Who Answers for What Your AI Agents Do
Only 7.2 percent of organisations have a named individual formally accountable for what their AI agents do.
That figure comes from Gravitee's State of AI Agent Security, fielded in April 2026 and published that June, which surveyed 750 senior technology leaders across the UK and US in financial services, healthcare, telecoms, manufacturing and transport. In the December 2025 wave it sat at around one percent. So it is moving. It is moving from almost nothing.
The answers that are not the headline
A single percentage invites you to file it and move on. The distribution is harder to put down.
Accountability unclear or situation-dependent: 32.4 percent. Shared but not formally defined: 29.9 percent. Not discussed or decided at all: 22.9 percent.
Read those three together and the picture is not an industry that has decided to accept a risk. It is an industry that has not yet had the conversation. "Shared but not formally defined" is the one worth sitting with, because it describes a room where everybody believes somebody is covering it.
What is arriving while that stays unresolved
Gartner expects at least 15 percent of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from zero in 2024.
Fifteen percent of ordinary decisions, inside two years, in organisations where nine in ten cannot say who answers for them.
Nothing about that gap closes on its own. Procurement will not close it, because the question is not which vendor you chose. Policy will not close it either, for the reason policy has never closed an execution gap: a document that says who should be accountable is not a record of who was.
A Tuesday, eighteen months from now
An auditor is in a meeting room with a laptop. She has one question, and it is the same question auditors have always had. Who approved this, and on what basis.
The answer involves an agent. It reconciled a set of exceptions, applied a tolerance, and cleared 400 of them. One was wrong, and the consequence has surfaced now rather than then.
Somebody pulls the logs. The logs are excellent. They show the agent authenticated at 09:14 under a service principal, called four endpoints, and wrote 400 rows. They show latency, retries and a clean exit.
What they do not show is which human being accepted that the agent could clear exceptions of that class, what they were looking at when they accepted it, or whether anybody has looked since. The service principal has a name, and it is not a person's name.
Why the machine log was never going to be enough
System logs answer what ran, when, and under whose account. That is what they were built for and they do it well.
They were not built to answer who undertook the work, on what evidence, or who signed off on which version. Those are different facts about the same event, and no amount of log retention turns one into the other. A record of execution is not a record of accountability, and the distance between them is exactly where an audit fails.
There is a second problem, which is quieter. Declared authority drifts from authority in force. An agent is permitted something in a console, a pilot becomes a workflow, a credential gets reused, and the thing downstream of that permission is replaced by something faster. Reading what an agent was permitted to do tells you very little about what it has been doing.
What the missing layer has to do
Three properties, and all three have to hold or the record does not survive contact with a dispute.
The acceptance binds to the act. Before a class of action can run, a named person accepts that class, and the acceptance is recorded against them with what they saw when they made it. Not a setting in a console that somebody may have changed since.
The record is produced by the work. Every exercise of that authority writes its own line, as it happens. Evidence assembled afterwards is the fragile kind, because it is assembled by whoever has an interest in how it reads.
What raises a hand is an event, not a scheduled review. The first use of a class of action. A permission that has gone quiet for a quarter. An agent doing work nobody accepted. A quarterly look finds the first two and misses the third entirely.
That layer has a name. Task Accountability Infrastructure. It is the thing that lets somebody state, in a form a party with no reason to trust you can check, what was undertaken and on whose authority, long after everyone involved has moved on.
Questions
Do system logs show who is accountable for an AI agent?
No. System logs record what ran, when, and under which account. They do not record which person accepted that the agent could take that class of action, what evidence they had, or who signed off on a given version. Those are separate facts, and log retention does not convert one into the other.
How many organisations have a named person accountable for their AI agents?
7.2 percent, according to Gravitee's State of AI Agent Security, fielded April 2026 and published that June, which surveyed 750 senior technology leaders in the UK and US. A further 32.4 percent said accountability was unclear or situation-dependent, 29.9 percent said it was shared but not formally defined, and 22.9 percent had not discussed it.
Is an AI governance policy enough to establish accountability?
A policy states who should be accountable. It is not a record of who was. When a decision is questioned months later, what is needed is evidence of a specific person accepting a specific class of action, and of what happened under that acceptance. A policy document cannot supply either.
What is Task Accountability Infrastructure?
A layer that records who was told, who acknowledged, what they did, and what proves it. It binds an authorisation to the act rather than to a setting, produces the record as the work happens rather than assembling it afterwards, and raises an event on first use or on unaccepted activity rather than waiting for a scheduled review.