The strongest first projects are valuable enough to matter and controlled enough to test. High frequency alone is not enough; the workflow also needs understandable inputs and a safe recovery path.
/ 01
Good first candidates
Start with work that is repetitive, document-heavy, and easy to review.
- Structuring form or email intake
- Drafting follow-up from approved templates
- Summarizing internal documents
- Moving approved data between systems
- Preparing recurring reports
- Classifying requests for human routing
/ 02
Projects that need more control
Healthcare, legal, employment, lending, safety, financial movement, and access-control decisions deserve additional expertise and oversight. The presence of AI does not remove existing responsibilities.
/ 03
Use a simple scoring model
Score each idea from one to five for frequency, manual effort, input consistency, output reviewability, reversibility, and business value. Then score consequence, data sensitivity, permission breadth, and exception variability from one to five.
Use a transparent formula: opportunity = frequency + effort + consistency + reviewability + reversibility + value. Risk = consequence + sensitivity + permission breadth + variability. Rank by opportunity minus risk, but treat any consequence score of five as a required governance review rather than something an average can erase.
/ 04
Do not automate a broken policy
If the team cannot agree on what should happen, automation will reproduce the disagreement faster. Resolve ownership, authoritative data, rules, and exceptions before building. If people repeatedly override the stated procedure for legitimate reasons, study those overrides; they contain the real policy.
/ 05
Run a 45-minute opportunity workshop
Invite the process owner, one frequent user, and the person accountable for risk or customer impact. List no more than ten candidates, score them independently, then discuss the disagreements. The disagreement is useful: it exposes hidden effort, risk, and exception knowledge.
Select one workflow and write a one-page charter: trigger, inputs, output, owner, human checkpoint, excluded actions, representative test data, success metrics, and stop condition. Do not select a second active workstream until the first produces credible evidence.
- 10 minutes: list recurring work
- 10 minutes: score opportunity
- 10 minutes: score risk
- 10 minutes: discuss score disagreements
- 5 minutes: name one owner and next evidence-gathering step
/ Practical takeaway
Prioritize a frequent, reviewable, reversible process with one owner. Avoid starting with your most consequential decision simply because it is painful.