Tell customers what happened in plain words, what data or systems were involved, what you have done, what they should do, and when you will update them. State only what you have confirmed, say clearly what is still unknown, and have legal and privacy review before sending.
This is general guidance, not legal advice.
What the GDPR text expects from a notification
Where personal data is involved, Article 33 lists what a notification to the supervisory authority should at least contain: the nature of the breach, including categories and approximate numbers of people and records, a contact point, the likely consequences, and the measures taken or proposed. Article 34 requires communication to affected individuals without undue delay when a breach is likely to result in a high risk to their rights and freedoms. Customer messages that follow the same structure are easier to write and easier to check.
Your contracts may set their own deadlines for telling customers, separate from any law. Read them before an incident, not during.
A simple structure
| Section | What to write |
|---|---|
| What happened | One or two factual sentences, including that an AI agent was involved if that is established |
| What was affected | Categories of data or services, and who is affected, as far as known |
| What we have done | Stopped the agent, revoked access, preserved evidence |
| What it means for you | Likely consequences, in plain language |
| What you should do | Concrete steps such as rotating a key or watching for phishing |
| What we are still working out | Honest list of open questions |
| Next update | A specific time, and a contact |
Wording that helps
Use active sentences. Name the thing that happened: "an automated assistant sent a file to an address outside our company" is clearer than "an anomalous event occurred". Do not blame the model, the vendor or an employee. Do not speculate about cause or attackers before evidence exists. Do not minimize with words like "limited" unless you can back them up.
Avoid these mistakes
Over promising ("no customer data was affected") before the investigation finishes. Saying "an AI did it" as if that explains responsibility. Using the agent's own summary as the basis of the statement. Sending different facts to different audiences. Going silent after the first message.
Steps for the first day
- Stop the agent and preserve logs.
- Assemble the facts you can verify, with timestamps, from downstream systems rather than the agent's description.
- Ask privacy and legal whether notification duties apply, to whom and by when.
- Draft the message from the table, keeping unknowns explicit.
- Agree one spokesperson and one source of truth.
- Prepare a support route for questions.
- Schedule the next update before sending the first one.
Prepare a holding statement now
A short pre approved template, with blanks for facts, saves hours when you are tired. Keep it in your incident runbook with the names of approvers.
After the message
Track questions customers ask and feed them into your review. Publish a clear follow up when the investigation ends, covering cause, fixes and what has changed in your controls. Customers tend to judge a company by how it handles the second message as much as the first.
Safeguard offers compliance and evidence automation and agent identity features that help you establish facts quickly. A free scan is at https://gold.safeguard.sh/scan.
Quick check
Should we say an AI agent was involved?
If confirmed and relevant, yes, plainly. Do not use it to deflect responsibility.
What if we do not know the numbers yet?
Say what is unknown and when you will update. The GDPR text allows information to be provided in phases.
Sources
- GDPR Article 33: Notification of a personal data breach
- GDPR Article 34: Communication of a personal data breach to the data subject
Last reviewed: 19 October 2026.
