Triage queue
Relay reads every inbound ticket, classifies intent, and proposes the next action. It acts alone only when it is confident and the action is reversible — everything else waits for you.
Insights
What the AI is actually doing to the operation — and whether its confidence means anything. Calibration is the metric that decides if a threshold can be trusted.
Is the confidence score honest?
If the model says 0.9, it should be right ~90% of the time. Below 0.70 it is systematically over-confident — which is why that band never auto-acts and always shows "AI is unsure" rather than a guess.
Volume handled, last 7 days
Solid bars are tickets Relay triaged end-to-end without a human touch. The gap between solid and outline is human work that still exists — the honest number to show a stakeholder.
Where humans still override the AI
Overrides are the product's most valuable input. Every one is captured with a reason and reviewed weekly — this table is the agenda for that review, not a scoreboard.
AI policies
Governance is a design surface, not a settings page buried in admin. These controls decide what Relay is allowed to do without a human — and every change is written to the audit log with a second approver.
Confidence gates
Two thresholds, three behaviours. Anything below the review floor is never given a suggested answer — Relay asks a clarifying question instead.
Hard limits
These override the confidence score entirely. A 0.99-confident action on a $400k contract still stops for a human — reversibility and blast radius are the real risk model, not accuracy.
Transparency to the customer
Trust is built on the customer's side too, not just the agent's. These are on by default and turning any of them off requires a second approver.
Undo window
Every automated action stays reversible. This is what makes auto-act acceptable to a risk-averse support org — the failure mode is "click undo", not "customer got the wrong answer".
Audit log
Every automated action, every human override, every policy change, every model degradation. Filterable, exportable, and written to before the UI updates — so a crash never loses the record.