Nobody wrote down who decides
By Dino Nokic · AUG 2026 · 4 MIN READ
IBM's Institute for Business Value asked executives a question most AI research skips. Not whether the technology works — whether anyone is allowed to overrule it. Seven in ten reported unclear decision rights and accountability in AI-enabled work, and 70% said that ambiguity has already caused problems in their AI projects.
That is not a technology finding. That's an org chart finding, and it explains more failed deployments than any benchmark does.
I've written before about the seam nobody owns — the handoff that belongs to both teams and therefore to neither. This is the same failure one floor up. When a system starts producing recommendations at machine speed, "who decides" stops being a philosophical question and becomes an operational one, several hundred times a day, at desks where nobody has been told the answer.
The detail that should bother you most: 63% of executives say AI has already shifted who makes key decisions, and nearly two-thirds say it's reshaping roles and workflows. So the authority moved. It just moved without anyone writing it down.
What ambiguity does to the people underneath it
The supporting numbers are worse than the headline. 43% say employees don't feel safe raising concerns. More than a third report no clear escalation path when an AI output looks risky. And 34% of organizations have no repeatable process for the moment that matters most — when a person and the system disagree.
Put those together and you get exactly the behavior you'd predict. When nobody knows who owns the call, and challenging the machine feels like a career risk, people defer. They stop arguing. And deference reads as adoption on a dashboard — usage up, overrides down, everything green. What it actually is, is an operation that quietly stopped checking its own work.
That's the failure mode I keep coming back to. Not the agent that breaks loudly. The agent that's right often enough to be trusted completely, inside a company where nobody was ever given standing to say no.
The measurement problem nobody planned for
93% of executives say AI-enabled work makes it significantly harder to evaluate contribution and accountability, and half of employees say it's harder to get fair credit for what they did. Meanwhile 65% report having guidelines for AI-enabled work — and the same share say those guidelines are already out of date. Among technology leaders, 77% say adoption is moving faster than their governance can follow.
Most companies are treating that as a policy backlog to be caught up on later. It isn't a backlog. It's the predictable result of deploying capability faster than you deploy authority, and no amount of catching up fixes it, because the gap regenerates every time you ship something new.
What the 15% do differently
IBM's split is the number worth taking to a board. Only 15% of organizations combine advanced AI maturity with strong change execution — and those report up to 73% higher revenue growth and an 11% improvement in operating margin than their peers. Same models. Same vendors. Different ground underneath.
Their framing is permission, practice, and proof, and it maps cleanly onto how autonomy should actually be earned.
Permission is deciding in advance — in writing, before deployment — when the system is followed, when it's questioned, and when it's overridden, plus who owns the call when human judgment and the output diverge. An escalation path that exists on a slide is not an escalation path.
Practice is treating judgment as a skill rather than an assumption. Only 27% of employees say their managers focus mainly on coaching interpretation and judgment — which is the one thing a manager will be needed for once the routine work is handled. Judgment is a bottleneck you can widen, but only if somebody is actually widening it.
Proof is the part almost everyone gets backwards. 81% of executives say their organizations reward AI-related skills. Far fewer say they recognize the person who pushed back on a wrong output and turned out to be right. You get the behavior you reward. Reward speed and adoption on their own, and you will get speed and adoption — including on the days the system is confidently wrong.
None of this is an argument for slowing down. It's an argument that the human gate isn't a training-wheel phase you remove once the tool proves itself. It's the load-bearing part. Write down who decides, before the machine starts deciding by default.
The technology arrived on a schedule. The authority to overrule it never got scheduled at all.