AI operations
Contract
Agent runtime contract
A practical contract for running agent actions in production: identifiers, approvals, audit trails, and safe reruns. This is the model used across Gestio AI workflows.
Non-negotiables
If an AI action cannot meet these, it must fail closed or route to manual review.
- Traceable: every action has an identifier (e.g. ai_task_id) and an audit trail.
- Reviewable: outputs include evidence/inputs; uncertainty is explicit and routes to review.
- Overrideable: humans can approve/reject and re-run safely without “hidden state”.
- Tenant-safe: actions are scoped by tenant isolation and never cross workspaces.
Inputs and evidence
Make reviews fast by attaching evidence where decisions happen.
- Store source references (documents, entity IDs, request payloads).
- Persist extracted/derived values as drafts, not final truth.
- Keep a “what changed” story from draft → reviewed → approved.
Approvals and policies
Commitments and money-impacting actions require explicit approvals.
- Approval gating for commitments (award decisions, PO issuance, payments).
- Fail closed on missing permissions or missing evidence.
- Keep approvals linked to the action identifier for auditability.
Deterministic fallback + safe reruns
Plan for failures and retries. Reruns must be safe and explainable.
- Retry only fetch-level failures; don’t loop on logical errors.
- Use idempotent patterns so reruns do not create duplicates.
- Record status history so operators can understand the run timeline.
Security boundaries
AI actions are permissioned operations, not “chatbot suggestions”.
- Least-privilege permissions (deny by default).
- Explicit tenant isolation (platform_user_id scoping).
- Audit logs for approvals, writes, and edge function invocation failures.
Related docs
Follow the model into workflows and product pages.