How do I work out whether an AI automation will pay back?
Short answer: Take the hours the work takes today, subtract the hours it will still take once AI does its part (including a person checking the output and fixing mistakes), and multiply what is left by what an hour costs you. Then subtract the monthly running cost and divide the one-time build cost by the monthly saving. That gives you payback in months. Because early savings are almost always smaller than planned, cut the saving by half to two thirds for the first year, and only go ahead if the payback still looks acceptable.
You have a 20-hour-a-week job in mind for AI and want a number you can trust before you spend money on it.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
If you would rather have someone find the best first process for you, the AI Strategy Sprint does that in 3 days.
The short formula
ROI means return on investment: what you get back compared with what you put in. For an AI automation, which is software that does a repeat task for you, the full version has four parts. To set the price against the payback, see what AI implementation costs a small business.
- Hours saved per week, after subtracting the hours a person still spends on the task.
- What an hour costs you, including payroll taxes and benefits, not just the wage.
- Running cost per month: the software subscription, usage fees, and the time someone spends looking after it.
- Build cost, paid once.
Monthly saving = (net hours saved per week x 52 / 12 x cost per hour) minus running cost. Payback in months = build cost divided by monthly saving. Most online ROI calculators stop at step one and multiply hours by wage. That is where the numbers go wrong.
What counts as a saving
Three things can fairly be counted.
- Staff hours you stop spending. This only turns into money if you do something with the hours: avoid a hire, cut overtime, drop a contractor, or move the person to work that earns revenue. Ten hours "saved" that disappear into the day are not a saving yet.
- Errors you stop paying for. If wrong invoices, missed follow-ups or typos cost you refunds, rework or lost jobs, put a rough cost on one mistake and count how many you see a month. Count them from your own records, not from a vendor's claim.
- Speed that earns money. Answering an enquiry in five minutes instead of five hours can win jobs. This is the hardest to measure, so I keep it out of the main number and treat it as a bonus.
Treat any "hours saved" or "percent faster" figure from a vendor as the vendor's claim about its own product. Use it as a hypothesis to test, not as an input. Messy data hurts the return, so see whether your data is good enough for AI.
The costs most guides skip
A saving of "80 percent of the work" nearly always hides these.
- Review time. AI output needs a person to check it. Reading and approving a draft is faster than writing it, but it is not free, and it never drops to zero for work that reaches customers. The European Union's AI Act (Article 14, on human oversight) names a related risk, automation bias: people tend to trust a system's output too easily, so the check has to be a real check. NIST's AI Risk Management Framework, a voluntary US standards-body guide, lists measuring as one of its four core functions, which is the point of timing the work before and after. My guide to where to put the human check covers how to size it.
- Exceptions. Some items are odd, unclear or high stakes. A person still handles those from scratch, and they often take longer than a normal item.
- Rework. When the AI gets something wrong and it is caught late, someone fixes it and sometimes apologises to a customer.
- Upkeep. Prompts need tuning, tools change, connected systems update, and someone has to notice when quality slips. Plan on a small, steady amount of time every week.
- Running fees. Subscriptions and usage fees. Check what you already pay for with an AI subscription audit before adding more.
- Your own time to set it up. If the owner spends 15 hours on the pilot, that is a cost too.
What an hour really costs
A wage is not the cost of an hour. The US Bureau of Labor Statistics (BLS, the federal agency that publishes pay data) reports, in its Employer Costs for Employee Compensation release for June 2026, that wages and salaries were about 70 percent of private-sector employer compensation costs, with benefits and other costs making up the other 30 percent. For scale, BLS puts the median pay for secretaries and administrative assistants at $23.23 an hour (May 2025). Divide a wage by 0.7 and you get a rough fully loaded hourly cost. The US Small Business Administration also lists employer Social Security taxes, workers' compensation and unemployment insurance among the costs of employing someone beyond pay. Use your own payroll figures where you can.
Worked example: one 20-hour-a-week process
All figures below are made up to show the method. Swap in your own.
The process: a small firm spends 20 hours a week answering supplier and customer emails and updating records. A loaded hour costs $30 and the firm works 50 weeks a year.
| Line | Assumption | Result |
|---|---|---|
| Current cost | 20 hours x $30 x 50 weeks | $30,000 a year |
| Items AI can draft | 16 of the 20 hours of work | Review takes a quarter of that time: 4 hours |
| Exceptions | 4 hours, still done by hand | 4 hours |
| Hours after AI | 4 review + 4 exceptions | 8 hours a week |
| Gross hours saved | 20 minus 8 | 12 hours a week |
| Rework and upkeep | 1 hour + 1 hour a week | 2 hours a week |
| Net hours saved | 12 minus 2 | 10 hours a week |
| Net labour saving on paper | 10 x $30 x 50 | $15,000 a year |
| Running cost | $200 a month, tool plus usage | $2,400 a year |
| Build cost | One time | $6,000 |
On paper the payback looks great: $15,000 minus $2,400 is $12,600 a year, or $1,050 a month, so $6,000 pays back in about 5.7 months. That is the number a vendor would show you. It is also too good to plan around.
The first-year haircut
New automations start worse than the plan. The first weeks go to fixing prompts and learning where it fails, volumes are uneven, and people check more carefully at the start. I cut the paper saving by 50 percent as a normal case and by 70 percent as a cautious case. This is my own rule of thumb, not a published statistic, and I would rather be pleasantly surprised than short of cash.
| Case | Labour saving | Minus running cost | Per month | Payback on $6,000 |
|---|---|---|---|---|
| Paper | $15,000 | $12,600 | $1,050 | 5.7 months |
| 50% haircut | $7,500 | $5,100 | $425 | 14.1 months |
| 70% haircut | $4,500 | $2,100 | $175 | 34.3 months |
The haircut changes the answer from "obvious yes" to "not at this price".
A go or no-go payback rule
Here is the rule I use for a small business with limited cash. Go ahead if the payback is 12 months or less in the 50 percent case and 24 months or less in the 70 percent case. If it fails, do not give up on the idea. Shrink it.
In the example, the build fails the rule: 14.1 months is over 12. So narrow the scope to the one or two email types that make up most of the volume and cut the build to $3,000. The payback becomes about 7.1 months in the 50 percent case and 17.1 months in the 70 percent case. That passes. In practice a narrower build saves a little less, so rerun the table with the smaller saving before you commit. The point is the method: change one input at a time and watch which one moves the answer most. Usually it is build cost, review time or volume.
When ROI is the wrong test
A payback table suits repeat work with countable hours. It fits poorly when:
- the process runs only a few times a month (the saving is too small to matter);
- the main benefit is not missing things, such as never losing a lead, which is real but is better measured as lost jobs per month than as hours;
- the work needs expert judgement every time, so review time eats the saving.
For the second case, estimate the value of one missed item (an average job's margin) and how often it happens today. It is still a number, just a different one.
How to measure it for real, step by step
- Time the current work. Record 30 real items: minutes per item, and how many needed rework. Ask the person doing it, then check against a stopwatch. People guess low about their own admin.
- Write the table above with your numbers and both haircut cases.
- Run a small pilot before building. My 30-day AI pilot plan gives you the weeks and the stop rules.
- Replace guesses with measurements. After the pilot, use the real review time, the real exception rate and the real cost of mistakes.
- Re-run the payback with measured numbers and apply the rule.
- Check again at three months after launch. Compare actual hours and errors with the plan.
If you are still choosing which process to automate first, the AI Strategy Sprint is a paid 3-day first step that finds where AI could help most, with the exact steps written down.
Common mistakes
- Counting 100 percent of the task as saved when a person still reviews it.
- Using a wage instead of a loaded hourly cost.
- Counting freed hours as cash when nothing changes in the payroll or the schedule.
- Ignoring upkeep, then finding the automation quietly broke two months ago.
- Comparing the build cost with year-one savings only, when a good automation lasts longer. Keep the payback rule, then look at the three-year picture as a second check.
For what a custom build costs, see my guide to custom software development cost. For a worked set of numbers in one industry, see the cost of AI automation for a Dubai brokerage. The wider picture is in AI implementation, AI workflow automation for small business and business process automation.
How I'd approach this
I would never sign off on an automation from the vendor's number. I would time the current work, build the table with both haircuts, and refuse anything that only works on paper. If the payback fails, I shrink the scope before I drop the idea, because a narrow automation on your busiest repeat task usually beats a broad one. I would also decide in advance who owns the number after launch, so someone checks the hours and the errors at three months.
Blur Studio fixes and finishes start around $1,500, which is one reason small, focused builds often pass the payback rule when a large one does not. I will tell you on a call if the answer is "do not automate this".
Questions people ask
What is a good ROI for an AI automation?
There is no standard figure. I use a payback rule instead: 12 months or less in a 50 percent first-year haircut case and 24 months or less in a 70 percent case. That suits a small business watching cash.
Is there an AI automation ROI calculator I can use?
A spreadsheet with the lines in the worked example is enough: current hours, hours after AI including review, rework and upkeep, loaded hourly cost, running cost and build cost. Most online calculators leave out review and upkeep, so check what a calculator includes before trusting its answer.
How do I calculate the ROI of an AI agent?
Use the same method. An AI agent is software that takes steps on its own toward a goal (see what an AI agent is). The more freedom it has, the more review and exception time you should expect, so measure it in a pilot first.
Should I count revenue growth in the ROI?
Count it separately and conservatively. Faster replies or fewer missed leads can earn money, but they are harder to prove. Base the go or no-go decision on cost savings, and treat extra revenue as a bonus.
How long before I know if it is working?
A 30-day pilot shows whether the idea holds. A real measure of hours and errors takes about three months after launch, once the early fixing is over.
What if my payback fails the rule?
Narrow the scope, lower the build cost, or pick a process with more volume. If none of those work, leave it as a manual task. Not automating is a valid result.
More in this topic: Problem solving: AI strategy and adoption (24 guides) →