Measuring value: coverage, correctness and review minutes

Most AI value claims rest on self-reported time savings. They are easy to collect and hard to believe, and they rarely connect to a budget line. Decision services can be measured directly because every item has an outcome.

Measure against three baselines

Run the same held-out cases through current manual handling, your existing deterministic rules and the proposed Jev-assisted workflow. If rules alone get most of the way, that is a cheaper result and worth knowing.

The measures that matter

• Coverage: the share of items policy was allowed to act on.
• Correctness among acted items, with the denominator shown.
• Review minutes per item, before and after.
• Time from arrival to first correct action.
• Error cost: what the wrong actions cost to find and fix.
• Unauthorised actions. The target is zero, and one breach is red regardless of the averages.

The pilot targets suggested on the evidence page are a starting point for negotiation, not claims: at least 95% correct routing among automatically routed items, at least 50% coverage and zero unauthorised actions.

A value line a finance director will accept

value = items handled × (review minutes saved × staff rate
                         + delay cost avoided)
      − whole-path cost
      − error cost

Every term can be measured in the pilot. None depends on asking people how they feel about the tool.

Report like an operator

Give each owner one output on a rhythm: the process owner weekly, the platform owner daily, the model evaluation lead every release, the commercial owner monthly. Show uncertainty and sample sizes. A small error-free sample does not prove a low error rate.