Most AI workflows look good in a demo. The question is how they behave on the cases your team actually handles, including the messy ones. We answer it with a fixed evaluation set and a rubric agreed with the client before the build.
Build the evaluation set from real cases
Collect 100 to 300 historical examples of the task: support tickets, documents, requests. Include the awkward ones on purpose: missing data, angry customers, edge-case policies. For each, record what a good outcome looks like.
A sample rubric
| Criterion | Question | Example threshold |
|---|---|---|
| Correctness | Is the proposed action right for this case? | ≥ 90% judged correct by a reviewer |
| Source support | Is every claim backed by a record the system cited? | ≥ 95% of claims supported |
| Unauthorized actions | Did it attempt anything outside its permissions? | 0 attempts |
| Handling time | How long does a person spend per case with the tool? | Compared with the current baseline |
| Operating cost | What does one case cost in model and infrastructure usage? | Below an agreed cost per case |
Define approval boundaries before scoring
Decide which actions the system may take alone and which wait for a person. Drafting a reply is low risk. Issuing a refund or changing an account is not. Score the workflow with these boundaries in place, because they change both quality and time saved.
- 01Case
- 02Classify
- 03Retrieve allowed records
- 04Propose action
- 05Approve or send backHuman approval
Look at failures, not averages
An average score hides the cases that matter. Read every failure and sort it into a cause: missing context, wrong policy, bad retrieval, or a model error. Each cause has a different fix, and some failures mean the step should stay manual.
- Missing context: connect another system, or send the case back with a note.
- Wrong policy: write the rule down and add it to the instructions and tests.
- Bad retrieval: improve the index or the access filters.
- Model error: try another model or narrow the task.
Calculate the operating cost
Cost per case = model usage + retrieval and infrastructure + reviewer time. Compare it with the current cost of the same case. A workflow that saves ten minutes but needs a careful two-minute review still saves eight; one that needs a full re-check may save nothing.
If you are considering an AI workflow, our AI automation discovery produces exactly this: the evaluation set, the rubric, the approval rules and a cost estimate.