Usually the business that deployed the agent. The agent is not a legal person, so its actions are the business's actions, to customers and regulators alike. A vendor may share the cost through the contract, but that does not move the duty owed to the customer. Inside the business, named people answer for the agent's authority and controls.
Why does the deploying business carry it?
Because the business chose to use the agent, decided what it could do, and acted through it. To a customer who was wrongly charged, it makes no difference whether a clerk or an agent made the error. The customer's relationship is with the business, and so is the regulator's.
This is the practical meaning of being a deployer. Under the EU AI Act, providers and deployers carry different obligations, and a business using an agent in its operations takes on deployer duties even if it bought the agent ready-made. Other regimes reach a similar place by different routes: conduct rules hold the firm responsible for the outcomes it delivers.
This is a general explanation, not legal advice; how liability falls in a specific case depends on the facts, the contract and the law that applies, so confirm with counsel.
Who answers for what?
Liability outside the business and accountability inside it are different questions. Both need answers before go-live.
| Party | What they answer for |
|---|---|
| The deploying business | The outcome to the customer and to regulators |
| The accountable executive | That agents are governed within the risk appetite |
| The agent's business owner | The agent's role, limits and performance in their process |
| The person who approved a held action | That approval, within their own authority |
| Risk and compliance | That the controls were adequate and tested |
| The vendor or model provider | What the contract says, such as service levels and defects in what they supplied |
In UK financial services, the Senior Managers and Certification Regime makes the accountable executive question concrete: a named senior manager answers for the area the agent works in. How that plays out in a bank is covered in who is accountable when a bank's AI agent makes a mistake.
How is accountability assigned and recorded?
Before go-live, write down the owner, the approver matrix and the accountable executive, and file them with the agent's approval. Each agent's authority should trace to a named person's decision.
At the time of each action, the record should show what the agent did, under which identity, version and certificate, which bounds were checked and what decided, and, for held actions, who approved. Without that record, a business cannot show that its controls were reasonable, which matters as much for defending a claim as for fixing the cause.
After a mistake, the record supports three things: putting the customer right, finding the cause, and showing a regulator or court what happened.
What does this look like in practice?
Take an online retailer's finance team. A customer service agent tells a customer that a late payment fee of 40 will be waived. The waiver was outside the agent's limits, but a gap in how the limit was configured let the message through.
The retailer honours the waiver. The customer acted on what the business told them, and arguing that "the agent was wrong" would be a poor answer to the customer and to any regulator reviewing complaints handling. Internally, the agent's owner leads the fix: the record shows the limit was checked against the wrong fee type. The configuration is corrected, other affected messages are found, and the capability is re-tested. The vendor contract is reviewed, but the configuration was the retailer's own, so no claim follows.
The mistake did not spread to the collections agent, which reads the same fee data, because each agent had its own limit check; see stopping one agent's mistake spreading to others.
Which controls reduce the exposure?
- Authority matched to evidence. Agents act only at depths they have proven on your own cases.
- Limits enforced at the moment of action, not described in instructions.
- A complete record that can reconstruct any decision.
- Named owners and approvers, written down before go-live.
- Contract terms that set out what the vendor answers for and what evidence they must give you.
- Insurance review, so cover reflects that agents now act on the business's behalf.
Where Pryme fits
Pryme does not take on the deploying business's liability; it gives the business the means to show its controls were sound. Each action passes five bounds at the moment of action, the verdict names the bound that decided, and every verdict is filed in an append-only record. More on accountability in who answers for an AI agent's mistake and our guide to agent governance.
Questions people ask
Can a business disclaim liability for its agent's statements?
A disclaimer may affect some claims, but it does not remove regulatory duties or make misleading a customer acceptable. Assume what the agent says is what the business says.
Is the approver personally liable for a held action they approved?
They are accountable inside the business for that approval. Personal legal liability is a separate question that depends on the role and the law; ask counsel.
Does using a certified vendor agent reduce liability?
It may help show diligence, but a vendor's certification is not evidence the agent works in your operation. Test it on your own cases.