Who owns AI governance in an enterprise?

Why 'everyone owns it' fails

Shared ownership is how AI governance quietly becomes nobody's job. The IAPP's AI Governance Profession Report finds primary responsibility scattered across functions, with privacy at 22%, legal and compliance at 22%, IT at 17% and data governance at 10%. No function holds a majority. The same research found that only 1.5% of surveyed organisations described their AI governance staffing as sufficient. That is a distribution, not a design. Meanwhile the mandate keeps arriving: Gartner's April 2026 survey reported that 80% of CEOs expect AI to force a high or medium degree of change to their operational capabilities. Deciding who owns this after the first agent is live is an expensive way to find out.

A split that holds up

Write the split as decisions rather than as responsibilities, and the gaps show up on the first read. A function owns something when it can say no on its own. Anything else is a consultation, and consultations do not carry accountability.

• The accountable executive decides whether a use case runs at all, and carries it when the outcome is wrong.
• Legal decides which obligations apply to that use case and what the organisation will commit to in writing.
• Privacy decides what personal information may enter a prompt and what has to be removed before it does.
• Security decides which systems an agent may reach and what authority it may hold inside them.
• The platform team decides what the runtime actually enforces, and refuses a use case whose controls it cannot apply.
• The business owner of the process decides what good output looks like and when to switch the thing off.

What changes once agents act

Governing chat is a content problem. Governing an agent is an authority problem, and it lands on a different desk. An agent that can send an email, update a customer record or release a payment is exercising authority a person granted it, and whoever granted that authority owns the consequence. OWASP published a dedicated Top 10 for Agentic Applications on 9 December 2025, and most of its entries describe authority failures rather than bad model output: an agent given more agency than its job needs, a tool used outside the purpose it was approved for, a privilege escalated through a chain of calls. So add a fourth question to the ownership split. Who decides which actions an agent may complete alone, which stop for a person, and which it should never be able to reach?

A test for whether the ownership is real

Run this before anyone writes a charter. If a question here takes longer than a minute to answer, ownership exists on paper only.

• Name the person who can stop a live agent today, without a meeting.
• Find the record of the last change to what an agent is allowed to do, and the person who approved it.
• Ask what happens at 2 am when an agent proposes an action nobody anticipated. Does the run stop, or does it proceed?
• Ask who reads the record after an incident, and whether they can get it without asking an engineer.
• Ask which function would be named in the incident report. If two functions each name the other, you have found the gap.

What the accountable owner needs to be accountable with

Most governance programmes discover late that they hold policy documents and no record of what actually happened. Difinity.ai keeps two records apart for exactly that reason. The run trail, which Hub calls the AI Trail, is append-only evidence of one run: the message, each guardrail verdict, each proposed action, the tool gateway's decision, approvals asked for and answered, and the outcome. The Configuration Log holds who changed what and who accessed what. Neither can be edited, because the application role holds no update grant and a repeated write is rejected rather than stored twice. Governed run records can contribute operational evidence to wider EU AI Act, ISO/IEC 42001, risk, and audit processes. Difinity does not determine that an organisation or AI system is compliant, and it does not provide ISO/IEC 42001 certification. Deciding which obligations apply stays with your legal and risk owners. How the evidence gets used inside a wider programme stays with them too. The accountable owner is a person, and no platform takes that job off them.

When this answer changes

Two things move it. If your AI only answers questions and touches nothing, most of the authority split above is heavier than the job needs, and a lighter arrangement under legal or privacy will hold for a while. And where a regulator names an accountable role for your sector, that naming beats any internal design you prefer. One caution on the numbers here: the IAPP has published more than one edition of its profession report, the figures move between editions, and at least one edition was co-published with Credo AI, a vendor in this market. Open the current edition before you put a percentage in a board paper.