Policy, approval, and audit

Every action goes through policy, the risky ones wait for a person, and all of it lands in one audit ledger: who, what, where, evidence, approver, outcome.

Policy on every action

Every action that goes through Datapace, whether an agent drafted it or a person requested it, is checked against policy first. Policy is designed to be set with your team and expressed in the graph's own terms: which databases, which environments, which entities, which kinds of change.

A read on a reporting replica and a resize of a production primary are not the same action, and policy is where that difference is written down.

Human approval on the risky ones

Policy sorts actions into what may proceed, what needs a person, and what may not happen through Datapace at all. Illustratively:

ActionOutcome
Read confirmed definitions for a dashboardAllowed
Resize a reporting replicaNeeds approval
Add an index on a production tableNeeds approval
Drop a tableBlocked

Who approves is part of the policy too. A FinOps proposal can wait for the platform team; a change to a confirmed definition can wait for the data lead who owns the entity.

The audit ledger

Whatever happens is written to one ledger. An illustrative entry:

json
{
  "id": "LG-20260917-0142",
  "when": "2026-09-17T14:02:11Z",
  "who": "FinOps agent",
  "what": "proposed resize of analytics-replica",
  "database": "analytics-replica",
  "evidence": ["cpu p95 30d: 11%", "connections p95: 14/400"],
  "policy": "needs approval",
  "approver": "platform team, approved 2026-09-17T16:40:03Z",
  "outcome": "applied, verified 2026-09-17T16:48:20Z"
}

Six things, every time: who, what, on which database, the evidence, the approver, and the outcome. Entries are designed to be appended, not edited, so the ledger reads as a record rather than a report.

What the ledger answers

The ledger answers the questions that follow an incident and the questions that precede an audit: what touched production, when, why, on whose approval, and with what result. For teams whose AI systems fall under the EU AI Act, it is also the kind of record the Act's logging article asks agent systems to keep.

Design intent

Policy, approval, and the ledger are mechanisms. How they are configured, where they run, and how their data is handled are decided with each partner, engagement by engagement. These pages describe what the mechanisms do, not a deployment.

Was this page helpful?