Lesson 12 / 12 · 10 min
Choosing what to automate: an ROI framework
Not every task should be an agent. A filter for picking the right workflow, the ROI math, and how to sequence your first build.
The fastest way to waste money on AI is to automate the wrong thing: a task that is rare, messy, or high-stakes. A clear filter saves you that, and points you at the work where an agent actually pays for itself.
The filter: five questions
- Is it high volume? Leverage comes from frequency.
- Is it repetitive? The same shape of task each time.
- Is it rule-ish? There are mostly clear right answers.
- Are mistakes cheap, or easy to gate behind a human?
- Can you measure success? If you cannot score it, you cannot trust it.
The more yeses, the better the candidate. A task that is rare, ambiguous, and irreversible is exactly where an agent will embarrass you. A task that is frequent, repetitive, and low-stakes is where it quietly saves you hours every week.
Where agents win, where humans win
Agents win on volume, speed, consistency, and tireless follow-up. Humans win on judgment, relationships, novel situations, and high-stakes calls. The best designs are not agent-versus-human, they are agent-handles-the-bulk and a person handles the edge cases the agent flags. Aim for that split, not full replacement.
The ROI math
Estimate the value, then the cost. Value is roughly time saved per task times tasks per month times the cost of that time, plus revenue you save by responding faster. Cost is the one-time build plus the monthly run (management fee plus AI usage on your own key). If value clears cost with margin, it is worth building. Most viable cases are not close, they are obvious once you put numbers on them.
Candidate: WhatsApp first-reply + qualify
volume ~600 chats / month
time saved ~4 min / chat handled
saved hours 600 x 4 / 60 = 40 hours / month
value 40 h x Rp 50k = Rp 2.000.000 / month (time)
+ faster reply = more booked jobs (upside)
cost build (one-time) + monthly run
verdict clears cost with margin -> build itSequence it
Pick a first workflow where volume is high and failure is cheap. Win there, collect the data and the team's trust, then expand to higher-stakes work. Starting with the hardest, riskiest process is how pilots die. Earn the right to automate the important things by proving it on the safe ones first.
What you take away
- Filter candidates: high volume, repetitive, rule-ish, low or gated stakes, measurable
- Design for agent-handles-bulk, human-handles-edge-cases, not full replacement
- Do the ROI math: time saved x frequency x cost of time, minus build and run
- Start where volume is high and failure is cheap, then expand to harder work
Curated sources
Hand-picked references for going deeper, reviewed by our team.