A privacy assessment should change system design, not merely decorate a procurement file. It identifies which people, data and rights a proposed AI workflow puts at risk and records how those risks are addressed.
Describe the processing before the model
Write down the purpose, lawful basis where applicable, affected individuals, input categories, data sources, output use and decision consequences. Draw the data flow, including vendor systems, connected retrieval databases, support personnel and logging. State why AI is necessary instead of assuming it is the most appropriate technology.
Evaluate necessity and proportionality
Consider whether a simpler rule, local search or human workflow could accomplish the goal with less personal-data exposure. Define retention, access, pseudonymization and deletion controls. Evaluate whether outputs could reveal details about individuals not intentionally included in prompts.
Identify groups and harms
Include risks of misclassification, unauthorized disclosure, re-identification, unfair treatment and reduced ability to challenge a decision. Seek relevant stakeholder input and evaluate whether consent or notice is meaningful in practice. The correct threshold and required form of an impact assessment vary by jurisdiction.
Convert concerns into controls
For each hazard, specify control owner, evidence, residual risk, open questions and approval condition. Examples include strict permissions for retrieval, explicit prohibition of unrelated training reuse, independently reviewed outputs and a deletion procedure verified against actual logs. A promise without a named test or owner is weak mitigation.
Keep assessment alive
Revisit after connecting an external web search, switching model providers, adding a new user population or changing data retention. A small configuration change can create a new disclosure route. Track dates and versions so historical approvals are not read as permission for all future features.
Regulator perspectives
The UK ICO, Australia’s OAIC and European regulators provide official resources, but terminology and legal triggers differ. This guide is a cross-jurisdiction process design; it does not claim a single assessment template satisfies every law.
Sources & Further Reading
- UK ICO guidance on AI and data protection — ICO.
- Privacy guidance for commercial AI products — OAIC Australia.
- Opinion 28/2024 on data protection aspects of AI models — EDPB.
Guidance varies by jurisdiction and application. This is general education, not legal, medical or other professional advice.