Data minimisation
A pilot uses only necessary fields and examples. Provenance, processing basis, and retention are recorded before transfer.
SECURITY · RISK GOVERNANCE
Security is designed with the AI workflow
This page describes an engineering approach; it does not claim a certification. Controls depend on the data, integrations, error cost, and agreed operating model.
01 · DIRECT ANSWER
Set the boundary before the pilot
Before a pilot, we establish data scope and processing basis, system owners, least privilege, providers, and high-cost errors. Consequential actions are constrained, human-approved, or excluded. Production adds observability, change control, incident procedures, and rollback.
02 · CONTROLS
Eight questions before production
A specific project may require additional controls. This is a minimum engineering frame, not a compliance audit.
Access and secrets
Services receive minimum roles; keys never enter client code, content, logs, or evaluation sets.
Providers
Each external model or API is assessed for processing terms, region, retention, limits, and replacement options.
Human review
Consequential actions require approval, while uncertain and unusual cases move to an accountable person.
Evaluation
A versioned set covers normal, failing, and adversarial cases, including instructions embedded in data and tool output.
Observability
Logs minimise sensitive data while preserving model, version, tool, decision, and safe-failure context.
Change and rollback
Model, data, prompt, or integration changes pass evaluation and have a path to the previous working version.
Incidents and retention
Production defines an owner, escalation path, response expectations, deletion, and recurring risk review.
RESPONSIBILITY BOUNDARY
What this page does not claim
- Certification, an SLA, or universal industry compliance.
- Security without reviewing actual data, architecture, and providers.
- Autonomy for decisions with material consequences.