The risk register missed the risk that mattered

Meta description: AI risk lives off the register, in unowned handoffs and verification gaps — the places where a consequential choice has no name on it.

The AI risk programs I encounter are serious documents. Model inventories, validation protocols, bias testing, usage policies — the architecture of caution, properly staffed. And in the versions of this story that matter, the harm arrived from outside the register entirely: a workflow that drifted, a decision nobody noticed the model had joined, a customer outcome that changed with no owner watching. The company was not careless. It was looking carefully, in a place where the risk no longer lived.

[AWAITING MIKLOS — scene: a model-risk framework in force while a quiet workflow is already changing customer outcomes with no owner]

What do AI risk programs typically inventory?

The inventory is a list of models. Each model gets a risk rating, a validation date, an owner from the model-risk team. This is the correct architecture for the risk of a decade ago, when the dangerous thing was a model making a regulated decision it should not make. I read these registers and notice what the inventory form cannot hold: the risk is in the model column, and the harm I keep seeing lives in the workflow column, which the form does not have.

Why does that inventory feel complete?

Miklos Roth, high-velocity AI consultant

The harms I would watch for sit outside the model inventory: mistakes that pass through because no one owns the end-to-end result, corrections that remain private to the person fixing the draft, outputs that look polished and are therefore trusted without examination. A locally successful operation does not remove the need for someone who understands the work. When verification is treated as a final ritual rather than part of the system, the register can be green while the repair work grows unseen.

The inventory feels complete because it is complete — of models. Every model the company knows about is listed. What I have learned to notice is the assumption inside the completeness: that risk attaches to models. In the failures I watch, risk attached to arrangements — a model plus a workflow plus an absent owner plus a team that trusted the output because a framework existed. No single element was off-register. The arrangement was, because registers count things, and arrangements are not things.

Where does harm show up off-register?

Read more: Miklos Roth’s profile.

Off-register harm has a consistent character in the stories I have seen. It is slow. No incident, no breach, no headline — a customer segment treated slightly differently for a year, a class of decisions drifting away from policy, a quality of judgment declining in a team that stopped exercising it. The harm is real and it is legible to the people inside the workflow, and it never generates a risk event, because the register watches models and the workflow was never registered as anything.

What pattern separates governance theater from the live risk?

The false explanation is that the company needs more policy — a thicker framework, broader coverage, a new committee. I have watched policy thickness double while the live risk sat in the same unowned workflow. Governance theater has a tell: it produces artifacts that are impressive to auditors and invisible to operators. The live risk, meanwhile, is visible to operators and unimpressive to auditors — a workaround, a habit, a quiet reliance. The two never meet, because they live in different documents.

What I require to be named is the person who owns the consequential choice. A business owner or specialist develops an integrated understanding through years of work; an AI assistant can provide valuable information about components of that picture, but it does not possess the company’s tacit knowledge, its relationships, its constraints, or accountability for the outcome. AI may do substantial preparatory work. Qualified people must interpret results and own consequential business choices.

[AWAITING MIKLOS — scene: second scene]

What question moves the search?

Risk committees in this pattern ask thorough questions about models and no questions about the space between models and customers. The space is where the decisions live. There is a question that relocates the committee’s attention into that space, and I have noticed it tends to arrive from outside the risk function, late, after something has already gone quietly wrong.

The question risk committees are not asking is the one with a name in the answer: for each consequential AI-assisted outcome, who — by name — interpreted the result, accepted it, and owns what happens next?