The agent can explain your privacy notice; it should not interpret it. From the approved notice, it tells customers what data you hold, why, who you share it with and for how long. When a question becomes a request to exercise a right, or asks about their own case, it logs the request and hands it to the privacy team.
What is the process, and who owns it?
Customers ask about their data more often than firms expect. Why do you need my income? Do you sell my details? Who did you send my information to? How long will you keep my recordings? Was a computer used to decide this? Can I see what you hold? Some of these are general questions answered by the privacy notice. Others are requests the law gives customers a right to make, with deadlines attached.
The data protection officer or privacy lead owns the privacy notice and all rights requests. Customer support owns the first conversation. The agent's job is to answer the first kind and recognise the second, which follows our guide to the AI support agent that knows when to say no. GDPR Article 5 makes transparency a principle of processing, and GDPR transparency: telling customers an AI agent processes their data covers what that means when the agent is itself processing.
What are the steps, and at what depth?
- Agent, depth 3: answers general questions from the notice. Categories of data, purposes, the lawful basis for each, categories of recipients, transfers, retention periods, and the customer's rights, quoting the notice section.
- Agent, depth 1: recognises a rights request. A request to see their data, correct it, delete it, object to a use, restrict processing, move it, or have an automated decision reviewed, in whatever words the customer uses.
- Agent, depth 3: logs the request with the date received and the customer's own words, and confirms receipt with the expected timescale from your approved wording.
- Agent, depth 1: verifies identity to the standard the privacy team sets for that request type before any detail is discussed.
- Agent, depth 2: drafts the request summary for the privacy team: what is asked, which systems likely hold the data, any related complaints.
- Person, privacy team: fulfils or responds to every rights request, including any refusal and its reasons.
- Agent, depth 1: routes case-specific questions. "Why did you share my data with that company?" or "Was my data in the breach?" go to the privacy team with the conversation.
- Agent, depth 3: tells customers how to complain to the data protection authority, in your approved wording, when asked.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| General questions | Quotes the privacy notice | 3 | Notice wording |
| Recognise | Spots rights requests | 1 | Nothing |
| Log | Records date and words, confirms receipt | 3 | Nothing |
| Verify | Runs identity check | 1 | Failed or disputed cases |
| Summarise | Drafts summary for privacy team | 2 | Nothing |
| Fulfil | Nothing | none | Every rights response |
| Case questions | Routes with context | 1 | Answering them |
Which controls matter, and what evidence does each leave?
The notice is the only source for general answers. The agent does not explain processing from training knowledge or guess at what "usually" happens. Evidence: the notice version and section cited in each answer.
Rights requests start their clock at first mention. Evidence: the timestamp of the customer's words, not of the case creation.
The agent does not read out personal data in response to an access request. The privacy team assembles and checks the response. Evidence: rights cases closed by named people.
Questions about automated decisions go to a person. GDPR Article 22 gives people rights around decisions based solely on automated processing with significant effects. Whether one applies is not the agent's to say. Evidence: routing records.
Every answer is filed. If the notice was wrong or out of date, you can find each customer told it. Evidence: the append-only record.
What should you measure?
- Questions answered from the notice, by topic. Frequent topics show where the notice is unclear.
- Rights requests logged by the agent, and how many were spotted in conversations not labelled as requests.
- Time from request to privacy team pick-up.
- Answers later corrected by the privacy team.
- Complaints about data handling that started with an agent conversation.
Where does it usually go wrong?
General answers drift into specific ones. "We keep call recordings for a set period" becomes "we have deleted your recording", which the agent does not know. Keep general answers general.
Requests are missed in passing. "I want to know everything you have on me" mid-way through a billing complaint is an access request. The agent must log it even while handling the other matter, as in answering billing questions with an AI agent.
The notice and the terms conflict. Customers quote one against the other. Answer from each with its citation and flag the conflict to the privacy team; answering questions about terms and conditions uses the same discipline.
The agent forgets it is part of the processing. Customers ask whether the agent itself stores what they say. Have an approved answer.
Bill and claim questions reveal data. Discussing a bill or claim with the wrong person is a data breach. Identity checks in can an AI agent answer questions about a customer's bill? apply here too.
Questions people ask
Can the agent delete a customer's data on request?
No. Deletion has exceptions, such as records the firm must keep by law, and the decision is the privacy team's. The agent logs the request and confirms receipt.
Should the agent tell customers which model provider processes their data?
Only what your privacy notice says about recipients, usually by category. If the notice names providers, the agent can quote it.
What if a customer asks whether an AI decided their claim?
The agent says it will pass the question to the team that can answer, and logs it. The decision process belongs to the claims function, as can an AI agent tell a customer the outcome of their claim? explains.