The business owner of the process the agent works in. Ownership follows the decision, not the technology: whoever would own a refund, a disbursement or a classification if a person made it owns it when an agent does. IT runs the platform, risk oversees, and internal audit checks independently, but none of them owns the agent's decisions.
Why not IT, or the team that built the agent?
Because they cannot judge whether a decision was right. IT can tell you the agent was running, the connection held and the data arrived. It cannot tell you whether a hardship arrangement was appropriate or a claim should have been paid. Only the process owner can, and only the process owner can change the procedure when the answer is no.
Giving ownership to IT also creates a quiet gap. The process owner assumes IT is watching the agent; IT assumes the business is checking its outputs. Decisions end up owned by nobody. The case for keeping it with the business is made in our guide on governing AI agents across a whole company.
How do the roles split?
| Role | Owns | Does not own |
|---|---|---|
| Process owner | The agent's decisions in their process; its role, limits and hand-over rules | The infrastructure |
| Accountable executive | That agents in their area are governed within appetite | Individual decisions |
| IT and platform team | Availability, security, connections, identities | Whether a decision was right |
| Risk and compliance | Oversight, challenge, approval of limits | Day-to-day decisions |
| Internal audit | Independent assurance that the above works | Anything it audits |
| Named approvers | The held actions they approve | The agent's other decisions |
In UK financial services, the accountable executive maps onto the Senior Managers and Certification Regime, which requires a named senior manager for each area of responsibility. Which one should own an agent is discussed in SM&CR: which senior manager owns an AI agent.
What about an agent that works across several processes?
Split ownership by capability, not by agent. An onboarding agent might verify identity documents, which compliance owns; open the account, which operations owns; and send the welcome pack, which marketing owns. Each capability has one owner. The agent as a whole has a lead owner who coordinates changes and certification, but each decision traces to the owner of its step.
The same holds when several agents work together under an orchestrator. Coordination does not create a new owner; each agent's decisions belong to the owner of its process.
How is ownership assigned and recorded?
In an ownership register, kept with the agent inventory. Each capability lists its owner, their cover, the accountable executive, and the date of the last review. The register is approved when the agent goes live and updated whenever a capability or owner changes.
At the moment of action, the record ties each verdict to the capability, and so to the owner. When a decision is questioned, there is no debate about who should answer it. Internal audit needs to be able to read this record independently, for reasons set out in internal audit needs its own view of agent decisions.
What does this look like in practice?
Take a payments company. Its onboarding agent checks business customers' documents, scores their risk and opens accounts. One week, a customer complains that their account was opened with the wrong trading name.
Under the old arrangement, the complaint went to IT, which found the agent had read the registered name rather than the trading name. IT had no authority to decide which name should be used. Under the register, the capability "record business names" is owned by the head of onboarding operations. She decides trading names must be recorded alongside registered names, updates the procedure, and asks for the capability to be re-tested. IT implements the source change she asks for.
The customer also asks for a fee to be waived as an apology. That is an exception to company policy, and whether an agent can grant one is a separate question; see whether an AI agent can make exceptions to company policy. Here a person grants it.
Which controls keep ownership real?
- An owner per capability, named, with cover.
- Owners sign off limits and hand-over rules for their capabilities.
- Owners receive the reports: refusals, holds, overrides and errors for their capabilities.
- Ownership reviewed at each recertification and when people change roles.
Ownership per capability is what makes agent governance workable: every bound has someone who can change it and answer for it.
Questions people ask
Can a vendor own an agent's decisions?
No. A vendor can be responsible for what it supplies under contract, but decisions in your process are yours.
What if the process owner does not want the responsibility?
Then the agent should not be deployed in that process. An agent without a willing owner is a risk nobody manages.
Does the owner need to understand how the agent works?
Enough to judge its outputs and set its limits. They do not need to understand the model; they need to understand the process and the evidence.