendor update.
  • When the diff shows divergence, open a QRQC with same-day containment and a named owner. Divergence between policy and deployment is a nonconformity, not a ticket.

Your AI policy just became an audited quality record while the artifact it governs still ships from an uncontrolled repository. That asymmetry is the finding. It will be written in our own language — document control, configuration management, reaction to nonconformity — which is why none of us will be able to argue with it. The regulator has chosen its yardstick and the engineering literature has confirmed the operating layer. Close the gap while it is still a paragraph in a quality manual, not a customer escalation with a containment bill attached.