What should you automate first? An AI automation prioritization formula for PMs
TL;DR
List your team's repetitive work → score 4 criteria (frequency × time per run, how rule-defined it is, risk when wrong, data readiness) → pick 1–2 high-score, low-risk items for a 2-week pilot → measure real hours saved before scaling.
After the first AI demo, every team has a list of "we could automate this and that". The problem is limited resources: pick the wrong first task and you spend weeks building something that saves a few minutes a week, and leadership loses faith in the whole programme. This post gives a simple scoring formula for choosing what to automate first, how to measure ROI and how to run a pilot.
Step 1: List repetitive work
Ask everyone on the team: "Which task do you do over and over every week and find boring?" Record the task, its frequency, the time per run and who does it. Don't filter at this stage; the goal is a list of 15–30 tasks.
Step 2: Score (1–5)
| Criterion | High score when… |
|---|---|
| Volume | Repeated daily, each run takes ≥ 10 minutes |
| Clarity | Inputs/outputs can be defined, sample examples exist |
| Cost of errors | Mistakes are easy to fix and don't affect customers or money (high score = low risk) |
| Data | Already digitised, accessible via API |
Add the four scores, or multiply volume by the sum of the other three if you want to favour big tasks. A task scoring 1–2 on cost of errors shouldn't be the first pilot, even with a high total.
Example in a software team
| Task | Assessment | Decision |
|---|---|---|
| Writing the weekly progress report | Repetitive, data already in Jira/Git | ✅ Do it now |
| Classifying support tickets | High volume, a human reviews hard cases | ✅ Pilot |
| Auto-replying to customers about refunds | Financial risk | ⚠️ Drafts only |
| Deciding roadmap priorities | Needs strategic judgement | ❌ Keep it human |
General rule: AI drafting and suggesting is acceptable for most tasks; AI deciding and executing should be reserved for low-risk work, or work with a human approval step.
Step 3: Estimate ROI simply
Hours saved / month = runs × minutes per run × automation success rate / 60
Cost / month = API tokens + infrastructure + maintenance hours
Break-even (months) = build effort / (value of hours saved − monthly cost)Two items often forgotten: maintenance hours (prompts need tuning, APIs change, users give feedback) and the automation success rate (the remainder still needs a human). Example: a weekly report takes 2 hours, 80% automated → saves about 1.6 hours a week. If it takes 12 hours to build, it breaks even after about 8 weeks.
Step 4: A 2-week pilot
- Pick 1–2 high-scoring, low-risk tasks.
- Record manual hours before automating (baseline).
- Run alongside humans in week one (shadow mode) to compare results.
- In week two, switch to AI drafting with human review.
- At the end, measure real hours saved, rework rate and token cost.
Common mistakes
- Picking "cool" tasks over "painful" ones: an impressive demo doesn't mean hours saved.
- Automating something too rare: a once-a-quarter task rarely breaks even.
- No baseline: without a "before" number you can't prove the "after".
- Ignoring users: a good tool nobody uses has zero ROI.
How to apply it
- PM: use the scoring table as the meeting document with your team; choose the pilot by score rather than opinion.
- Tech Lead: assess technical difficulty in parallel; a high-scoring task that needs complex integration may go later.
- Leadership: report in measured hours saved and actual cost.