How to calculate the ROI of AI automation
Most AI business cases only count the upside. A credible one counts the hours honestly, values them correctly and subtracts every cost. Here is the method, step by step, with a worked example you can reproduce in the calculator.
The formula
Each input deserves care. Small errors in hours or time-saved percentages multiply through everything else.
Step 1: Baseline the hours — measure, don't guess
List the specific tasks you want to automate, not job titles. "Drafting replies to routine customer emails" is measurable; "customer service" is not. For each task, estimate hours per person per week. Better still, have the team log time on those tasks for one or two normal weeks. People routinely misjudge how long repetitive work takes, in both directions.
Only include work that genuinely changes. If an AI tool drafts emails but every draft still needs careful checking, the reviewing time stays in the baseline.
Step 2: Use the fully loaded hourly cost
An hour of an employee's time costs more than their hourly pay. Add employer contributions, benefits, equipment, software, office space and management overhead. A common rule of thumb is base hourly pay × 1.25 to 1.4. If a team member earns the equivalent of $30 an hour, a loaded cost of roughly $38–42 is a reasonable planning figure.
Step 3: Choose a defensible time-saved percentage
This is where most business cases go wrong. Published studies of AI assistance on real work have found productivity gains from about 14% to 25% on specific tasks, with much bigger gains for less experienced staff — and at least one rigorous trial found experienced professionals got slower. See what the research says for details.
- 20% — conservative; suitable for complex or judgement-heavy work.
- 35% — moderate; routine drafting, summarising and formatting.
- 50%+ — only for highly repetitive, rule-based tasks, or where you have pilot data.
Then apply an adoption rate. Not everyone will use the tool, and not every time. 70–90% is a realistic planning range for a supported rollout.
Step 4: Count every cost
Separate one-off setup costs (design, integration, data cleanup, training, the pilot itself) from running costs (licences per user, usage-based API charges, maintenance, support). Our guide to the hidden costs of AI automation lists the ones most often missed.
Step 5: Allow for ramp-up
Nobody is fully productive with a new tool on day one. The calculator counts benefits at half value during a ramp-up period — two months by default — which pushes payback out realistically.
Worked example: a five-person admin team
| Input | Value |
|---|---|
| People doing the tasks | 5 |
| Hours per person per week on target tasks | 13 (email 4, reports 3, data entry 3, research 2, meeting notes 1) |
| Fully loaded hourly cost | $40 |
| Working weeks per year | 46 |
| Time saved on those tasks | 35% |
| Adoption | 80% (4 licensed users) |
| Ramp-up | 2 months at half benefit |
| Setup cost | $5,000 |
| Licences | $30 per user per month |
| Other running costs | $150 per month |
| Result | Value |
|---|---|
| Hours saved per week | 18.2 |
| Hours saved per year | 837 |
| Value of time per year | $33,488 |
| Running costs per year | $3,240 |
| Year-1 ROI | 273% |
| Payback | 4 months |
| 3-year net value | $82,953 |
The project looks clearly worthwhile — but notice how sensitive it is to the time-saved assumption. At 20% the three-year net value falls to $41,093; at 50% it rises to $124,813. That spread is the reason to pilot before scaling.
Step 6: Validate with a pilot
- Pick one task with high hours and clear rules (see which tasks to automate first).
- Measure the baseline for one to two weeks.
- Run the AI-assisted process for two to four weeks with a few volunteers.
- Measure time per task and quality — error rates, rework, customer complaints.
- Re-run the calculator with the measured figures before committing the wider budget.
Turn hours into money — deliberately
Value of time is not cash in the bank. It becomes real only if the freed hours are redeployed: more work handled without hiring, overtime cut, faster turnaround that wins business. Write down how the time will be used before the project starts, then check whether it happened.
Sources
- Brynjolfsson, Li & Raymond, "Generative AI at Work", NBER Working Paper 31161 (2023)
- Dell'Acqua, McFowland, Mollick, Lakhani et al., "Navigating the Jagged Technological Frontier", Harvard Business School working paper (2023), as reported by The Harvard Crimson
- METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (July 2025)