Most AI debates are really data debates. Users want to know what happens to their information, whether it is stored, and how to turn features off. Organisations need controls that are auditable, not just statements in a policy document.
A consent and preference centre is a product pattern that turns those expectations into concrete UI, data flows, and operational rules.
Separate consent from configuration
Users often need two kinds of control:
- Consent controls. Whether certain data can be processed at all, and under what conditions.
- Preference controls. Tone, language, and personalisation choices that improve experience (see assistant memory).
Mixing these into one toggle usually confuses users and complicates compliance.
Define what users can opt out of
Opt-outs should map to real system behaviours, not just UI states. Common opt-outs include:
- Personalisation and memory. Disable long-term storage of preferences and user facts.
- Telemetry content. Reduce or disable content-bearing logs while keeping operational metadata.
- External processing. Restrict data flows to certain vendors or regions (see data residency).
Implement data rights as workflows
Data rights requests are operational workflows, not legal text. At minimum, ensure you can:
- Explain. What data was processed, where it flowed, and which vendors were involved (see telemetry schema).
- Export. Provide a user-visible view of stored preferences and memory.
- Delete. Remove memory and content-bearing data from stores, indexes and caches (see retention and deletion).
Make consent decisions auditable
Consent and opt-out decisions should be captured as structured decisions with reason codes and policy versions. This supports investigations and audits (see decision logging and compliance audits).
Keep the UX honest
Trust improves when controls are transparent:
- Explain what a toggle actually changes in the system.
- Show when a feature is unavailable due to policy or region (see policy localisation).
- Communicate changes via release notes when behaviour shifts (see release notes).
Consent is a trust feature. Treat it as product design plus operational enforcement, not a checkbox.