Kadre Labs

Platform

See what happened. Know what needs you.

Every case is a record of what the agent used, what it wrote, who decided, and what followed. The dashboard shows the three numbers an operator cares about, and alerts say what changed and what Kadre is doing about it.

See what happened. Know what needs you.

Every case is a record of what the agent used, what it wrote, who decided, and what followed. The dashboard shows the three numbers an operator cares about: cases handled, cases needing review, and reviewer time.

One run, span by spanIllustrative example · click a span
Run · QT-1042 · quote follow-up4.1 s of model time · one human gate
model callDraft follow-upGrounded in the two reads above
Scored onlineGroundedness0.96Tone0.91Policy1.00
readpolicy checkmodel callhuman gatewrite-back
Quote follow-up · this weekIllustrative example
Cases handled14
Cases needing review3
Reviewer time1h 40m
Cases needing review, per week12 weeks

Illustrative example, not measured customer results.

alertRenewal drafts needed more edits for two days. Kadre is reviewing the cause.
ApprovalsWho decided, and why
ActivityRules and corrections
OutcomesWhat followed

Illustrative example, not measured customer results.

Evaluated before it ships, and while it runs

Every prompt or rule version is scored against a golden set of cases your reviewers already approved. In production, evaluators score every run and alert when a rate crosses its threshold. A new version becomes the default only when a person promotes it.

Experiments · quote follow-upgolden set · 240 approved cases
VersionGroundednessTonePolicy
v4.0baseline · Aug 12Groundedness0.90Tone0.84Policy0.99
v4.1shorter opener · Aug 21Groundedness0.91+0.01Tone0.88+0.04Policy0.990.00
v4.2account fact first · Sep 2Groundedness0.93+0.03Tone0.90+0.06Policy1.00+0.01
Promote v4.2 to defaultBeats baseline on every metric. Nothing changes until a person decides.
Waiting for a person
Online evaluatorsevery production run · last 14 days
Unsupported claim rate1.4%threshold 3%
Two deployments marked. The spike on day 5 crossed the threshold and that version was rolled back within the hour.
Off-tone rate2.2%threshold 5%
One deployment marked. The rate has fallen since the tone rule was tightened on day 6.
Annotation queueCases a reviewer scores by hand, sampled weekly into the golden set12 to review
Alert · P95 latency over 12 sFired today 13:58. Kadre is reviewing the cause. No case was affected.Firing

Illustrative example. Metric names and thresholds are set per workflow with each client.

Every correction improves the next draft

Your team reviews the work. Kadre uses approved corrections to improve the next run. Patterns across many corrections become insights: an observation with evidence, a proposed change, and a person deciding. Applied changes are measured, and rolled back if the number does not move.

QT-1042 · reviewedAdd a correction
Your quote from the 27th
Hi Dana, checking in on the quote we sent on the 27th. Your last order shipped on time and we can hold the same lead time if you confirm this week. Happy to walk through the line items on a short call.
QT-1058 · next caseStarts with what you taught it

Quote to Harbor Line Logistics, quiet for 11 days

Pick a correction on the left and save it. This case will start from it.

Open insights3 proposed · 0 applied without a person
Observed

Reviewers edit the closing line in 30% of drafts. Evidence: 120 drafts · 30 days.

Recommended

Tighten the closing instruction so fewer drafts need a hand edit. Expected: Fewer edits per draft.

Proposed change
End with an offer to help.End with one specific next step and a date. No open-ended offers.
Decision
ApplyDismissWaiting for a person

The weekly evidence packet

Counts, not claims. Every number traces to the cases behind it, and the definitions do not drift from week to week.

Cases handledEvery case the workflow opened in the week, by trigger.
Cases needing reviewHow many reached a person, and why.
Reviewer timeTime your team spent deciding, edited drafts counted separately.
OutcomesWhat followed each decision, recorded against the case it came from.
DefinitionsFixed in week one. If a definition changes, the packet says so and shows both.

Illustrative example. The packet views on this page use synthetic records. Real packets are defined with each client in the first week of an engagement.

See your workflow as a record

Bring one recurring workflow. We will show what a week of it looks like as cases, decisions, and outcomes.

See how we work