The ROI Math We Actually Use Before Recommending Automation

Clients often come to us with a task in mind they want automated — "can you build a bot for this?" — before anyone has checked whether it's actually worth automating. Sometimes the honest answer is: leave it alone. Here's the framework we use to figure out which is which.

Step 1: Measure the real cost of the manual process

Not a guess — an actual number. Time per instance, times frequency, times fully-loaded hourly cost. A task that "feels" annoying but happens twice a month rarely clears the bar; a five-minute task done 200 times a day almost always does.

Monthly cost of manual process =
  (minutes per instance / 60) × hourly cost × instances per month

Step 2: Estimate the automation's real success rate

No automation is 100% reliable. A realistic estimate — usually 80–95% depending on task complexity — matters because the remaining cases still need a human, and that exception-handling cost has to be built into the math, not ignored.

Step 3: Add the ongoing cost of the system itself

Automation isn't free to run. Factor in:

• API/model costs at your actual volume
• Monitoring and occasional maintenance
• The cost of the initial build, amortized over a realistic lifespan (we typically use 18–24 months)

Step 4: Compare, honestly

If the automated column doesn't clearly beat the manual one within a reasonable payback window — we tell clients that directly, even when it costs us the project. A task that's borderline today is often worth revisiting in six months once volume grows, and we'd rather earn that follow-up conversation than oversell an automation now.

The tasks that actually clear the bar

In practice, the automations with the clearest ROI share three traits: high frequency, low ambiguity (clear rules, not constant judgment calls), and a cost of error that's tolerable if something occasionally needs human review. Document intake, data entry between systems, and routine customer questions hit all three consistently. Highly judgment-heavy decisions — the ones with real consequences for getting it wrong — usually stay human-led, with AI assisting rather than deciding.