Data policy meets AI
Data governance rules were written for databases and reports. AI systems add new paths for data to travel: collection for training, storage in embeddings, retrieval into prompts and generation into outputs. Each path can violate a rule that was satisfied elsewhere. A policy control plane helps by expressing data rules once and applying them at each stage.
Stage one: what goes into training and tuning
Before data is used to train or fine-tune a model, the pipeline should check its origin, license or consent basis, classification and permitted purpose. Data collected for one purpose may not be reusable for another, and personal data may carry deletion obligations that are hard to honor once information is embedded in model weights. Record provenance for every dataset, and make approval a gated step rather than an informal agreement.
Stage two: retrieval and context
Retrieval-augmented systems pull documents into the prompt at request time. The key rule is that retrieval must respect the caller's entitlements. If a user cannot open a document directly, an assistant should not summarize it for them. Apply the source system's access rules at query time, or carry entitlement metadata into the index and filter on it. Enforcing access only at ingestion leaves stale permissions in the index after rights change.
- Carry classification and owner metadata with every chunk.
- Filter results by the requesting identity before they enter the prompt.
- Re-evaluate when permissions change, not only when content changes.
- Log which documents were supplied to which request.
Stage three: outputs and downstream use
Generated text can reproduce sensitive input or combine safe fragments into something sensitive. Apply output checks for personal identifiers, secrets and restricted categories, and decide what happens on a hit: block, redact or route for review. Also control where outputs flow. A summary written into a shared workspace inherits a wider audience than the private prompt that produced it.
Retention, deletion and evidence
Prompts, responses and logs are data too. Define how long they are kept, who may read them and how deletion requests propagate to caches, indexes and backups. Keep evidence that the rules ran: dataset approvals, retrieval filters and output decisions. Where regulation applies, such as privacy law, consult qualified counsel to confirm which obligations attach to your use case.
Questions for a data owner
- Which datasets may be used for training, and what approval is required?
- Which documents may an assistant retrieve for which groups of users?
- How long are prompts, outputs and embeddings retained?
- How is a deletion or correction request carried through indexes and caches?
- Who reviews a sample of outputs for leakage each month?
If the data owner cannot answer these, the pipeline is not yet governed, whatever the written policy says.
Roles and responsibilities
Assign a named data owner for every dataset, a platform owner for the pipeline and a reviewer who can approve exceptions. Without named people, rules about consent, retention and access fall between teams. Hold a short review each quarter in which owners confirm that the datasets, indexes and outputs in use still match what was approved, and record the result alongside the policy version.
