Choose a first AI agent workflow that has a clear owner, repeatable work, accessible systems and an outcome you can measure. High volume alone is not enough. A process with missing permissions or unresolved exceptions can consume more implementation effort than it saves.
This scorecard is a starting point for an operations team selecting a pilot. The weights are a proposed decision aid, not a validated prediction of return. Use observed work and a process owner's review to fill it in.
How do you prioritize AI use cases?
Prioritize AI use cases by defining a specific workflow, checking implementation readiness and comparing the evidence for value. Start with recurring tasks that have an accountable owner and a measurable result. Then compare system access, exceptions and human review requirements before choosing the first pilot.
- List candidate workflows from approved observations, process records and interviews.
- Define the trigger, eligible cases and finished outcome for each candidate.
- Hold workflows with unresolved access, ownership or exception handling.
- Score the remaining candidates using the same evidence and weights.
- Select a bounded pilot and record why it is ready to test.
The scoring method below is WIQ's proposed approach. For broader use-case discovery, OpenAI's workflow prioritization matrix provides another framework. If you already have a shortlist, use the automation opportunity assessment template to record scores and evidence.
Define one workflow before comparing opportunities
Write down the trigger, finished outcome, systems, owner and boundaries. “Automate customer service” is too broad. “Read a return request, check eligibility and prepare a return authorization for approval” gives the team something it can observe and test.
Record active handling time separately from elapsed time. An invoice that waits two days for approval may require only six minutes of human work. Reducing the wait and removing the work are different benefits. Event logs show recorded transitions; observations and interviews help explain what happens between them.
Check readiness before scoring value
Hold a candidate if any of these conditions is unresolved. A high score does not override a missing prerequisite.
- A process owner can define a correct result and accept the test outcome.
- The team has permission to use the required data and systems.
- Representative examples can be handled under the organization's privacy rules.
- A person owns policy exceptions, failed actions and requests for help.
- The pilot can run within a limited scope, with review or rollback where needed.
Score each candidate using evidence
Score each row from 0 to 2. Multiply the score by its weight, then add the results. The maximum is 20. Record the evidence and its date beside every score; mark an unknown as unknown rather than assigning zero.
| Criterion | 0 | 1 | 2 | Weight |
|---|---|---|---|---|
| Measured work | Little recurring work | Recurring work; small measured burden | Recurring work with a substantial measured burden | 3 |
| Process evidence | No representative examples | Normal cases documented | Normal cases and exceptions reviewed | 2 |
| System access | Access unresolved | Approved, with integration work remaining | Required actions tested in a controlled environment | 2 |
| Exception handling | No owner or policy | Owner assigned; some routes undefined | Escalation and recovery paths tested | 2 |
| Measurement | No baseline | Baseline exists; outcome definition incomplete | Baseline and acceptance criteria agreed | 1 |
Define “substantial” with your own team before scoring. Use hours, errors or service delays that matter to the process owner. This keeps the assessment consistent without inventing a universal cutoff.
Work through an illustrative comparison
The following figures are invented to show the method. They are not WIQ customer results.
- Return authorization preparation: 600 eligible requests per month, eight active minutes per request, documented policy and a review owner. Scores of 2, 2, 2, 2 and 2 produce 20/20.
- Invoice exception investigation: 200 cases per month, twenty active minutes per case, with examples available but integration work and exception routes unfinished. Scores of 2, 1, 1, 1 and 2 produce 14/20.
The returns workflow offers 80 hours of monthly handling time before automation; the invoice workflow offers about 67. Neither figure is savings. If a returns pilot removes five minutes per eligible request but adds two minutes of review, its gross capacity release is 30 hours per month: 600 × (5 − 2) ÷ 60. Deduct maintenance and exception work before estimating net capacity. Cash savings require a separate staffing or spending decision.
Choose the returns pilot first in this example because its access, exceptions and acceptance criteria are ready. Revisit the invoice workflow after the missing implementation work is resolved.
Use the matching 20-point scorecard worksheet to record readiness, scores and evidence for each candidate. Collect missing measurements with the human baseline template.
Turn the selected workflow into an agent specification
Document inputs, tools, permitted actions, approval points, exception routes and expected outputs. Keep the evidence behind each rule so the owner can correct an assumption before deployment. WIQ's process discovery helps establish how work runs; blueprints organize the steps and controls the team reviews.
Begin with a controlled test. Include normal cases, missing information, policy exceptions, unavailable systems and failed actions. Compare the result with the original baseline using eligible cases, completion accuracy, human review time and recovery work. Expand scope only after the owner accepts the outcome.
Common questions
Should the highest-volume process always go first?
No. Volume increases potential value, but access, process evidence, risk and human handling determine whether the pilot can deliver it.
Do we need process mining or task mining?
Use event data for recorded system transitions and task evidence for work that happens between them. Many workflows need both. The right combination depends on the question and approved capture scope.
What if two candidates have the same score?
Prefer the candidate with fewer unresolved assumptions and a faster controlled test. Keep the score and the written reason; the number supports the decision rather than making it.
Sources and further reading
Microsoft's process-mining overview explains event-data analysis and task mining. For platform selection, see the Bardeen process-intelligence comparison. The scorecard and worked example above are WIQ editorial guidance.
For a support workflow, use the customer-service agent handoff test to define acceptance criteria before deployment.
