Automation ROI compares the value a process improvement produces with its full cost over an agreed period. Start with measured work, include the effort that remains after automation, and distinguish time released from money the business will actually stop spending.
What should you measure before estimating ROI?
Measure transaction volume, active handling time, exception work and the current definition of a completed outcome.
Choose a representative operating period and include normal and difficult cases. Time spent waiting in an approval queue is cycle time; it is not automatically labour time. Record both because a faster customer response can matter even when the financial saving is modest.
Give the process owner a short baseline sheet: volume by case type, handling minutes, rework, systems involved and the person responsible for each exception. Document how the sample was taken. A memorable worst-case task should not become the average for every transaction.
If the process has not yet been selected, evaluate ownership and data readiness before calculating a return. An attractive spreadsheet cannot make an undefined process implementable.
Which costs belong in the business case?
Include implementation, recurring operation and the human work the new process still requires.
One-time costs can include discovery, data cleanup, integration, test cases, training and transition support. Recurring costs can include platform fees, model usage, hosting, monitoring, maintenance and vendor changes. Human review and exception handling are operating costs too; count them once, consistently.
Compare realistic options: improve the manual process, configure an existing product, build a deterministic workflow, or add AI where interpretation is necessary. Estimate value for the least complex option that can meet the need before pricing a more ambitious system.
- State the period and currency used in the calculation.
- Keep one-time and recurring costs separate.
- Show what review and exception effort remains.
- Identify which benefits are cash, capacity or service quality.
- Record assumptions that require a pilot to validate.
What does a worked example look like?
Use transparent assumptions and carry the remaining manual work through the calculation.
This is a hypothetical order-review process, not a PROMTURE customer result or a savings promise. Assume 1,000 cases per month, six active minutes per case before automation, and a loaded labour value of £30 per hour. After automation, assume 800 cases need one minute each and 200 exceptions still need six minutes each.
The baseline is 100 hours per month. The new process needs about 33.3 hours: (800 × 1 + 200 × 6) ÷ 60. It releases about 66.7 hours, valued at £2,000 of monthly capacity. With £600 in recurring non-labour costs, net monthly capacity value is £1,400.
If implementation costs £8,400, simple capacity-value payback is six months after reaching steady operation. Over twelve steady-state months, the capacity-value return is (12 × £1,400 − £8,400) ÷ £8,400 = 100%. This excludes ramp-up, financing, taxes and discounting; include the relevant factors in an actual investment decision.
| Case | Hours released / month | Net capacity value / month | Simple payback |
|---|---|---|---|
| Conservative: 600 cases at 2 min; 400 at 6 min | 40.0 | £600 | 14 months |
| Base: 800 cases at 1 min; 200 at 6 min | 66.7 | £1,400 | 6 months |
| Improved: 900 cases at 1 min; 100 at 6 min | 75.0 | £1,650 | About 5.1 months |
Interactive estimate
Test your own assumptions
Adjust the example to reflect the work, review and operating costs in your process.
Process assumptions
Scenario estimate
What the assumptions suggest
- Hours released each month
- 66.7 hrs 100 hrs current · 33.3 hrs after automation
- Net monthly capacity value
- £1,400 £2,000 capacity value less £600 ongoing cost
- Simple capacity-value payback
- 6 months Uses steady-state net monthly capacity value.
- Capacity return over 12 months
- 100% £8,400 net capacity value over this period.
This is a scenario, not a forecast or a PROMTURE customer result. The starting assumptions match the article’s hypothetical example. Replace them with measured data. Released capacity is not payroll savings unless spending changes or the time produces evidenced additional value. Ramp-up, tax, financing, quality changes and benefits not entered here are excluded.
Estimate updated. 66.7 hrs released each month; net monthly capacity value £1,400; simple capacity-value payback 6 months.
When does released time become a financial saving?
Only when the business changes spending or uses the capacity to create an evidenced economic benefit.
If the same team receives the same pay and no additional valuable work is completed, 66.7 released hours are not £2,000 of cash saved. They may still reduce pressure or improve service. Report that benefit honestly instead of labelling it payroll reduction.
A cash case might involve reduced overtime, avoided contractor spend or a demonstrable hiring requirement that no longer arises. Revenue benefits need a separate contribution-margin assumption and evidence that the extra work can be sold and delivered. Do not count the same hour as both a cost saving and new revenue without explaining the allocation.
Track quality separately: errors, reversals, missed service expectations and customer complaints can change the result. They are especially important when a model interprets ambiguous requests.
How should the pilot change the estimate?
Replace assumptions with observed outcomes before increasing the commitment.
Run a bounded pilot with an agreed period, process owner and failure path. Compare equivalent case types and include unsuccessful attempts, review time and rework. Record adoption: a workflow that people avoid will not deliver its theoretical throughput.
For AI tasks, inspect whether the intended outcome was achieved, not merely whether the assistant produced a plausible response. Anthropic's agent-evaluation guidance makes this distinction explicit. Use that principle to define a successful business transaction and the evidence required to count it.
Agree a decision before the pilot starts: expand if the net value and quality thresholds hold, improve the weak step if the gap is fixable, or stop if the economics no longer work. Keep ownership, monitoring and maintenance in the operating plan after launch.
Sources: Anthropic — Demystifying evals for AI agents, January 2026
Common questions
Should every automation project have a short payback period?
No single threshold fits every business. Mandatory controls, service commitments and strategic capability can matter alongside return. State those reasons explicitly and compare the costs of realistic alternatives.
Can a small business use this method?
Yes. A small team can start with a representative sample, an hourly capacity value and a list of ongoing costs. The method is useful because the assumptions are visible, not because the spreadsheet is complicated.