To calculate the ROI of AI automation, measure what the task costs today, estimate how much the automation will remove, then subtract every cost of building and running it, including human review and maintenance. ROI is net benefit divided by total cost. Payback period is build cost divided by monthly net savings. Always model a pessimistic case too.
Most automation business cases go wrong in the same two places: they overestimate how much work disappears and they forget the running costs. This guide walks through a method that avoids both, with an illustrative worked example you can copy. It pairs with the project-selection advice in our practical AI automation playbook.
Step 1: What exactly are you automating?
Define one task with a clear start and end, not a department. "Automate support" cannot be measured. "Draft first replies to order-status emails" can.
Write down:
- The trigger (an email arrives, a form is submitted, an invoice is received).
- The steps a person takes today.
- The output (a reply sent, a record created, data entered).
- Who does it and how often.
If you are not sure which task to pick, our list of 25 repetitive business tasks you can automate with AI includes a scoring table for choosing a starting point.
Step 2: How do you measure the baseline?
Measure for at least one or two typical weeks before you build anything. Estimates from memory are almost always wrong, usually in both directions.
Collect:
- Volume: how many times the task happens per week.
- Time per instance: time a sample of 20–30 occurrences, including switching between apps, not just typing.
- Loaded hourly cost: salary plus benefits, taxes, equipment and overheads, divided by productive hours. This is higher than the raw wage.
- Error cost: how often mistakes happen and what each one costs to fix (a wrong invoice, a missed follow-up, a refund).
- Delay cost: whether slowness loses revenue, for example enquiries that go cold.
The delay cost is often the biggest number and the hardest to pin down. If you pay for leads, you already know what each enquiry costs. Across five wedding photography ad accounts we manage, for instance, the average was ₹58.55 per enquiry. Slow replies waste that spend, but how much revenue faster replies recover depends on your sales process, so measure it rather than assume it.
Step 3: How much of the work will automation actually remove?
Assume partial removal, not elimination. AI usually handles the bulk of routine cases while people still review outputs, handle exceptions and fix errors.
Be honest about:
- Exceptions: what share of cases are unusual enough to need a person anyway?
- Review time: if every output needs a quick check, that time is a cost.
- Adoption: will the team actually use it, or keep doing things the old way?
A useful habit is to estimate three reduction levels: optimistic, expected and pessimistic. If the project only works at the optimistic level, it is a risky investment.
Step 4: What are all the costs?
Count four kinds of cost: build, run, maintain and supervise. Leaving any out flatters the result.
- Build: discovery, development, integration and testing. Typical market ranges are roughly US$6k–20k (₹1.5–5 lakh, A$10k–30k) for a simple internal tool or automation, US$3k–10k for an FAQ or rules chatbot, US$10k–40k for an AI chatbot on your own documents, and US$40k+ for an AI agent with integrations. These are estimates; scope decides the real number.
- Run: model usage, hosting, and any platform subscriptions.
- Maintain: roughly 15–25% of the build cost per year for updates, model changes and fixes as connected systems change.
- Supervise: staff time reviewing outputs and handling escalations.
Platform choice changes the running cost a lot, especially at volume. Our comparison of Zapier vs Make vs n8n vs custom automation explains how per-step and per-execution pricing affect this line.
Step 5: Which formulas should you use?
Three numbers are enough for a decision: annual net benefit, ROI and payback period.
- Annual benefit = (hours saved per year × loaded hourly cost) + error costs avoided + revenue recovered
- Annual cost (year 1) = build cost + run costs + supervision time
- ROI = (annual benefit − annual cost) ÷ annual cost
- Payback period (months) = build cost ÷ monthly net savings, where monthly net savings = (annual benefit − run costs − supervision) ÷ 12
Year 2 usually looks much better than year 1 because the build cost is behind you, though maintenance is added.
What does a worked example look like?
Here is a hypothetical invoice-entry automation, using made-up but plausible inputs to show the method. Replace every figure with your own measurements.
Scenario: three finance staff each spend about 10 hours a week keying supplier invoices into accounting software. Loaded cost is US$25 an hour. A custom extraction workflow with a review screen is quoted at US$12,000. Model usage is about US$100 a month. Someone spends 2 hours a week reviewing flagged invoices. The team works 48 weeks a year.
| Line item | Expected case (70% of entry time removed) | Pessimistic case (40% removed) |
|---|---|---|
| Hours saved per week | 21 | 12 |
| Hours saved per year (48 weeks) | 1,008 | 576 |
| Annual benefit at US$25/hour | US$25,200 | US$14,400 |
| Build cost (one-off) | US$12,000 | US$12,000 |
| Model usage per year | US$1,200 | US$1,200 |
| Review time per year (96 hours) | US$2,400 | US$2,400 |
| Year 1 total cost | US$15,600 | US$15,600 |
| Year 1 net benefit | US$9,600 | −US$1,200 |
| Year 1 ROI | about 62% | about −8% |
| Monthly net savings | US$1,800 | US$900 |
| Payback period | about 6.7 months | about 13.3 months |
| Year 2 net benefit (adds 20% maintenance, US$2,400) | US$19,200 | US$8,400 |
What the example shows:
- The expected case pays back within the first year; the pessimistic case does not, but still turns positive in year 2. That is a reasonable investment with manageable downside.
- If the pessimistic case stayed negative in year 2, you would want a cheaper approach, such as the invoice capture built into your accounting software, or a narrower first phase.
- The numbers ignore error reduction and faster supplier payments. If those matter in your business, add them, but only with measured figures.
What benefits are hard to put a number on?
Some of the best reasons to automate do not fit neatly in a spreadsheet; list them separately rather than inflating the main figures.
- Faster response to customers and leads.
- Staff spending time on more valuable work, which can help retention.
- Capacity to grow volume without hiring at the same rate.
- Better data, because structured records replace notes and memory.
Treat these as tie-breakers. If the hard numbers are roughly break-even, these can justify going ahead. If the hard numbers are clearly negative, they rarely should.
How do you check the ROI after launch?
Re-measure the same baseline numbers two to three months after go-live. Compare actual time per task, volume handled, review time and error rates against your expected case.
If results fall short, find out why before expanding. Common causes are more exceptions than expected, staff not trusting outputs, or a messy input source. Each has a different fix.
Frequently asked questions
What is a good ROI for an AI automation project?
There is no universal benchmark, and it depends on your cost of capital and risk appetite. A practical test is whether the project pays back within a year or two in your expected case and stays reasonable in your pessimistic case. Shorter payback periods mean less risk.
Should I include staff time saved if I'm not reducing headcount?
Yes, if that time goes to useful work such as more sales calls or faster delivery. Be realistic, though: saved minutes scattered across a week do not always turn into productive hours. Note what the freed time will be used for.
How do I estimate costs before getting a quote?
Use typical ranges for the type of build as a first pass, then refine with a quote after a discovery session. Our guide to what a custom AI chatbot costs breaks down one common category in detail.
Is off-the-shelf software usually better ROI than custom?
For standard processes at low volume, often yes, because there is little build cost. Custom tends to win when the process is specific to your business, volumes are high or subscription fees scale steeply with usage.
How long should I measure the baseline?
At least one or two representative weeks, avoiding unusually quiet or busy periods. For seasonal businesses, take samples from both peak and off-peak months if you can.
Build the business case first
A two-hour baseline exercise can save you from an automation that never pays back, or give you the confidence to go ahead with one that clearly will. If you would like help measuring a task and modelling the numbers, ask us for a fixed-price scope and we will include the ROI model with it.