Erasure is mostly a question of what you are obliged to keep. The agent finds every record about the person, matches each one to your retention schedule and any legal hold, and drafts a plan of what to erase and what to keep. A privacy specialist decides; system owners carry out the deletions; the agent verifies they happened.
What is an erasure request, and who owns it?
GDPR and UK GDPR give individuals a right to have their personal data erased in certain circumstances, such as when it is no longer needed or when they withdraw consent and there is no other lawful basis. The right is not absolute. In regulated financial services, a large share of customer data must be retained for set periods under anti-money laundering, tax, accounting and conduct rules, and data subject to a legal hold or an investigation cannot be deleted.
The privacy team owns the response. System owners own the deletions in their systems. Records management owns the retention schedule. The difficulty is that a customer who asks to be forgotten usually expects everything to go, and the honest answer is usually that some data will be erased, some restricted, and some kept for years. That answer must be right and explained clearly. Whether an AI agent can handle a right to erasure request explains the boundary.
What does the agent prepare, and what does a person approve?
- Agent, depth 1: recognises and logs the request. From any channel. The receipt date starts the clock.
- Person, privacy team: verifies identity and validity.
- Agent, depth 1: locates the data. Every system on the data map, the same search as for an access request; see handling subject access requests.
- Agent, depth 1: checks holds and restrictions. Legal holds, open complaints, live investigations, outstanding balances, regulatory retention periods.
- Agent, depth 2: drafts the erasure plan. For each data category and system: erase, restrict, anonymise or retain, with the rule or reason for each.
- Person, privacy specialist: decides the plan.
- Person, system owners: carry out deletions in their systems, or approve a scripted deletion run that the agent prepares.
- Agent, depth 1: verifies. Re-runs the search and confirms the agreed data is gone or restricted, including copies in agent memory, caches and exports.
- Agent, depth 2: drafts the response explaining what was erased, what was kept and why, and the person's rights. A person sends it.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Intake | Logs the request | 1 | None |
| Validity | Nothing | None | Identity and grounds |
| Locate | Searches all mapped systems | 1 | Adding unmapped ones |
| Holds | Checks every reason to keep | 1 | None |
| Plan | Drafts erase, restrict or retain per item | 2 | Deciding the plan |
| Delete | Prepares deletion run if asked | 2 | Approving and executing |
| Verify | Re-searches after deletion | 1 | Signing off |
| Respond | Drafts the explanation | 2 | Sending |
Which controls apply, and what evidence does each leave?
Deletion is a hard-to-undo action. It sits at depth 4 in our scale, and the agent here does not hold it. A system owner executes or approves each deletion run. Evidence: the approval and the run log.
Reasons for retention recorded. Every item kept names the rule or hold that requires it. Evidence: the plan, which is what the regulator will ask for if the person complains.
Legal holds checked first. Data under a legal hold cannot be erased, whatever the request, and the hold register must be one of the sources the agent reads. Evidence: the hold check result on each plan.
Verification after deletion. A second search confirms deletion. Evidence: the before and after search results.
A minimal record of the request. You must keep enough to show you handled the request, and to avoid re-adding the person from another source. Evidence: a suppression record holding only what is needed.
What should you measure?
- Requests completed within the statutory time limit.
- Items found in the verification search that should have been deleted.
- Plans the specialist changed, and how.
- Complaints from customers about the outcome, often caused by poor explanations rather than wrong decisions.
Where does it usually go wrong?
Agent memory forgotten. Agents that retain conversation history or summaries hold personal data outside the main systems. GDPR erasure requests and AI agent memory covers how to find and remove it.
Backups. Data in backups cannot always be deleted immediately. The usual approach is to ensure it is not restored into live use and is overwritten on the normal cycle. Say so in the response.
Third parties not told. Where the firm has shared the person's data with others, such as a credit reference agency or a service provider, the erasure may need to be passed on. The agent should list known recipients from the data map, and the specialist decides which must be informed.
Erasure that breaks retention. Deleting transaction records an AML rule requires you to keep is itself a breach. The agent's hold check must run before any deletion is planned.
Confusing erasure with closing an account. A customer who wants to leave and a customer who wants their data erased are not the same request, though they often arrive together. Product requests follow their own route, such as handling returns requests in retail.
Explanations nobody understands. Retention rules are technical. The response should say in plain words what was kept, for how long and why. Where the requester has asked for an accessible format, see whether an agent can handle accessibility requests.
The stance on compliance agents overall is in AI compliance agents.
Questions people ask
Can the agent delete data directly once the plan is approved?
Not in this design. Deletion is hard to undo, and errors can breach retention rules. The agent prepares the deletion run; a system owner approves and executes it.
What if the customer has an open complaint?
Data needed to handle the complaint is kept until it is resolved and for any period your rules require afterwards. The plan should say so, and the response should tell the customer.
Does anonymised data count as erased?
Truly anonymised data is no longer personal data, but anonymisation is harder than it looks. Treat it as a decision for the privacy specialist, not a default the agent applies.