The business is, and inside the business a named person is. To the customer, it makes no difference whether a person or an AI agent gave the wrong information: the business must put it right. Internally, named owners for the agent, its sources and its limits are assigned in advance, so any error can be traced to whoever fixes it.
Why can't responsibility sit with the software?
Because software cannot apologise, compensate, or change its own instructions. Customers and regulators deal with the business, and the business acts through people. In the UK, the Senior Managers and Certification Regime makes named individuals in financial services accountable for their areas, which includes the systems used in them. The principle applies more widely: if nobody owns the agent, nobody fixes it.
Responsibility also needs to be specific. "The AI team" or "IT" is not an owner. A named person for each part of the agent's operation is.
Who owns what?
| Owner | Answers for | Typical errors traced to them |
|---|---|---|
| Agent owner | The agent's brief, bounds and certification | Agent acted outside its role; agent at the wrong depth |
| Content or policy owner | The knowledge base and policy text | Article wrong, out of date or contradictory |
| System owner | The data the agent reads from live systems | Wrong status, wrong balance, failed integration |
| Support lead | Escalation rules and customer remediation | Missed handover; slow or poor fix for the customer |
| Risk or compliance | The limits and the "never say" list | A harmful statement the rules did not anticipate |
Each wrong answer usually has one primary cause, and the record should show which. That is only possible if the agent's log records the sources it used, the versions in force, and the verdict it reached. Escalation responsibilities sit in the escalation matrix for customer support agents, and the broader principles of ownership are covered in what responsible AI is.
How should an error be handled?
- Put the customer right first. A person decides on the remedy: correcting the information, honouring a reasonable reliance on it, compensating where appropriate. Evidence: the remedy recorded on the case.
- Trace the cause. From the log, identify whether the source was wrong, the system data was wrong, or the agent strayed from a correct source. Evidence: the cause, recorded.
- Assign the fix to the owner. The content owner fixes an article; the system owner fixes an integration; the agent owner changes the brief or reduces depth. Evidence: the fix, linked to the error.
- Test the fix. The case joins the test set and is replayed. Evidence: the replay result.
- Look for others. Search for other conversations affected by the same cause, and contact those customers if needed. Evidence: the list and the outcome.
An illustrative example
Take a retail bank's chat agent that tells several customers that a fee for international card payments has been removed. It has not; a help article announcing a planned change was published early.
The support lead arranges refunds of the fee for customers who relied on the agent's answer. The log shows the agent quoted the article correctly, so the content owner is responsible for the cause. The article is withdrawn, a check is added so articles cannot be published before their effective date, and the case joins the test set. The agent owner is not blamed, because the agent did what it was certified to do: answer from approved sources. A search finds every conversation that cited the article, and each customer is contacted.
Had the log shown the agent inventing the claim with no source, the responsibility would sit with the agent owner, and the agent's depth on that topic would drop until it passed again.
What makes accountability work in practice?
- Named owners recorded before launch, not after the first error.
- Logs that show cause: sources, versions, verdicts.
- Measurement that does not hide errors. Teams rewarded on containment may not look for wrong answers; the argument is in containment rate is the wrong goal for a support agent. Measure what matters, as in how to measure an AI customer support agent.
- Clear rules on what the agent must never say, so the riskiest errors are prevented, not just traced; see what an AI support agent should never say to a customer.
Where Pryme fits
In Pryme, every verdict is filed with the sources and their provenance, the certificate in force and the bound that decided it, so the cause of a wrong answer can be traced to a source, a system or the agent. Each agent has a named owner in the Workspace. More is in the AI support agent that knows when to say no.
Questions people ask
Is the model provider responsible for wrong answers?
Your contract may allocate some risks, but to customers and regulators the business that deployed the agent is responsible for what it says.
Should the agent owner be disciplined for an agent's error?
Only as you would treat any process owner. Most errors are system faults to fix, not individual failures.
Do customers need to be told an AI gave the wrong information?
Tell them what was wrong and how it is being put right. Being open about the cause supports trust.