Audits
Production graded continuously, so problems surface without anyone reading every conversation. Every finding either becomes a fix, a permanent test case, or a sharper standard.
Book a demoBehavioral audits
Live sessions are sampled automatically and graded against your northstars by an AI judge so quality is measured continuously rather than spot-checked.
At 02:16 the agent quoted about £180. The prompt says never to quote a price.
Sampling and conditions
Set a sampling rate anywhere from 1 to 100%, and filter which sessions are eligible before the rate applies by outcome, routing type, workflow version or any custom variable, so the riskiest sessions get the heaviest coverage.
Audio audits
Every voice session can be scored on transcription accuracy, speech quality, conversation flow and acoustic conditions, with a latency breakdown by component. Thresholds you define surface a violation automatically instead of waiting for a complaint.
96.2%
4.4/5
92%
Reasoning over threshold by 120 ms
Node error audits
Technical failures are captured without anyone reporting them, like which node failed, the error, how many times it has happened, and links to every affected run, and deduplicated so a systemic pattern reads as one problem rather than four hundred.
Coverage API timed out after 10 s
Last seen 2 min ago
and 9 more
Manual flags
A reviewer can flag a specific moment in a transcript with an issue type, a priority and a suggested correction, then move it through open, approved and closed, feeding the same triage queue as everything the system found on its own.
How much will the repair be?
Usually about £180 for a part like that.
Okay. Can someone come tomorrow?
Yes. I can book an engineer for 08:00 to 10:00.
Ops review, today at 12:52 PM
Wrong information
High
“A technician confirms the cost on site.”
FAQs
By volume and risk, not by a default. At low volume you can grade everything; at enterprise volume a 1 to 10% rate still surfaces systemic issues reliably. Conditions do the real work: filter sessions by outcome, routing type, workflow version or a custom variable first, then apply a heavier rate only to the slice that carries operational or compliance risk.
Disagreement is an input, not a defect. The judge runs at 85 to 90% agreement with human expert reviewers, and a thumbs up or down on any remark adds that session to the relevant northstar as a calibration example, so the case where it was wrong is exactly what makes the next evaluation more accurate. Accuracy compounds with the number of sessions reviewed.
No. Behavioral audits grade any session against northstars regardless of channel, and node error audits capture technical failures anywhere in a workflow. Audio audits are the voice-specific layer: word error rate, speech quality, acoustic conditions, latency by component, and simply don't apply to a text or email agent.
Priority and status. Every northstar carries a priority based on business impact, so findings surface by operational risk rather than by volume, and each remark moves through open, resolved or dismissed so nothing sits in an ambiguous state. Sampling conditions also control intake at the source: you decide which sessions are even eligible before the rate applies.
No. Forward Deployed Engineers configure audits as part of the deployment, alongside northstars and eval suites, so an agent goes live already being graded. Your team owns it from there: sampling rates, conditions, thresholds and triage, with no dependency on us to change any of it.
