The frame: impact × effort × reversibility
Every AI opportunity gets three scores. Impact on the unit you care about (operator capacity, revenue, retention, decision quality). Effort in calendar weeks to first value, including adoption. Reversibility — how cheap is it to undo if it doesn't work? Anything irreversible needs a higher impact bar.
Step 1 — Generate the candidate list from workflows, not from tools
Start from a workflow you actually run, not a tool you saw. For each meaningful step in the workflow, ask: where could AI compound this? This produces opportunities rooted in operating reality instead of vendor positioning.
Step 2 — Score impact in the units your business measures
Don't score in "AI-ness." Score in operator hours back per month, in expected revenue or retention lift, in decision latency reduced from weeks to days. If you can't translate an opportunity into one of those units, it isn't ready to be on the list yet.
Step 3 — Score effort honestly, including adoption
Most AI projects under-estimate adoption. A two-week build often needs six weeks of change management. Include the people work: training, documentation, role redefinition, the conversations with the team members whose work is changing. If you skip this, your prioritization will look great on paper and fail in production.
Step 4 — Sequence by dependency, not by score
Some of the highest-impact opportunities sit on top of redesigns that haven't happened yet. Sequence the redesigns first. A great opportunity blocked on poor data is not a great opportunity; it's a great opportunity after the data layer is fixed.
Step 5 — Limit the backlog to what can ship in a quarter
Most companies pick too many. Pick three. Ship them well. Measure them honestly. Then pick the next three. A working backlog of three beats a beautiful backlog of thirty every time.
Anti-patterns to refuse
- Tool-first prioritization. Picking opportunities to justify a license.
- Demo recency bias. The vendor who pitched last week wins.
- Wishlist length as ambition. A 40-item backlog is a way of avoiding execution.
- Skipping the redesign. Automating a frictioned workflow accelerates friction.
- Optimizing for output, not outcome. "We shipped 12 automations" is not the goal.