A wrong fact written into a customer record by an AI agent leaves the firm holding inaccurate personal data, which engages the GDPR accuracy principle. Article 5 requires personal data to be accurate and, where necessary, kept up to date, with every reasonable step taken to correct or erase errors. Prevention means agents write only verified facts, labelled by origin.
What does the accuracy principle require?
Article 5 of GDPR sets six principles plus accountability. Accuracy is one of them. In substance, personal data must be:
- accurate and, where necessary, kept up to date;
- erased or rectified without delay when inaccurate, having regard to the purposes for which it is processed, with every reasonable step taken to make that happen.
People also have a right to have inaccurate data about them rectified, and to have incomplete data completed, including by a supplementary statement. And the controller must be able to demonstrate compliance with the principles, which is the accountability duty.
Accuracy is judged against purpose. A note recording what a customer said is accurate if it records what they said, even if what they said was wrong. A note stating as fact something that is not true is inaccurate.
Who does it apply to?
Every controller processing personal data, which includes any firm whose agents read or write customer, employee or counterparty records. Processors acting for the controller must follow its instructions, but the duty to keep data accurate sits with the controller.
Why are agent outputs an accuracy risk?
Language models can produce fluent statements that are not supported by their inputs. When an agent only answers a question in a chat, the error reaches one person and fades. When an agent writes into a system of record, the error persists and spreads.
Four routes from agent output to inaccurate personal data:
- Unsupported inference written as fact. An agent summarising a call notes "customer confirmed they are unemployed" when the customer said they were between contracts.
- Stale sources. An agent reads an old address from a cached copy and writes it into a new account record.
- Misattribution. An agent handling several cases in one session writes a detail from one customer into another's file.
- Declared facts promoted to verified ones. A customer says their income is a certain amount; the agent records it in a field that downstream systems treat as verified.
The fourth is the most damaging, because it looks correct and travels furthest. A credit process, a collections strategy or a fraud model may later rely on the field. Our article on how to measure an AI agent's accuracy treats grounding as its own score for this reason.
Accuracy also overlaps with the EU AI Act. For high-risk systems, Article 15 requires appropriate accuracy, and Article 10 requires quality data. A wrong fact written by a credit agent can breach both laws at once; see EU AI Act Article 15 accuracy and cybersecurity for AI agents.
How do you prevent agents writing inaccurate data?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Accurate data | Agents write only to fields designed for agent output, never directly to verified fields | Field permissions per agent |
| Distinguish fact from statement | Every written item carries its origin: system of record, named party, customer statement, or outside source | Provenance label on each record |
| Up to date | Agents read live systems of record, not cached copies | Source register; retrieval logs |
| No unsupported inference | Summaries quote or reference the source passage; inference flagged as such | Records showing source references |
| No misattribution | One case per agent context; case identity checked before each write | Session design; write logs with case IDs |
| Reasonable steps to correct | Changes to key fields held for a person; easy correction route for staff and customers | Approval records; rectification log |
| Respond to rectification requests | Process covers agent-written fields and downstream copies | Request log and outcomes |
The control that matters most is the second row. If every fact an agent writes carries its origin, downstream users and systems can tell a ledger balance from a customer's estimate. Without that label, the accuracy problem is invisible until it causes harm.
A related principle: write less. An agent that writes only what the task needs creates fewer chances to be wrong, which is where accuracy meets GDPR data minimisation for AI agents.
For changes to key personal data, a second check by a person is the classic control. Our explainer on the four-eyes principle explains how it works when one pair of eyes is an agent.
What will a supervisor or auditor ask?
- Which records can your agents write to, and which fields?
- How do you tell a fact the agent verified from one a customer stated?
- Show us how a wrong agent-written fact was found and corrected.
- How do corrections reach downstream systems that copied the data?
- What reasonable steps do you take to keep agent-written data current?
What does the accuracy principle not cover?
It does not require perfection; it requires reasonable steps, judged against purpose. It does not govern the accuracy of an agent's answers that are never recorded as personal data, though conduct rules may. It does not replace the right to rectification, which has its own process and timelines. And it does not decide what an agent may do with the data, which is a matter for bounds and oversight. For the full set of rules, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme, every fact reaching an agent carries a provenance class: sealed, signed, declared or external. That label travels with anything the agent writes, so a customer's stated income stays declared and is never presented as verified. Writes to key fields can be held for a named person before they take effect.
Questions people ask
Is an agent's call summary inaccurate if the customer was wrong?
Not if it accurately records what the customer said and is labelled as their statement. It becomes inaccurate if it presents their claim as a verified fact.
Who is responsible when an agent writes a wrong fact?
The controller, which is the firm. The agent is a tool. Responsibility for reasonable steps to prevent and correct errors stays with the firm and the people who set up the agent.
Do we have to tell the customer when an agent wrote wrong data about them?
The accuracy principle requires correction. Whether to notify depends on the effect: if the error led to a decision about them, other duties, including fairness and possibly breach rules, may require you to tell them.