Skip to content

Evidence

What to measure before expanding authority

Compare three baselines on the same cases: current manual handling, existing deterministic rules and the proposed Jev-assisted workflow. Separate development examples from held-out evaluation and later shadow-production samples.

Evidence status legend

Proposed design

Architecture, thresholds or flows recommended in this guide. Not yet run in a customer tenant.

Synthetic demo

Illustrative values, Contoso-style scenarios or lab fixtures. Useful for teaching the loop; not a result.

Measured pilot

Held-out or shadow results with denominators, error costs and agreed acceptance gates. Replace proposal labels only when this exists.

Production case study

Live authority with operational ownership, kill switch and ongoing scorecard. None claimed on this site yet.

Suggested pilot targets

Negotiate rather than represent as achieved results:

  • At least 95% correct routing among automatically routed items
  • At least 50% automation coverage
  • Zero unauthorised actions
  • p95 end-to-end latency from the customer’s process requirement

Report uncertainty intervals and denominators. A small error-free sample does not establish a low incident rate.

Scorecard rhythm

OwnerRequired outputReview rhythm
Business process ownerAccepted routes and error-cost assumptionsWeekly in pilot
Platform ownerAvailability, latency, queues, recoveryDaily operations
Security and data ownerData boundary, retention, accessBefore launch and on change
Model evaluation leadVersioned labelled set and threshold reportEvery model or policy release
Commercial ownerOutcome value and complete service costMonthly

Mark Red when an agreed safety or service limit is breached. Do not average away a critical red control.

Cost note (illustrative)

Vendor documentation checked around the verification date listed model inference in the order of cents per hundred thousand items at stated token assumptions. Review labour and error remediation dominate the business case. Invented planning assumptions in related materials are labelled as such and are not customer results.