How to assess personal data breaches for notification with an AI agent

Use the agent to establish the facts fast and keep the notification decision with your data protection officer. The agent reads the incident record, identifies what data and which people were affected, drafts a risk assessment against your criteria and prepares draft notifications. The DPO decides whether the regulator and the individuals are told, and a person sends.

What is the process, and who owns it?

Under GDPR and UK GDPR, a controller that suffers a personal data breach must assess the risk it poses to individuals, notify the supervisory authority within a short statutory window unless the breach is unlikely to result in a risk, and tell the affected people directly when the risk is high. Every breach, notified or not, must be recorded with its facts, effects and remedial action.

The data protection officer, or the privacy lead where there is no DPO, owns the assessment. Information security owns the containment. Legal advises on exposure. The clock is the hard part: the window starts when the organisation becomes aware, not when the investigation ends, so the first hours are spent working out what happened while the deadline runs. That is where an agent earns its place.

What are the steps, and at what depth?

1. Person, whoever finds it: reports the incident. Through the normal incident channel. Awareness is timed from here, and the agent records the timestamp it receives.
2. Agent, depth 1: assembles the facts. From the incident ticket, system logs and the reporter's account: what happened, which systems, when it started, when it stopped, whether it is contained.
3. Agent, depth 1: identifies the data and the people. Which data categories were involved, how many records, which customer groups, and whether any special-category data such as health information was included. This is classification work of the kind covered in whether an agent can classify personal data.
4. Agent, depth 2: drafts the risk assessment. Likelihood and severity of harm to individuals against your written criteria: identity theft, financial loss, discrimination, distress, loss of confidentiality. Each factor cites the fact it relies on.
5. Person, DPO: decides. Notify the authority or not; notify individuals or not; what further investigation is needed.
6. Agent, depth 2: drafts the notifications. The regulator notification with the facts known so far and what is still being established, and plain-language messages to individuals that say what happened, what it means for them and what they can do.
7. Person, DPO and legal: approve the wording. A person submits to the authority and sends to individuals.
8. Agent, depth 1: writes the breach record and keeps it current as facts change, including follow-up information sent in phases.

Which controls apply, and what evidence does each leave?

Awareness timestamp fixed at intake. The agent records when the organisation became aware and does not move it. Evidence: an immutable intake time against which every later step is measured.

Facts separated from assumptions. Each statement in the draft is marked as confirmed from logs, reported by a person, or assumed pending investigation. Evidence: a fact sheet a regulator can test, and a clear list of what is still open.

Criteria-based risk drafting. The agent assesses against your written risk criteria, not its own sense of seriousness. Evidence: each factor in the draft points to a criterion and a fact.

Named decision. The DPO's decision, with reasons, is filed whether the answer is notify or do not notify. A decision not to notify is the one most likely to be questioned later, so its reasoning matters most.

Minimum data in drafts. Notification drafts describe categories and counts, not the breached records. If extracts are needed, see how agents redact personal data from documents. Evidence: drafts that contain no copies of the affected data.

What should you measure?

• Time from awareness to DPO decision, against your internal target, which should sit well inside the statutory window.
• Changes the DPO makes to the agent's risk draft, and in which direction.
• Facts later found to be wrong in a notification, and why.
• Completeness of the breach record at close, including breaches not notified.
• Customer contacts after individual notices, a rough measure of whether the message was clear.

Where does it usually go wrong?

Waiting for certainty. Teams delay the decision until the investigation is complete and miss the window. The law allows information to be provided in phases; the agent's draft should say plainly what is known now.

Counting records instead of people. Ten records can belong to one person, or one shared record to many. Ask the agent for both counts and say which one drives the assessment.

Missing special-category data. Free-text fields such as notes or complaint narratives often hold health or financial details that a column-based classification misses. The agent should scan free text and say that it has.

Treating agent-caused breaches differently. A breach caused by an AI agent, such as an agent sending one customer's details to another, is assessed exactly like any other. GDPR data breaches caused by AI agents covers how to reconstruct what the agent did from its record.

Forgetting the regulatory angle. Some breaches also breach other rules, such as conduct or operational resilience requirements. Route those to the process for logging and assessing regulatory breaches, so the two assessments agree.

The broader position of compliance agents, preparing and never deciding, is set out in AI compliance agents.