An agreed baseline and acceptance definition help evaluate a deployment. Their absence makes results harder to assess; no prevalence or cost claim is made here.
Four numbers to agree up front
- Baseline: hours, error rate and cycle time for the workflow today, measured, not estimated.
- Target: a range, not a promise, for what the system should remove.
- Floor: the result below which you stop or change approach.
- Cost per run: model, infrastructure and human review, per case handled.
How to test it
Build a representative evaluation set sized to the workflow and risks, using authorised redacted or synthetic cases. Include normal, edge, ambiguous and high-risk situations. Have accountable reviewers label outcomes and evaluate correctness, escalation, tool choice and cost on relevant changes.
If you can't say when you'd stop, you don't have a business case. You have a hope.