Evidence

Practice

• Agents do the work here, and execution advances only once a human has cleared the gate.

  Read in fullCollapse+

  The framework is our own, human-gated, and at work against a live production service. No part of it runs unattended. Constant supervision is the honest description of our role in it, and autonomy is not the claim. It has shipped a handful of changes over a period of weeks, which makes it young, in active use, and not yet a record of anything. Six mechanics decide whether its gate holds, and all six are published on /agentic-routing.

• The gate has failed closed and stopped execution in practice.

  Read in fullCollapse+

  That is the whole of what we publish about the occasion, and the bare fact is the part worth having. A gate nobody has watched refuse is a gate nobody has tested. Fails closed costs nothing to write into a design, and it means something only after a run has actually been left stopped by it. If you are assessing a gate of your own, ask whether anyone can name an occasion when it refused, then ask what happened to the run afterward. Ours refused, and execution stopped there.

• You cannot check any of this.

  Read in fullCollapse+

  The service is not named and the trail is not linked, both by our decision, so nothing in this section is something you can go and confirm for yourself. Discount it accordingly. What we can put in front of you is the mechanism. The six gate mechanics on /agentic-routing are written to be read as a specification and judged on whether a gate built that way would hold, whatever you decide about ours.

Patterns

The first thing we find, nearly every time, is that nobody owns it. Decisions about AI are being made across the organization and no single person is accountable for any of them. That is a governance failure and not a technology one, which is why it opens this list. The four gaps below are the ones that recur underneath it.

1. 01

   No inventory of what is in use

   Nobody can produce a list of the AI tools actually running.

   Read in fullCollapse+

   Every other control is then applied to a set nobody defined, which is how a policy can be carefully written and cover none of what people are using. The inventory is the precondition for the rest of this list.

2. 02

   No decision owner on record

   Approvals do happen.

   Read in fullCollapse+

   They happen informally, by several people, and nothing records who decided. Asked afterward who authorized a particular tool for a particular use, an organization reconstructs an answer instead of producing one, and a reconstruction is not something anyone can stand behind when an auditor or a customer asks.

3. 03

   Data classification not tied to AI

   A classification scheme exists, and it predates AI.

   Read in fullCollapse+

   Nothing in it connects a tier to which models may process data in that tier, so the scheme is real and unused for this purpose. What is missing is the rule attached to each tier.

4. 04

   No AI-specific incident path

   An incident runbook exists, and it has no answer for a bad model output, a leaked prompt, or an agent that acted wrongly.

   Read in fullCollapse+

   None of the three fits the categories the runbook was built around, so all three fall through it. The practical effect is that whoever meets one of them first is deciding the response while it is happening.

• Two beliefs come up reliably at the executive level and both are wrong.

  Read in fullCollapse+

  The first is that their people are not really using AI. They are, and more broadly than leadership estimates. The second is that their people are not using it well. That judgment is wrong too, and it is wrong because it was formed without an inventory of what is in use. It is a verdict on tools leadership cannot list. The two sit together, since a leader who believes usage is narrow has no reason to look hard at how good it is.

• The obvious next heading is the assumptions that turn out to be wrong, and we do not have one to publish.

  Read in fullCollapse+

  The variation across organizations is too wide to generalize from honestly, and past the four gaps above nothing recurs often enough that we would print it as a rule. A fifth pattern would be easy to write and it would be invented. What we can tell you is where the generalization stops, which is here.

Describing client work without attribution is standard practice here, which is why nothing above names an organization or attaches a date to one.