AI accountability is falling behind AI authority. Forty per cent of companies now let AI execute decisions without human approval, at least to a moderate degree. Only 15 per cent report a significant shift in decision-making authority for their own people. Read the two figures side by side and the direction of travel is obvious: the workflow is acquiring authority faster than anyone is redesigning who holds it.
Both figures come from Deloitte’s State of AI in the Enterprise report, the global professional services network’s survey published 29 October 2026, covering 3,500 director-to-C-suite leaders across 25 countries and six industries. The same report found that 60 per cent of companies have redesigned jobs around AI capabilities to at least a moderate degree, and that 23 per cent cannot say who is accountable when an AI system produces a negative outcome.
A separate study backs the same pattern. IBM, the technology and consulting company, surveyed 3,221 executives across Europe and the Middle East and found that when asked who is accountable when an AI agent causes harm, the most commonly cited answer, the CIO or CTO, was named by only 17 per cent of respondents.
Decision Rights Are Not the Same as Accountability
A capability and a right are different things, and most organisations are granting the first while assuming the second follows automatically. An AI system that can approve a refund, route a support ticket, or release a payment has been given a decision right the moment it goes live. Nobody votes on that separately. It arrives bundled with the deployment, tested for accuracy and throughput, rarely tested for who answers for the outcome.
Compare that with how the same organisation hands authority to a person. A new programme director does not get a budget line, a hiring mandate, or a sign-off limit on day one. A governance board agrees it, usually after weeks of drafting, review, and a named limit attached to a named role. The AI workflow that approves a thousand transactions an hour rarely gets the same scrutiny in a single meeting, let alone weeks of it.
The gap shows up fastest in the AI workflows nobody thinks of as decisions at all. A credit line adjustment, a fraud flag released without review, a support ticket closed with a refund attached: each one looks like routine processing until the outcome is wrong, and then it is indistinguishable from a decision a person would have needed sign-off to make.
That same ambiguity recently surfaced in public. When one large employer tied a round of job cuts to both agility and AI readiness, its CEO and CFO gave two different public reasons for the same restructuring on the same day. Nobody had agreed in advance which answer was the real one, which is exactly the kind of gap a working AI accountability register is meant to close.
A survey of enterprise risk, compliance, and security leaders puts a number on how often that gap bites: 34 per cent of organisations have already acted on an inaccurate decision or output an AI agent produced, and only 9 per cent have redesigned their workflows to manage that risk, despite 96 per cent agreeing that workflows need to change.
The AI Accountability Register Nobody Is Keeping
Every organisation with functioning governance already keeps a delegation of authority schedule: who can approve what, up to what value, and under whose name. Few of them have added a line for AI. The workflow sits outside the schedule that exists specifically to answer the question it raises.
The fix needs no new governance machinery, just an extra tab on the one that already exists. List every autonomous AI action next to the human delegations already there: the action it is permitted to take, the value or consequence limit it operates within, the named individual who owns the outcome, and a date by which that delegation is reviewed. A single row might read: claims-triage workflow, approves claims under $2,000 without review, owner the head of claims operations, reviewed quarterly. Nothing in that row demands new technology, just someone prepared to put their name in the owner column. A workflow with no name in that column has been granted a right nobody actually signed for.
What to Ask Before the Next AI Workflow Goes Live
Three questions anchor AI accountability before any workflow goes live, and they catch most of what the Deloitte figures are pointing at. What decision is this system actually allowed to make, stated as plainly as a human delegation would be? What is the ceiling on it, in value, in volume, or in the type of case it can touch? And who, specifically, signs their name against the outcome if it goes wrong, rather than leaving the answer to the vendor or to the platform team by default?
None of those questions need new software, a new committee, or outside consultants. They need a willingness to put AI decisions through the same paperwork a person would have to clear, and a refusal to let a capability keep operating once it fails that test.
Deloitte’s own number for how often that test currently gets applied is 23 per cent unclear, meaning roughly a quarter of organisations running autonomous AI have no answer to the accountability question when asked directly.
Pick one AI workflow in your organisation that currently acts without approval. Find the register it belongs in and check whether a name is written next to it. If nobody signed, the workflow does not have a decision right. It has an unclaimed one.
