Here is an uncomfortable observation – the moment a discipline becomes a product category, executives stop building it and start shopping for it. This past week, Boomi launched an Agent Control Plane aimed at the governance gap stalling enterprise agents, Credo AI reported surging demand for the same shelf, and a report out of India put enterprise AI investment up 119% while governance trailed the money. Three headlines, one message: oversight now has a SKU. I have watched this film on a shop floor. Every time a company could buy the instrument instead of building the system, someone in a boardroom confused the purchase order with the capability. The licence arrives on time. The governance never ships.

What a control plane sees and what it can't sign for

I deploy agent systems myself, so I will not pretend the telemetry is worthless. MultiPS, the multi-model orchestration platform I designed and built, runs 63+ models in parallel with consensus synthesis – when the models disagree, the platform surfaces the disagreement rather than letting one confident answer stand. I would not run an agent fleet without it. But my ethical-hacking background trained me to read systems the way an adversary does: hunt the seam between what is monitored and what is owned. A control plane monitors the seam. It does not own it.

The dashboard will give you policy hits, prompt trails, drift flags, spend and latency. All of that answers the morning-after question. The 2 a.m. question is different. When an autonomous agent releases the wrong lot, commits spend with the wrong supplier or quietly rewrites a work instruction nobody reviews until Tuesday, the control plane will render the decision in beautiful colour. It will not tell you who answers for it. Who holds line-stop authority. What happened to the four batches that ran after the decision. Consensus among models is a measurement. Somebody still has to adjudicate what the measurement means, and that somebody is not in the box.

Quality learned this decades ago

Quality management ran this experiment at scale in the 1990s and 2000s, with ring binders instead of APIs. Companies bought QMS software, printed ISO manuals nobody opened and called themselves governed. Auditors were never fooled for long, because an audit never certified the software. It certified the system: roles, competence records, escalation paths, the CAPA log, the discipline of the loop. We called the unread manuals shelfware. Shelfware now ships with a login screen and an annual true-up.

When I built a greenfield QA/QC department for a 900-plus-employee automotive plant, nothing about governance was procured. It was organisation design, line by line: who signs first-off approval, the QRQC cadence that put every failure on a shop-floor board within 24 hours, A3 discipline for the systemic few, the internal audit rhythm, andon authority – any operator could stop the line without asking permission. Governance was a meeting schedule with named owners, a competency matrix and the nerve to stop shipments. Defect costs fell 70%. Customer satisfaction hit 98%. We ran a full quarter without a single critical customer escalation. The software was the cheapest line in that budget; a whiteboard and a competent hierarchy would have carried most of the system.

A dashboard is a witness, not a defence.

Wire it into the NC/CAPA loop or don't buy it

None of this argues against buying a platform. It argues against buying one instead of the system it plugs into. Treat the tool the way a quality organisation treats any new instrument: verify it, then wire it into corrective action. A non-conformity without a CAPA is a scrap report. An alert without a named owner is a fire alarm in an empty building.

Three gates before any purchase order

  • MSA the tool like a gauge. A governance dashboard is a measurement system. Run the equivalent of a gauge R&R: two auditors, the same agent outputs, the same policy toggles. If their verdicts diverge, you are governing noise.
  • Audit agents like critical suppliers. Approved agent list, incoming inspection of outputs against a sampling plan, scorecards, an 8D when they fail, the standing right to de-source. Two decades in automotive and aerospace taught me that grammar under VDA 6.3; an agent that commits spend or releases work deserves no less rigour than a press shop.
  • Price the roles in €, not licences. The subscription is the small number. The real cost is the adjudicator for model disagreements, the escalation owner, the auditor who pulls the thread. If the business case does not carry their salaries, you have not bought governance – you have bought reporting with ambition.

Key takeaways

  • Map the human seats any platform assumes – approver, auditor, escalation owner. Empty seats mean shelfware with telemetry.
  • A governance tool is a gauge: validate it with measurement-system analysis before trusting a single verdict it renders.
  • Autonomous agents are suppliers. Qualify them, sample their output, scorecard them and keep the right to de-source.
  • Budget oversight payroll in € alongside the licence. If the salaries are not in the business case, governance is not either.

The vendors are not lying. Boomi's control plane and Credo AI's dashboards instrument real gaps, and instrumentation is worth paying for. But when the first agent makes the first expensive mistake, the board will not ask which platform witnessed it. They will ask who signed, who escalated, who closed the loop. Governance you can buy is compliance theatre. Governance you build – roles, escalation cadence, CAPA discipline, people paid to say no – survives its first escape. India's 119% says the money is already moving. The organisations that win will be the ones that moved accountability at the same speed.