Ask a leadership team who owns the AI decisions in their organisation and you’ll usually get a confident answer. The CTO. The Head of Data. The AI Centre of Excellence.
Ask the same question a level or two down and you’ll get a different answer. Or several different answers. Or a long pause followed by a name nobody at the top mentioned.
This is decision rights in most organisations: clear in theory, ambiguous in practice, and entirely untested until something goes wrong.
What decision rights actually means
It’s not an org chart question, though people usually treat it like one. Org charts describe hierarchy. Decision rights describe something more specific: who can say yes, who can say no, who gets consulted, and who needs to be told after the fact.
In well-functioning organisations these things are mostly documented and mostly followed. In most organisations they’re partially documented, inconsistently followed, and heavily supplemented by institutional memory — the people who’ve been around long enough to know that you actually need to get so-and-so in the room before anything moves.
That informal layer is fragile. It doesn’t survive reorganisations. It doesn’t scale. And it is completely invisible to an AI system.
Why AI makes this worse
AI doesn’t navigate informal decision-making. It operates on whatever is explicit, documented, and structured. If your decision rights are in someone’s head rather than written down, the AI doesn’t have access to them. If the criteria for a decision shift depending on context that’s never been codified, the AI will apply the wrong criteria. If accountability for an output is genuinely unclear, nobody will pick it up when it’s wrong.
I’ve seen AI initiatives stall not because the model was bad, but because nobody could agree on who was allowed to act on what it produced. The output existed. It was good. And it sat there, because acting on it required a decision that wasn’t anyone’s job to make.
What I actually look for
When I assess decision rights as part of a diagnostic, I’m looking at a few things. Whether accountability is explicit and not just assumed. Whether the decision-making process for AI outputs matches the speed AI is supposed to operate at — there’s not much point in real-time AI recommendations if sign-off takes two weeks. Whether there’s a clear answer to what happens when the model is wrong, including who owns it, who fixes it, and what the escalation path is.
The last one is the most revealing. In organisations with clear decision rights, the answer to “what happens when the AI is wrong” is specific and practical. In organisations where decision rights are murky, the answer is usually “it depends” or “we’d figure it out” — which means nobody’s actually thought about it.
The consultancy-specific version
For boutique consultancies, this problem shows up differently than in larger organisations. There usually isn’t a formal governance structure, because there hasn’t needed to be one. Decisions get made by the founder, or by whoever’s closest to the client, or through an informal consensus that’s worked fine until now.
AI changes the calculation. The consultancy is using AI to support client work, which means the quality and accountability of AI outputs is a client-facing problem, not just an internal one. Who reviews the AI output before it goes to a client? Who’s responsible when it’s wrong? These questions need answers before the AI is in the workflow, not after.
The organisations that get this right tend to have documented it before they needed it. Not because they expected problems, but because writing it down is how you find out whether your assumptions about who decides what are actually shared.