A workflow is ready for serious assessment when the team can name the work, the bottleneck, the owner, and the evidence that would distinguish improvement from activity. Model selection comes later.
Begin with a unit of work.
Choose a recurring decision, handoff, review, or production step. “Improve operations” is too broad. “Route a submitted claim to the right reviewer with the relevant policy evidence attached” is narrow enough to map and test.
If you cannot identify where the work begins and ends, you cannot define what the system is allowed to change.
Map the current path.
Capture the real path, including the workarounds. A useful map answers these questions:
- What event starts the work, and what state marks it complete?
- Who owns the decision and who carries its consequences?
- Which systems, documents, policies, and informal sources shape the answer?
- Where do people wait, rework, escalate, or bypass the stated process?
- What happens when information is missing, conflicting, late, or wrong?
Define the bottleneck in observable terms.
Separate the symptom from the constraint. “The team is slow” says little. A measurable statement names the affected unit, present behavior, desired behavior, and time window.
For [unit of work], [current condition] should change to [accepted condition], measured by [evidence].
Test whether AI earns a place.
Ask the intervention question after the current state is visible.
Does the work contain uncertainty a model can address?
Language, unstructured evidence, pattern recognition, ranking, or prediction may justify evaluation. Stable known logic usually belongs in deterministic software.
Can representative examples be assembled?
You need ordinary cases, edge conditions, disagreements, and harmful failures. A polished demo set is not an evaluation set.
Can a person or rule judge the output?
Acceptance needs a reference: a policy, a resolved case, an expert judgment process, or another defensible standard.
Can the intervention be bounded and reversed?
Start where access, actions, users, and consequences can be limited. Preserve a safe path when the model is uncertain or unavailable.
Write the assessment record.
- Name the workflow, start event, end state, decision owner, and operating owner.
- Record the baseline, bottleneck, desired change, and evidence window.
- List source systems, data boundaries, permissions, and known quality gaps.
- Describe the proposed intervention and the simpler alternatives considered.
- Define acceptance criteria, evaluation cases, review gates, fallback, and stopping conditions.
Choose the next decision.
The assessment can support four honest outcomes: design a bounded evaluation, gather missing evidence, fix a non-AI constraint first, or stop. A stop decision is progress when the economics, risk, data, or operating ownership do not hold.