Split a subject access request into finding, filtering and deciding, and give the agent the first two. It confirms the request is valid, searches every system holding the person's data, assembles what it finds, flags third-party data and possible exemptions, and drafts the response. A privacy specialist decides what is withheld, and a person releases the pack.
What is a subject access request, and who owns it?
Under GDPR and UK GDPR, individuals have the right to obtain a copy of the personal data an organisation holds about them, along with information about how it is used. The organisation must respond within a statutory time limit, verify the requester's identity proportionately, and may withhold some information under specific exemptions or to protect other people's data.
The data protection officer or privacy team owns the response. The work is mostly search and review: personal data sits in the core system, the CRM, email, call recordings, case files and, increasingly, AI agent logs. GDPR subject access requests and AI agent logs covers that last source in detail. A thorough search by hand can take days; an agent can do it in hours, but cannot decide exemptions.
What are the steps?
- Agent, depth 1: logs and recognises the request. Requests arrive in any channel and need not use the words "subject access". The agent records the date received, which starts the clock.
- Agent, depth 2: drafts the acknowledgement and any identity request. Proportionate to what the firm already knows about the requester. A person approves the template once; the agent sends at depth 3 within those bounds.
- Person, privacy team: confirms validity and scope. Identity sufficient, any clarification needed, any manifestly unfounded or excessive request handled under the rules.
- Agent, depth 1: searches every system on the data map for the person's identifiers, including agent logs and conversation records.
- Agent, depth 1: assembles and de-duplicates. One copy of each record, with its source system.
- Agent, depth 2: flags third-party data and candidate exemptions. Other customers, staff names, legal privilege, information relating to crime prevention or ongoing investigations. Flags only; no removals.
- Person, privacy specialist: decides. What to disclose, what to redact, what to withhold and why.
- Agent, depth 2: applies the decided redactions and drafts the cover letter, including the information about processing the requester is entitled to.
- Person, privacy specialist: checks the final pack and releases it through a secure channel.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Intake | Recognises and logs the request | 1 | None |
| Acknowledge | Sends approved acknowledgement | 3 | Approving the template |
| Validity | Nothing | None | Identity and scope |
| Search | Queries every mapped system | 1 | Adding unmapped sources |
| Assemble | Collects and de-duplicates | 1 | None |
| Flag | Marks third-party data and exemptions | 2 | Deciding each one |
| Redact and draft | Applies decisions, drafts letter | 2 | Final check |
| Release | Nothing | None | Sending the pack |
Which controls apply, and what evidence does each leave?
Received date fixed. The clock starts on receipt. Evidence: the intake timestamp, unchangeable.
Search coverage recorded. Every system searched, the identifiers used, the hits found. Evidence: a search log that shows the response was complete, or what was not searched and why.
Exemption decisions by a person. Each withheld or redacted item has a named decision-maker and a reason. Evidence: a redaction schedule, which is also what you will need if the requester complains to the regulator.
No investigation leaks. If the person is the subject of a financial crime investigation or report, disclosing that could amount to tipping off. The agent flags any link to restricted cases without revealing the case content in the pack. Evidence: the flag and the decision.
Secure release. The pack goes only to the verified requester. Evidence: delivery record.
What should you measure?
- Requests answered within the statutory time limit.
- Systems searched per request against the data map.
- Items the specialist added that the search missed, which shows gaps in the map.
- Redaction errors found after release, which should be none.
- Complaints to the regulator about access requests.
Where does it usually go wrong?
Requests not recognised. A customer writes "send me everything you have on me" in a complaint, and nobody starts the clock. The agent reads all inbound contacts for this.
Unmapped systems. The data map misses a team's spreadsheet or an agent's memory store. The search is only as complete as the map.
Over-redaction. Removing anything that looks sensitive, including the requester's own data. Redaction should protect others, not hide awkward material about the requester.
Raw exports instead of an intelligible copy. Dumping database rows and system codes into a pack technically discloses the data but leaves the person unable to understand it. The agent should translate codes into plain descriptions and group records by what they are, such as transactions, calls and notes, so the pack can be read.
Bulk requests used as a tactic. Requests made during disputes or by claims firms are still valid. The agent processes them like any other; whether a request is excessive is a person's judgement.
Confusing access with erasure or cancellation. A request may combine several rights or service requests. Split them: erasure follows handling erasure requests, and a request to close a product follows handling cancellation requests.
Requests from regulators for information are a different process with different rules; see whether an agent can answer regulator information requests. The boundary for access requests generally is in whether an AI agent can handle a data subject access request, and the stance on compliance agents in AI compliance agents.
Questions people ask
Do the agent's own logs count as personal data?
Where they contain or relate to the requester, yes. Conversation records, reasoning traces that refer to the person and decisions about them are in scope and must be searched.
Can the agent decide that a request is manifestly unfounded?
No. It can note patterns, such as repeated identical requests, but the decision to refuse or charge carries legal risk and belongs to the privacy team.
Should the agent send the pack once it is approved?
Release is a held action for the specialist. A pack sent to the wrong address or containing someone else's data is a breach, so the final release stays with a person.