Repetition is a reason to inspect a task, not automatic proof that it should be automated. A frequent task may depend on inconsistent inputs, sensitive decisions or an exception that changes the outcome.
Begin by defining the work precisely enough to compare alternatives. Then decide whether the next step should be process improvement, conventional automation, AI assistance or more investigation.
Download the candidate worksheet.
Define one task at a time
“Automate procurement” is a program, not a testable task. “Draft a supplier-information request from an approved list of missing fields” is narrow enough to examine.
Record the trigger, input, expected output, owner, systems and human review. Use the roles, duties and tasks worksheet where those details are unclear.
Examine six decision factors
Input quality: Is the needed information available, understandable and permitted for the proposed use?
Variability: Does the task follow a stable pattern, or do exceptions change the route?
Reviewability: Can a person or a reliable check determine whether the output is acceptable?
Consequences: What happens when the task is performed incorrectly, late or twice?
Operating value: What burden or constraint would improve, and what evidence supports that expectation?
Ownership: Who can authorize the experiment, review it and stop it?
Do not combine these into a universal score that overrides a serious unresolved risk. A single permission or safety issue can matter more than several attractive efficiency estimates.
Choose the simplest suitable change
If requests arrive without required information, improve the intake before generating faster replies. If two systems need a predictable transfer, evaluate a conventional integration. If a person spends time turning permitted source material into a reviewable draft, AI assistance may be worth testing.
A recommendation to keep a human decision or defer a project can be a useful assessment result. The goal is not to maximize the number of tasks labeled “AI.”
Example: preparing a customer update
Illustrative: An employee gathers approved status information and drafts a customer update. A candidate pilot might assist with drafting while a person checks accuracy and decides whether to send. It should not quietly expand into promising delivery dates or changing the underlying order.
Measure the original work, the review effort and the quality of the final result. A faster draft is not automatically a faster end-to-end workflow.
Keep the value model honest
Separate measured volume from employee estimates. Distinguish hands-on effort from elapsed waiting time. Record software cost, review cost and implementation work alongside the potential benefit. Do not add overlapping time estimates from several stages of the same task.
At an early stage, “we need to measure this” is a more useful answer than a precise-looking return based on invented inputs.
Turn a candidate into a pilot brief
The brief should name the work owner, permitted input, expected output, human checkpoint, prohibited actions, test cases, evaluation method and stop conditions. Include normal cases and the exceptions participants described.
Use an AI opportunity assessment to compare candidates. Use an AI readiness assessment to examine the prerequisites for the selected one.
Start with discovery when the task is still vague
DutyGraph gathers leadership and workforce context to help an advisor understand what is actually performed. The resulting record can support candidate selection; it does not deploy an agent or authorize its actions.
When the problem spans people and departments, begin with a business operations audit rather than a software shopping exercise.