Ask a leadership team who owns the AI decisions in their organisation and you’ll get a confident answer. The CTO. The head of data. The steering group. Ask a level or two down and you’ll get a different answer, or several, or a pause followed by a name nobody at the top mentioned.
That gap is the whole problem. Decision rights in most organisations are clear in theory, vague in practice, and untested until something needs deciding quickly.
It isn’t an org chart question
People treat it like one, but an org chart describes hierarchy, not decision rights. Decision rights are more specific: who can say yes, who can say no, who gets consulted, and who finds out afterwards.
In well-run organisations these things are mostly written down and mostly followed. In most organisations they’re partly written down, inconsistently followed, and heavily propped up by the people who’ve been around long enough to know you need so-and-so in the room before anything actually moves.
That informal layer is fragile. It doesn’t survive a reorganisation. It doesn’t scale. And it’s invisible to an AI system, which only works with what’s explicit.
Where AI exposes it
AI operates fast. That’s the entire pitch. A model produces a recommendation in seconds. If acting on that recommendation needs a sign-off that takes six weeks, you haven’t sped anything up. You’ve just added a faster front end to a slow machine.
I’ve watched AI initiatives stall not because the output was bad but because nobody could agree who was allowed to act on it. The recommendation was there. It was good. It sat untouched, because acting on it needed a decision that wasn’t clearly anyone’s to make.
The speed of the tool makes the slowness of the organisation obvious in a way it wasn’t before. When everything was slow, a six-week decision looked normal. Next to a model that answers instantly, it looks like what it is.
What I look at
When I assess decision-making as part of a review, a few things tell me most of what I need. Whether accountability is explicit or assumed. Whether the decision process matches the speed the AI is supposed to run at. And what happens when a delegated decision turns out to be wrong, who owns it, who fixes it, what the escalation path is.
That last one is the most revealing. In organisations with clear decision rights, the answer is specific. In organisations without them, it’s “it depends” or “we’d work it out,” which means nobody has.
For firms where the partners are the decision-makers, this shows up differently but no less sharply. Decisions get made by whoever’s closest to the client, or by informal consensus that’s held up fine until now. AI changes the maths, because the speed and the client-facing stakes both go up. Who reviews the AI output before it reaches a client? Who’s accountable when it’s wrong? Those need answers before the tool is in the workflow, not after.