Skip to content

Production AI governance

Enterprise

AI features are production features when they can see tenant data or trigger writes. Use this checklist before enabling a real provider.

  • List which fields can enter prompts, tool arguments, extracted documents, and model responses.
  • Decide whether each field may leave the tenant’s network or must use a local endpoint.
  • Keep prompts and responses out of logs unless the data classification allows it.
  • Treat assistant conversations and tool-call records as user or tenant data.
  • Define retention, export, and erasure behavior before collecting real turns.

The cost control guide covers step policy and the levers that cut provider cost.

  • Mount only the handlers and providers the app needs.
  • Add descriptions and risk declarations to every handler the agent may see.
  • Test the role-filtered catalog with a normal tenant user and an operator.
  • Start in read-only or approval mode.
  • Keep high-risk handlers out of unattended approval rules.
  • Verify that accepted writes dispatch as the calling user.

Use Make your app agent-ready and the Assistant walkthrough for the concrete approval model.

  • Store provider API keys in the deployment secret store.
  • Select providers per tenant only after the data boundary is documented.
  • Set model, endpoint, and output-token limits deliberately.
  • Use the provider matrix to decide between a hosted, compatible, local, or mock backend.
  • Test timeout, retry, rate-limit, malformed-response, and provider-outage behavior.
  • Use the mock provider for deterministic handler and integration tests.
  • Keep golden inputs and expected outputs for pipeline steps.
  • Draft policy or prompt changes before activation when the workflow supports it.
  • Run a dry run without a domain write.
  • Check the proposed output with the same schema and access rules used by the domain handler.
  • Record a rollback procedure for prompts, policies, and provider changes.

Track at least:

  • request count, latency, timeout, and provider error rate;
  • token usage and tenant cap consumption;
  • prompt revision, provider, model, and step outcome;
  • proposal, approval, rejection, and executed tool-call counts;
  • extraction validation failures and pipeline dry-run mismatches.

Alert on provider failures and cap exhaustion separately. A provider outage is not the same incident as a tenant reaching its budget.

AI output is a candidate. Keep a normal form, queue, or handler path available when a user rejects a proposal, validation fails, a provider is unavailable, or an extraction is incomplete. The application owns the final domain write.

You can implement these controls with public Kumiko features: tenant-scoped config and secrets, custom handlers, jobs, audit, cap-counter, and your own provider adapter. Enterprise packages the common revision, approval, provenance, and evaluation patterns so each app does not have to rebuild them.