AI use policy, DPIAs and incident rehearsals
Write down which data may go into which kind of AI tool and have data screening hold to it, keep a data protection impact assessment for each use, and record the incident rehearsals an auditor asks about.
These are part of Compliance Mode, on Business and Enterprise, beside the AI register. They are your organisation's own records, kept for audit; they are not legal advice.
The AI use policy
Under Governance → AI use policy, one grid says which data may go into which kind of AI tool. The rows are public, internal, confidential, restricted, personal data, special-category data. The columns are the kinds of AI tool:
- Consumer tool: A personal or free account: no contract, and the vendor may keep or train on what is sent.
- Enterprise tenant: The vendor's service under your contract: a data processing agreement, no training on your data, agreed retention.
- Private deployment: A model running in your own cloud account or data centre, in a region you chose.
- Each cell is yes, no, or conditional. A condition is met either by a person approving each time, or by the record having a current, signed-off DPIA.
- Each record in the AI register says what kind of AI tool it is (Edit record → Kind of AI tool). The page lists the records with none, and the ones approved for data the policy says no to in their column.
- Two lines sit under the grid in your own words: where you draw the GDPR Article 22 line, and the uses that are out.
- Every save is in Records → Audit log with the whole policy, so you can show what it said on any day of an audit period.
Holding agents to it
Tick Apply it in data screening and every check Deliverd runs before text goes to a model also reads the grid: the cell for what it found, in the column of the AI system it is going to. A no is handled like data the record is not approved for (redacted, put to a person or refused, as your screening policy says). A condition puts the text to a person, unless it is a DPIA condition and the record has a current one. The grid only ever makes a check stricter. A record with no kind of AI tool, and a model outside the register, are read as a consumer tool.
A DPIA for each use
Article 35 is about processing operations, so the unit is the use, not the tool: a record in the AI register. A record that uses personal data, or makes decisions on its own, asks for one. It has these parts:
- The processing and its purpose (Art. 35(7)(a)): What personal data goes to the model and back, whose it is, where it goes (provider, region) and why: the use, not the tool.
- Necessity and proportionality (Art. 35(7)(b)): Why this data is needed for the purpose (Art. 5: purpose limitation, data minimisation), and what is left out or redacted before it is sent.
- Risks to the people concerned (Art. 35(7)(c)): What could go wrong for them: disclosure, a wrong output acted on, a decision about them, data kept or used for training, a transfer outside the EEA.
- Measures to address the risks (Art. 35(7)(d)): Contract (DPA, no training, retention), data screening, a person approving before anything irreversible, logging, how an erasure request reaches indexes, caches and logs.
- The data protection officer's advice (Art. 35(2)): What the DPO advised, and whether it was followed. If you have no DPO, say who gave privacy advice instead.
- It records where the use stands on Article 22: no decision about people; a person decides; the output informs them; decides on its own, with legal or similarly significant effect. A decision based solely on automated processing needs one of the Article 22(2) grounds.
- An owner, administrator or compliance admin writes it. Someone else signs it off.
- A version keeps what the record said when it was written. When the purpose, the provider, the model, the model version, the data classification, whether it uses personal data, the deployment class, what its output feeds into changes on the record, or its review date passes, it shows as due for review and says why.
- A changed assessment is a new version. Nothing written is edited.
Incident rehearsals
On Governance → AI overview, record each tabletop exercise with an AI system in it: the day, what it rehearsed (prompt injection, personal or confidential data disclosed, harmful or wrong output acted on, an agent took an action nobody approved, provider outage or model withdrawn, other), who took part, whether it ran the GDPR Article 33 72-hour clock, what it found and what will change. A rehearsal counts for twelve months; the overview says when there has been none. A mistake is withdrawn with a reason and stays in the list.
Try it on your own agent.
Free to start, no card. Sign up with Google, Microsoft or GitHub, and the first page after hands you an API key.