Mostly, yes. An AI agent can log a data subject access request, find the person's data across systems, collate it, propose redactions and draft the response pack against the deadline. Accepting the requester's identity, deciding what is exempt and approving release stay with the data protection or compliance officer, because a wrong disclosure is itself a breach.
What makes subject access requests hard?
Under GDPR and the UK GDPR, a person has the right to obtain a copy of their personal data and information about how it is used, within a fixed statutory period. The request does not have to use any particular words. A complaint email that says "send me everything you hold on me" is a request.
The effort is in the search. Personal data sits in the core system, the CRM, email, call recordings, chat transcripts, case notes, spreadsheets on shared drives and, increasingly, the records of what AI agents did with the customer. Finding all of it, removing duplicates and separating out other people's information takes days per request. That is work an agent does well. How agent logs fit into the search is covered in GDPR data subject access requests and AI agent logs.
What may the agent prepare?
| Stage | The agent | The officer |
|---|---|---|
| Intake | Recognises the request in any channel, logs it and starts the clock | Confirms it is a valid request |
| Identity | Checks the requester against the customer record and requests proof if needed | Decides identity is sufficient where it is not clear-cut |
| Search | Searches each system in scope and records which systems it searched | Adds systems the agent cannot reach |
| Collation | Removes duplicates and organises by source and date | Nothing routine |
| Redaction | Proposes redactions for third-party data and marks material that may be exempt | Decides exemptions and approves redactions |
| Response | Drafts the cover letter and the information about processing | Approves release |
| Tracking | Watches the deadline and flags extensions | Decides to extend, where the law allows |
The agent reaches systems through access it has been granted for this purpose, not through broad standing access. Setting that up is covered in how to let an AI agent request access through your access management system.
Which calls stay with a person, and why?
Three decisions carry most of the risk.
Third-party data. Customer files contain other people: joint account holders, staff, people named in complaints. Whether to disclose, redact or seek consent depends on context the agent may not see.
Exemptions. Some material can lawfully be withheld, such as legally privileged advice or information whose disclosure would prejudice the prevention of crime. A request from a customer who is the subject of a suspicious activity report is the sharpest example: disclosing internal notes about the suspicion could tip them off. The officer decides.
Release. Sending the pack is irreversible. A pack sent to the wrong address, or with someone else's data in it, is a personal data breach.
What does a request look like?
Take an insurer. A former policyholder in a dispute over a declined claim emails asking for "all information you hold about me, including call recordings and internal notes."
The agent logs it, matches the email to the customer record and starts the clock. It searches the policy system, claims system, call recording store, complaints system and email archive. It finds 340 items, removes duplicates to 210, and transcribes eleven calls. It proposes redactions for a loss adjuster's personal phone number and the name of a neighbour who gave a witness statement. It marks two documents as possibly privileged: emails between the claims team and external lawyers about the dispute.
The officer confirms the privileged documents are exempt, keeps the neighbour's name redacted, and approves release. The agent sends the pack through the secure portal and records what was disclosed and what was withheld, with reasons. The playbook version is in how to handle data subject access requests with an AI agent.
Which controls make it safe?
- A search record. The response file lists every system searched and every system not searched, so a later complaint can be answered.
- Redaction applied to copies. Originals are never altered.
- Release held. Sending the pack is held for a named person.
- Exemption reasons filed. Anything withheld has a reason a regulator could read.
- Linked requests spotted. A subject access request that arrives alongside an erasure request or a police request is flagged so the two are handled consistently.
The wider principle is in AI compliance agents: second line, never the last word.
Where Pryme fits
In Pryme, every verdict an agent reached about a customer is filed in an append-only record, so a subject access search can find what agents did with a person's data as easily as what staff did. Sending the response pack is a held action: the agent prepares it and a named officer releases it.
Questions people ask
Can the agent refuse a request it thinks is excessive?
It can flag the request with its reasons. Refusing, or charging for, a request is a legal judgement for the officer.
Should the agent disclose its own records about the customer?
Records of what an agent did with a person's data are generally within scope. The officer decides how to present them.
Can the agent answer requests from a third party acting for the customer?
It can check the authority provided and flag it. Accepting that someone may act for the customer is a decision for a person.