AI strategy
AI ROI: how to calculate the return on an AI project
A practical method for calculating AI ROI before and after you launch a project. With a full worked example, a break-even calculation that fits on one page and the mistakes that inflate the numbers.
By Olivier Buigues · Published , updated
What AI ROI measures and how to calculate it
An accounting firm in Nottingham has its Copilot licences coming up for renewal, and the managing partner asks whether they have made a difference. The team says work feels quicker, but nobody has a figure, so the decision comes down to gut feeling. Calculating AI ROI gives you that figure, provided someone measured how the task was done beforehand.
ROI, or return on investment, compares what a project delivers with what it costs over a set period, usually twelve months. The formula is the same as for any investment: benefit minus cost, divided by cost. The payback period is the number of months it takes for the project to earn back the money spent at the start.
The formula is simple, but it needs data, and nearly all of it comes from your own team. That is why the calculation starts well before you buy anything.
It is worth getting right. In the Harvard Business School and BCG study of 758 consultants, those using AI produced work of over 40% higher quality and finished 25% faster. The same study found that when AI was used outside its strengths, results got worse. Measuring tells you which situation your project is in.
Start by measuring the baseline
Without a baseline there is no ROI to calculate. Pick one specific process, such as logging supplier invoices at an accounting firm or answering customer emails at an online shop, and record these four figures for two to four weeks.
Time per task
The minutes one person spends on each unit of work. Time a real sample, because estimates from memory tend to come in low.
Volume
How often the task happens each month, and whether there are peaks such as quarter-end. A small saving per unit adds up quickly at high volume.
Errors and rework
What share of cases has to be corrected or redone, and how long each correction takes. Add indirect costs, such as a late-filing penalty.
Hourly cost
The cost of the person doing the task, including employer contributions. Multiplied by the time, it gives you the money figure you will compare against later.
The costs of an AI project to add up
An AI project costs a good deal more than the monthly subscription. Separate one-off costs, paid at the start, from recurring costs, which come back every month.
Licences and usage
Subscriptions to ChatGPT, Claude, Gemini or Microsoft Copilot, the automation tool and model usage. Count every seat, including the ones people barely use.
Build and integration
Designing, connecting and testing the solution with your email, CRM or accounting software, whether your team or a provider does it. It is a one-off cost in year one.
Training
The hours the team spends learning instead of producing. Without training the expected savings don't arrive, so this is a poor place to cut.
Maintenance
Adjustments when a process changes, tool updates and prompts that stop working. Budget a realistic monthly amount.
Human oversight
The time someone spends checking what the AI produces, handling exceptions and monitoring quality. It is a permanent cost and comes off the gross saving.
The benefits you can measure
Value each benefit on the same basis as the baseline. Anything you can't measure gets a separate mention, kept apart from the figures.
Hours saved, valued in money
The time saved per task, multiplied by volume and by the hourly cost. Only count time that genuinely goes on other useful work, such as serving more clients or closing the month sooner.
Fewer errors
Fewer corrections mean less rework. Add any avoided costs you can document, such as penalties or refunds.
Faster response
Answering in an hour instead of a day improves service. Measure it as response time and, where you have the data, as conversion. Otherwise, record it as a qualitative benefit.
Revenue effects, handled with care
More quotes sent can turn into sales, but season and pricing play a part too. If you include them, use a conservative estimate and show it apart from the time savings.
Example
The sums for an accounting firm automating invoice logging
Back to the Nottingham firm, which wants to automate the logging of supplier invoices. The figures are hypothetical round numbers chosen to show the method. They are not Obinova's rates and not a result you should expect.
For three weeks, the team times the task. It handles 1,000 invoices a month at 6 minutes each, which is 100 hours, and corrections take another 20 hours. At a hypothetical cost of £25 an hour, the process costs £3,000 a month.
After a pilot, each invoice needs 3 minutes of checking and corrections drop to 10 hours. That saves 60 hours a month, worth £1,500 a month and £18,000 a year. Year one costs £6,000 in one-off costs (£5,000 to build and £1,000 for training) and £500 a month in recurring costs (£250 for licences, £125 for maintenance and £125 for oversight). In total, £12,000.
First-year ROI is (18,000 − 12,000) / 12,000 = 50%. In year two, with no one-off costs, it becomes (18,000 − 6,000) / 6,000 = 200%, provided volume and costs stay the same. The net benefit is £1,000 a month, so the £6,000 spent at the start is paid back in 6 months. The freed-up hours go on advising more clients, and that is what the partner checks at each quarter-end.
Break-even: how many minutes AI has to save
Estimating benefits invites optimism. This calculation turns the question round. It starts from the minimum saving the project needs to avoid losing money, then asks the team whether that is realistic. With the firm's figures, a calculator is enough.
- 01
Add up the total monthly cost
The recurring costs (£500) plus the one-off costs spread over twelve months (another £500). The project has to cover £1,000 a month in its first year.
- 02
Turn it into hours
Divide by the hourly cost. At £25, that is 40 hours a month AI must free up just to break even.
- 03
Bring it down to one unit of work
Spread those hours across the volume. 40 hours is 2,400 minutes, which across 1,000 invoices comes to 2.4 minutes per invoice out of today's 6.
- 04
Ask the person who does the task
Show the figure to whoever logs the invoices. If saving that much sounds easy, the project has headroom. If it sounds impossible, rethink the scope or the costs before you invest. A pilot on real invoices will give you the answer with data.
Mistakes that inflate AI ROI, and a first process to measure
Some figures inflate ROI without reflecting any real saving. The most common is freed-up hours that nobody reassigns, a benefit that exists only on paper. Licences activated, prompts sent to an assistant or workflows built don't count either. They measure activity and say little about results.
Avoid counting the same benefit twice, such as the hours saved and the salary of whoever did the task. Don't use the pilot's best week as your reference or credit AI with a rise in sales that coincides with a campaign.
To get started, pick one high-volume process, time a sample for two weeks and work out your break-even point. If the minimum saving looks achievable, launch a pilot with an end date and a written success criterion.
In an Obinova AI audit, each use case is scored on impact and feasibility, and the roadmap sets out how it will be measured. During the follow-ups through the year we review the agreed indicators with you. If you then want to automate the process, we build the n8n workflow with your team.
Frequently asked questions
What is a good ROI for an AI project?
There is no single figure that suits every company. It depends on the project's risk, what the money would earn elsewhere and how long leadership is willing to wait. The useful step is to set the threshold before you start, for example paying back the investment within a year, and then check it against real data from the pilot.
Is there an AI ROI calculator I can use?
A spreadsheet is enough. Enter the time per task, the monthly volume, the hourly cost and the project's total monthly cost, and it gives you the break-even saving per task, as in the steps above. An online AI ROI calculator works the same way and is only as reliable as the figures you feed it. Time a real sample first and use your own costs, not the tool's defaults.
How do you measure the ROI of ChatGPT or Microsoft Copilot?
Pick three or four routine tasks for each role, such as drafting emails, summarising meetings or preparing reports. Time them before and after on a real sample over a few weeks. Compare the time saved, valued at the hourly cost, with the price of the licences and the training. Surveys help you interpret the data, but they are no substitute for measuring.
How long does it take to see a return on AI?
It depends on the project. An assistant used well can save time within the first few weeks, while an automation has to be built and tested before it pays its way. The sensible approach is to measure during a pilot of a few weeks and use those figures to work out the payback period. If it is far longer than leadership will accept, rethink the scope.
What if the hours saved don't reduce costs?
That is the most common case. Freed-up time has value when it goes on other useful work, such as serving more clients, cutting overtime or taking on more volume without hiring. Decide before the project what that time will be used for, and measure it. If it isn't reassigned, count only the benefits you can show, such as fewer errors or faster replies.
Can Obinova help calculate the ROI of AI?
Yes, it is part of the AI audit. We look with your team at how the work is done today, score each use case on impact and feasibility, and set out in the roadmap how the results will be measured. During the follow-ups through the year we review those indicators. You receive a quote for the audit after a free 30-minute first meeting.