The agent's named owner is accountable. An agent acting alone has done so because a certificate granted it depth for that capability and the business's own limits allowed the action. The owner answers for both: for seeking that certificate, for setting or accepting those limits, and for watching what the agent did with them.
Why can't the agent be accountable?
Accountability means someone can explain a decision, answer for its consequences and change what happens next. An agent can do none of these in a way a regulator, court or customer will accept. It has no authority of its own; whatever it does, it does under authority a person granted.
That is why every agent needs a named owner inside the business, as set out in who should own an AI agent inside a business. Without one, accountability falls into the gap between IT, the vendor and whoever happened to be using the agent.
What allowed the action?
When an agent acts without a person, every action has still passed five bounds at once: role scope, certificate, sources, the customer's own limits and the platform guardrails. The lowest bound wins, and the verdict names which bound decided. That gives a precise chain of who answers for what:
| What allowed the action | Who answers for it |
|---|---|
| Role scope | The owner, who defined the agent's job |
| Certificate at that depth | The owner, who sought certification; the builder, who prepared the agent |
| Sources and their provenance | The owner of each source system |
| The customer's own limits | Whoever set them, under change control |
| Platform guardrails and withheld actions | The platform provider |
The owner sits at the centre because they chose to put the agent into work, accepted its limits and are responsible for watching the result.
What does the record show afterwards?
Every verdict is filed in an append-only record. For an action taken alone, the record shows what the agent saw and where each fact came from, the depth it acted at and the certificate that granted it, the limit it acted within, and the bound that decided.
Take an insurer whose claims agent pays a 4,200 claim on its own. The auto-pay limit is 5,000; the agent is certified for claims payment at depth 4. The claim later turns out to be fraudulent. The record shows the repair invoice arrived from an approved repairer, signed, and everything was within limits. The owner answers for the outcome: was a 5,000 limit appropriate for this claim type? The builder answers for whether Attest's cases covered this pattern. Nobody blames the agent; the chain shows where to act.
How does regulation frame this?
Regulators place accountability on organisations and named people, not on systems. In the UK, the Senior Managers and Certification Regime assigns named individual accountability; see SM&CR reasonable steps when an AI agent acts. Under the EU AI Act, providers and deployers carry different obligations, including human oversight under Article 14 and deployer duties under Article 26; see the EU AI Act and agentic AI. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Who should the owner be?
Usually a manager in the business function the agent serves, not IT. IT owns connections and access; the function owns outcomes. The case is made in operations, not IT, should own agents in production. The wider role map is in AI agent skills by role.
Where the Academy fits
An owner must understand depth, certification and the record well enough to answer for them. Pryme Intelligence Academy's Foundations track covers ownership, data and trust in about four hours. Owners who run the workspace daily take the Operator track, and those leading regulated, multi-agent deployments take the Architect track, which ends in a board review with two Pryme assessors.
Questions people ask
Is the vendor accountable for the agent's actions?
The vendor answers for the platform's guardrails and withheld actions. Decisions the business allowed through its own limits and certification remain the business's.
What if the certificate had lapsed when the agent acted?
A governed platform drops a lapsed agent's authority, so the action should not have been allowed. If it was, that is a platform failure to investigate, and the record will show it.
Is the person who set the limit accountable?
They answer for the limit through change control. The owner remains accountable overall, because the owner accepted the limit for the agent's work.