Start with work you can describe
The best first candidate is usually a recurring task that follows stable rules and produces an output someone can check. Examples include preparing a weekly update from a structured task list, creating an onboarding checklist from an approved form, or drafting a response to a known type of inquiry.
A vague goal such as “use AI for sales” is too broad. “Prepare a response draft and assign the next task for every completed inquiry form” is something you can test.
Compare three candidates
| Candidate | Good first version | Reason to wait |
|---|---|---|
| Client onboarding | Prepare a checklist from a complete intake | No standard service or owner |
| Proposal preparation | Draft an outline from approved notes | Every price or commitment needs expert judgment |
| Weekly reporting | Summarize structured updates for review | The source data is not trusted |
Use this five-part decision
- Volume: count how often the task happens in a normal month.
- Effort: time several examples, including review and corrections.
- Clarity: write what a correct result contains.
- Access: verify that the information may be used in the chosen tools.
- Ownership: identify who will review errors and keep the process useful.
If the task has little volume, a checklist may be a better investment. If the rules are unstable, simplify before building.
A worked planning example
Suppose a team prepares 40 client updates a month, taking 15 minutes each. A draft-and-review workflow that takes 7 minutes per update would recover about 5.3 hours a month. At an assumed $50 per hour, that represents about $267 of capacity—not cash automatically saved.
If tools and maintenance cost $150 monthly, only about $117 of modeled capacity-inclusive benefit remains. A $3,500 build would take roughly 30 months to recover its cost on those assumptions. That may be a poor first project. More valuable work, better scope, or a simpler setup could change the decision.
Test before expanding
Use ten ordinary examples and several deliberately awkward ones: a missing field, a duplicate, an unsupported request, and a tool outage. Record the expected outcome before running the test. Expand only after a responsible person can explain the result and recover from mistakes.
Check your candidate’s readiness, then use the value calculator with conservative inputs.
Prepared by CCAIHelp for small teams evaluating AI and automation. Review the approach against your own process, data permissions, and current tool capabilities.
Suggest a correction ↗