Can an AI agent handle a right to erasure request?

It can do everything except the deciding and the deleting. An AI agent can verify the request, locate the person's data in every system and map each record against its retention rule. Deciding what must be kept and approving the deletion stay with a data protection lead, because erasure is permanent and some records must be kept by law.

Why is erasure harder than it sounds?

People have the right, under GDPR and the UK GDPR, to ask for their personal data to be erased in certain circumstances. The right is not absolute. Data may need to be kept to meet a legal obligation, to establish or defend legal claims, or for other lawful reasons.

In regulated firms, that exception covers a lot. AML rules require customer due diligence and transaction records to be kept for a set period after the relationship ends. Complaint files are kept for regulatory review. Some records are under a litigation hold. So a typical erasure request ends in a mix: marketing data deleted, the account closed, but identity and transaction records retained and restricted. Working out that mix is where the effort goes, and where the agent helps.

What evidence does the agent assemble?

For each request the agent builds an erasure map:

• Identity check: the requester matched to the customer record, using the same discipline as when an agent requests identity checks on customers. Erasing the wrong person's data is a serious breach.
• Data inventory: every system holding the person's data, including backups, archives, analytics copies and agent records.
• Retention basis per record: the rule that applies, such as AML record-keeping, an open complaint, a tax requirement or none.
• Holds: litigation holds, open investigations or reports that prevent deletion.
• Downstream recipients: processors and partners who received the data and need to be told.
• Proposed outcome: for each record, erase, restrict, or retain with reason.

Agent memory deserves a note. If an agent has kept anything about the person between conversations, that memory is personal data and falls within the request. How to design for this is in GDPR erasure requests and AI agent memory.

What does the person decide?

• Whether the request is valid and the grounds apply.
• Which retention bases override erasure for each category.
• Whether any exemption, hold or investigation prevents action, and how that is explained to the person.
• Approval of the deletion run.
• The wording of the response.

The last step is subtle. If a customer is under an AML investigation, telling them "we cannot delete your data because of an investigation" tips them off. The officer chooses wording that is accurate without revealing that.

What does an erasure run look like?

Take an online lender. A former customer who repaid a loan two years ago asks for all their data to be deleted.

The agent verifies the request against the email and phone on file. It finds data in nine places: the loan system, CRM, marketing platform, call recordings, the credit reference reporting file, the document store, two analytics tables, and a support agent's conversation memory. It maps each: the loan agreement, repayment history and identity documents are within the AML retention period and the lender's own limitation period for claims, so it proposes restricting them. Marketing data, call recordings beyond the retention rule, analytics copies and the agent memory have no basis to keep, so it proposes erasure. It notes that the credit reference record is held by an external agency under its own rules.

The data protection lead approves. The agent runs the deletion on the approved items only, then re-runs its search to confirm the erased items are gone and reports the result. The response explains what was deleted, what was kept and why, in general terms. The full method is in how to handle erasure requests with an AI agent.

Which controls make it safe?

• Deletion held for approval. The agent proposes; a named person approves. Deletion is irreversible, so it is never routine.
• Only approved items are deleted. The run matches the approved map exactly.
• Verified afterwards. A second search confirms the result, and failures are reported, not hidden.
• Restricted records are restricted. Retained data is moved out of everyday access so agents and staff cannot use it.
• Agents read fresh data. Agents should draw customer data from systems of record at the time they need it, not from copies; see how AI agents get the right data at the right time.

If the same person also asked for a copy of their data, handle that first; see can an AI agent handle a data subject access request? The general principle is in our guide to AI compliance agents.