The short answer
ROI is the cost of the hours an agent gives back, minus the cost of the exceptions, reviews, upkeep and oversight it creates, minus software and model costs, set against one-time build and parallel-run costs. Measure today's time before you start.
The inputs
- VolumeItems per monthWhere it comes from: System report
- Handling timeHands-on minutes per item todayWhere it comes from: Stopwatch, before the pilot
- Exception rateShare of items routed to a personWhere it comes from: Pilot runs
- Exception timeMinutes per exceptionWhere it comes from: Timed in the pilot
- Review shareShare of clean items still checkedWhere it comes from: Your review policy
- UpkeepHours on rule edits and screen changesWhere it comes from: Run history
- OversightHours reading results and alertsWhere it comes from: Calendar
- Hourly costLoaded cost of whoever does the workWhere it comes from: Payroll
- SoftwareMonthly tool and model costWhere it comes from: Quotes and usage
- One-timeBuild hours, parallel run, feesWhere it comes from: Pilot plan
Five formulas
-
Baseline hours = volume x minutes per item / 60.
-
Exception hours = volume x exception rate x exception minutes / 60.
-
Review hours = clean items x review share x review minutes / 60.
-
Hours returned = baseline hours, minus exception, review, upkeep and oversight hours.
-
Payback (months) = one-time cost / (monthly labor savings minus software). Only valid when that number is positive.
Hours are planning estimates until real runs replace them.
The stopwatch comes first
The whole case rests on handling time. A guess from memory is not a measurement. Time 20 consecutive items across two people, and count hands-on minutes only, not time spent waiting in a queue.
An overstated baseline hurts the case more than a doubled exception rate. In a typical scenario, cutting baseline time by a third cuts savings nearly in half, while doubling exceptions cuts it by about a quarter.
Baseline time
Exception rate
Review share
Upkeep
Never zero
Upkeep and oversight belong in month one. Vendors redesign screens, thresholds change, new exception types show up. Unattended is not unobserved: someone reads the morning summary and clears the exception list.
The parallel run is a real cost too. For weeks, a person does the work and also checks the agent.
Model cost moves
Price per token keeps falling, but spend per task often doesn't, because newer agents use more tokens per job. Keep software and model cost as its own line, check it quarterly, and send every step with one right answer to plain rules, where it costs nothing to run.
What not to count
Error savings, unless you measured the old error rate and cost per error
Hours that just move to a supervisor
Faster cycle time as dollars, unless it captures a specific discount or avoids a specific fee
A vendor's case study in place of your own stopwatch
The process owner's build time as "free"
Hours returned are capacity, not cash, until someone decides how to use them.
Measured, not guessed
LaunchAI's run dashboard tracks success rate, exceptions, pending reviews and hours returned per workflow, so your pilot estimates get replaced by real numbers. Product →
Not sure which workflow to price first? Which Workflow First →
Questions
Payroll data or national averages?
Your payroll whenever you have it. National averages are fine for a first rough pass.
How do I count an avoided hire?
Count the loaded cost of that role from its planned start date, and stop counting hours returned for the same work. Counting both is double counting.
Keep learning
- Which Workflow First Start with a workflow that is frequent, slow by hand, low in variation, spread across a few systems, cheap to get wrong and rarely off the happy path. Read the guide
- Agent-Ready SOP Put one action on each line. Read the guide
- Human in the Loop Put a person where a wrong action is expensive and hard to undo, and where the agent's evidence is weak. Read the guide
No prompts. No babysitting. No Integrations.