A good process for AI agents is high volume, follows clear rules, ends in a result recorded in a system, can be checked and reversed, carries a real compliance need, and does not require licensed judgement or customer contact. Score each candidate on those six questions, check the data is there, and start with the highest scorer.
Picking the wrong first process is expensive. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. A process with a measurable result and clear controls avoids the most common reasons for failure.
Step 1: List candidate processes
Ask each team lead for the recurring work that eats the most hours. Write down each process, monthly volume and roughly how long each item takes. Back-office processes usually top the list: invoices, orders, certificates, policy checks, reconciliations.
Step 2: Score each on six questions
Score each question from 0 to 2.
- Volume. Is it done hundreds of times a month?
- Rules. Could the steps be written on one page?
- Recorded result. Does the result land in a system, like a posted invoice or an issued certificate?
- Recoverable. Can a mistake be caught and fixed before it reaches a customer?
- Compliance weight. Does someone need to prove it was done correctly?
- No license, no customer. Can it be done without licensed judgement or customer contact?
A score of 10 to 12 is a strong first candidate. Below 7, look for a smaller slice of the process that scores higher.
Step 3: Check the data before you commit
For the top two candidates, confirm the agent can get what it needs: read access to the source documents, write access to the system where results land, and at least a few months of history to set a baseline. Missing access is the most common reason a good process stalls.
Step 4: Pick one and define the outcome
Choose one process and write its outcome definition: what counts as done, what evidence proves it and how long it must hold. See how to write an outcome definition. Measure the baseline, then run it.
This scorecard is the same one we use to select work for Agentic MSP. If you would like a second opinion on your list, a Business Process Workshop scores your processes against your own data.