Most bad decisions are not failures of intelligence. They are failures of attention. The question no one thought to ask. The default no one chose. The decision that got made by absence. A framework is just a way to make attention reliable, to put the important question in front of you before the moment has passed rather than after the cost has landed.
What follows is not theory. It is the thinking I use, stripped to what you can apply on your own. The deeper, tailored versions of this work live in what I do with organizations, but you do not need those to get most of the value. Take these, adapt them, make them yours.
Everything else rests on five reframes. They are not controls or procedures; they are the way of seeing that makes controls obvious later. Get them wrong and every tool choice, policy, and review inherits the error. Get them right and most of what looks like an AI problem turns out to be a decision problem you already know how to handle.
Most AI governance goes looking in the wrong place. It asks which tools are approved, which models are allowed, which prompts are permitted. But a tool no one has acted on has produced neither value nor risk. Both appear at a single moment: when an output stops being a draft and becomes the basis for a decision someone carries out. That decision is the thing to govern. You can allow every tool in the building. What matters is what happens when the output leaves the screen and shapes a choice.
The instinct is to manage AI by how often it is wrong. That is the wrong axis. A model can be right ninety-nine times, and the hundredth, a client-facing number, a legal position, a pricing call, can cost more than the ninety-nine saved. Frequency is comforting and mostly irrelevant. Consequence is the axis that matters. Put your attention where a single error is expensive, not where errors are common. Most of them are cheap. A few are not, and those few are the whole game.
If you have not decided how AI influences your decisions, it is tempting to believe you have stayed neutral. You have not. You are running a system already: the tool's default settings, the habits of whoever happens to be using it, the judgment of the person under deadline. That is a system, just an unexamined one, and it was designed by no one. The only real choice is whether the system that governs your decisions is one you chose, or one that got installed while you were not looking.
AI errors do not sit politely where they are made. A misclassification becomes a line in a report. The report becomes a recommendation. The recommendation becomes a decision, and by then the original mistake is three steps away and wearing a suit. This is why checking the tool is not enough. By the time an error is visible in the outcome, it has already moved through everything in between. You catch it at the decision, where it can still be stopped, or you catch it afterward, in the cost.
A policy cannot be held accountable. A document does not answer the phone when something goes wrong. Only a person does. For any decision that carries real consequence, a name has to be attached before it executes: not a committee, not a function, a person who owns the outcome and has the authority to change it. Ownership assigned after the fact is not ownership, it is blame. The decision that belongs to everyone belongs to no one.
None of this requires new software. It requires seeing AI where it actually operates, at the decision, and refusing to let the consequential ones happen by default.
Download the one-page reference (PDF)Put the third principle to work. The Absence Audit finds the decisions your organization is already making by not making them. It takes about an hour and a willingness to hear uncomfortable answers. Run it on any process where AI now touches the outcome, and ask five questions in order.
You finish with a short list of decisions your organization was making silently, each now carrying a name, a cost, and an owner. That list is worth more than most AI policies, because it is specific, and it points at exactly what to fix first.
Download the fillable worksheet (PDF)They map straight onto the Audit. The Record gives every consequential decision an owner. The checklist stops the next default from being installed in the first place. Copy them as they are.
One entry per AI-influenced decision that carries consequence.
Run this before rolling out any AI tool.
This is the personal habit under all of it, the one behind "someone has to ask." Good decisions are rarely lost on the hard questions everyone is already debating; those get answered. They are lost on the one question no one thought to raise. This is a way to find it on purpose, before the decision instead of after.