How to work out the return on an AI project

Owners often ask whether an AI project will pay back. The sums aren't hard, and you can do most of them before you talk to anyone, including me. The numbers below are made up for illustration, not results from any project. Swap in your own.

I won't promise a result. If the sums say no, I'll tell you so.

The method in five steps

1. Pick one task. One job, such as writing quotes or entering supplier invoices, rather than "AI in the business".
2. Measure the time it takes now, in hours per week. If you're not sure, track it for two weeks.
3. Put a value on an hour: a wage, or what you could earn doing billable work instead. This is a time value. It becomes cash only if the freed hours reduce paid hours or are spent on paid work. If the hours saved wouldn't turn into anything useful, they aren't worth much.
4. Add up all the costs: the build or setup, the running costs, and the costs people forget (below).
5. Work out the payback period, which is how many months until the savings cover the costs.

A worked example (hypothetical numbers)

Picture a small trades business (hypothetical, not a client or an observed result). The owner spends 5 hours a week on quoting and wants to know if automating most of it makes sense. These are the assumptions, and you can change any of them.

• Time spent today: 5 hours a week.
• Value of an hour: $60, treated as a time value and not as cash received.
• Working weeks per year: 46, allowing for holidays and slow weeks.
• Share of that time the tool removes: all 5 hours. That's generous, and I come back to it below.
• Build cost: $6,000. Made up for the sums, not a quote. A real figure comes from the builder's written quote before work starts.
• Your own time to explain the job and test the result: 20 hours, valued at $60.
• Hosting and upkeep: $50 a month.
• Your time checking the output: 30 minutes a week.

On these assumptions the value of the time freed covers the costs in about seven months, leaving roughly $4,620 of time value ahead after the first year. That is not cash in the bank. Of the costs, only the $6,000 build and $600 a year of hosting are cash outlays. The rest is the owner's time valued at $60 an hour. Realised cash savings arise only if the freed time reduces paid hours or is used for billable work. Check each line. If one of my assumptions doesn't match your business, change it and redo the sum.

The costs people forget

Most back-of-envelope sums leave out the same few items, which is how a good-looking return ends up smaller than expected.

• Your time at the start. Explaining how the job really works and testing the result takes hours, and they're yours.
• Checking the output. A person has to look at what AI produces, and that review time never goes to zero.
• Hosting and subscriptions. Usually small, but they run every month.
• Upkeep when things change. A connected system updates or your process changes, and the tool needs adjusting.
• Showing staff how to use it. Even a simple tool needs a short walk-through.
• The cost of mistakes. If a wrong quote costs you a job or a wrong invoice annoys a customer, put that on the list.

For more on what drives build costs, see what custom software costs in NZ now that AI writes the code.

Someone has to check what the AI produces, so count that time. It never goes to zero.

When the answer is "don't bother"

Same method, different job. Suppose the task takes 1 hour a week. The gross value of the time is 1 hour x $60 x 46 weeks = $2,760 a year. Keep the $600 hosting and the $1,380 checking time to be cautious, and the net value is $2,760 minus $600 minus $1,380 = $780 a year. Against an up-front cost of $7,200, payback is $7,200 / $780, about 9.2 years, even before asking whether those hours become cash. That doesn't pay.

Leave it alone when:

• The task takes little time today.
• It happens rarely, so there's little to automate.
• The saved time wouldn't be used for anything.
• The process changes constantly, so the tool would need frequent rework.
• An existing subscription already does the job well enough.
• A wrong answer is costly and checking takes as long as doing the task by hand.

Sometimes there's a simpler answer: a better spreadsheet, a template, or an AI subscription you use yourself. That beats building something that doesn't pay.

Benefits that aren't hours

Time saved is the easiest benefit to count, but it is a time value and not the only benefit. Answering enquiries after hours might win jobs you'd otherwise lose, which would be additional revenue with its own costs and margin. Fewer errors may mean fewer credit notes, and faster quotes mean customers hear back sooner. These are harder to put numbers on and we have not assumed any here, so leave them out of the sum at first. If the project pays on time savings alone, anything extra is a bonus. If it only pays once you add them, be careful.

Using the calculator

The ROI calculator on this site does this arithmetic with your own figures. Try a few variations, including pessimistic ones. If a project still looks good, it's worth a conversation about AI consulting or a first AI automation job. The invoice automation demo shows invoice details prepared for checking. We usually begin with one job.

Next step: try your own figures in the calculator.