Reference

Insights

Insights is a read-only analytics layer over your workforce's execution records. It measures the outcomes, value and reliability of your agents over a date range, both across the whole organisation and per agent.

What it's for

Insights answers the question "is our digital workforce earning its keep?" Every figure is derived at query time from records you already have (agent runs, the side-effecting actions they proposed, and the approval decisions made at the gate), so the numbers reconcile with what you see on the dashboard and in billing.

There are two views:

• Organisation overview (Insights in the sidebar): headline outcome cards, a daily run trend, approval performance, actions by type, and a per-agent breakdown.
• Per-agent performance panel: opened from an agent's page (or by clicking an agent in the overview breakdown). It adds average run duration, token usage and the agent's recent rejection reasons.

The key metrics

The headline cards lead with outcomes and value, not raw provider cost:

• Reliability: completed runs as a share of runs that reached a definitive outcome (completed ÷ completed + failed). Cancellations are a deliberate human stop, not a failure, so they are left out of this rate.
• Actions executed: side effects that actually ran, with the share that succeeded. Proposed or pending actions still awaiting the approval gate are in flight, not yet outcomes.
• Estimated time saved: executed actions multiplied by an assumed minutes-saved-per-action figure.

The per-agent panel adds average run duration (across completed runs), token usage (input/output split) and the same reliability and time-saved figures narrowed to that one agent.

Choosing a date range

Both views share a date-range filter. By default they show the last 30 days. Set a from and to date to widen or narrow the window; the range is inclusive of both ends.

• If you leave the range empty, it falls back to the previous 30 days ending today.
• The span is capped at 366 days so the daily trend chart always stays bounded — a longer request is clamped to that limit.
• If you see "No activity in this range", try widening the dates. Newly created agents that have not run yet will show nothing until their first run.

Estimated figures, shown honestly

Estimated time saved is an estimate, not a measured fact. It is calculated as executed actions × minutes saved per action, where the per-action assumption is configurable (the default is 15 minutes). The assumption in force is always printed beneath the metrics so the estimate is never mistaken for a measurement.

Use it as a coarse guide to the manual effort offloaded, not as an invoiced or audited number. Reliability, action counts, token totals and approval timings are all measured from the records and are not estimates.

Approvals, rejections and governance

Insights shows how your governance gate is performing: the approval and rejection rates over decided requests, the average time from request to human decision, and how many actions are still awaiting approval. This ties Insights directly to the approval centre — the same decisions you make there are summarised here.

On the per-agent panel, the recent rejection reasons list shows the most recent rejected proposals with the reviewer's note. That's the signal an operator reads to decide whether to improve an agent's instructions or retire it.

Permissions and tenant scope

Insights reuses the existing runs.view permission. There is no separate Insights permission. Anyone who can view runs can view both the overview and the per-agent panel; the analytics are simply a view over run data.

Every figure is tenant-scoped: all counts flow through your organisation's data only, and a per-agent panel for an agent in another tenant resolves to a 404. There is no write path here. Insights reads, projects nothing to the audit log and triggers no approvals.

Next steps