When AI acts on your company’s behalf, “who approved this?” needs a better answer than a Slack scroll. Rills records every decision, human or automatic, as a permanent record: the proposal as it stood, who decided, when, what they changed, and the confidence reasoning that routed it. That’s the accountability layer under everything else in this section.
#What’s on a record
Each proposal and its decision are stored together. The record captures:
| Stored | Details |
|---|---|
| The proposal | The exact action that was proposed: action type, parameters, and description, frozen at the moment it was queued. |
| Context | The trigger data and earlier step outputs the reviewer saw on the card. |
| Confidence | The overall score plus the full per-input reasoning: each input’s score, whether it was deterministic, the validation/anomaly/containment checks it passed or failed, and the model’s rationale for non-deterministic inputs. |
| Routing | Why it reached a human (or didn’t): always-ask, below the bar, a failed check, the circuit breaker, a spot check, a scoring error, or auto-approved. |
| Who decided | The specific team member, or auto-approval, marked as such, never ambiguous. |
| The decision | Approved, rejected, or edited. Edits store the modified action alongside the original proposal, so you can see exactly what changed. |
| Feedback | Any comment or feedback category the reviewer attached. |
| Timing | When the proposal was created, when the reviewer first saw it, and when they decided. |
Auto-approved proposals get records too, marked auto-approved, with the score and reasoning that cleared the bar. Autonomy doesn’t create a gap in the trail.
#Records are immutable
Once made, a decision is recorded permanently : there is no update and no delete, by design, at the database layer. Nobody, including workspace admins, can revise a decision after the fact. If a decision was wrong, the correction is a new action with its own record, not an edit to history.
#Where to see them
- On the run: open any run in run history to see its Human Review steps, what was proposed, and how each was decided.
- In the queue: decided cards leave your pending queue, but the decision and its reasoning stay attached to the run.
- Across the workspace: the Activity tab is the workspace-wide timeline of who-did-what, approvals alongside workflow edits, integration connections, and settings changes. It’s available on every plan, including Starter, and kept for 1 year regardless of your plan’s data retention period.
#Why this matters
If you’re putting AI in front of prospects and customers, someone will eventually ask what went out, who signed off, and why the system trusted it. Decision records answer all three without archaeology:
- Prove sign-off: every send, update, and refund traces to a named approver or a scored auto-approval, with the proposal as it stood at decision time.
- Explain the routing: the confidence reasoning shows why a run paused for review or cleared on its own, not just that it did.
- Audit the autonomy: as you switch steps to Confidence gated , spot checks and their records keep the autonomous path inspectable.
Approvals, and the records they create, are always free. The decision trail is not a metered feature.
#Related
- Approvals : the queue and how to approve, reject, or edit
- Modes & Confidence : the scoring and routing captured on each record
- Learning & Insights : how your decisions shape future routing
- Runs & Versions : run history and the run detail view
- Account & Workspace : the Activity tab and team roles