PwC's January 2026 report summarised responses from 4,454 chief executives surveyed from 30 September to 10 November 2025. 56% said they'd seen no significant financial benefit to date. Only 12% said it had delivered both lower costs and higher revenue.
That finding does not tell us whether measurement, weak adoption or disappointing performance caused the gap. It does tell a buyer to demand a business case they can test. Start with a baseline and make the costs visible before claiming a return.
The short answer: AI ROI is (total benefit minus total cost) divided by total cost. The formula is simple. Measuring it properly means three things: a baseline taken before the AI goes live, every cost included (not just the licence), and benefits counted only when they turn into cash or capacity you actually use. Track payback period alongside the percentage.
Key points:
- Most AI ROI claims count the licence and ignore the people time. That's where the inflated numbers come from.
- Hours saved are not money saved. They only count if the time goes somewhere useful.
- Deloitte found most organisations reach satisfactory ROI on a typical AI use case in two to four years, not months.
- Set your kill criteria before launch, while you're still objective.
What is AI ROI and how do you calculate it?
AI return on investment is the same thing as any other return on investment. You compare what you got with what you spent:
ROI (%) = (total benefit − total cost) ÷ total cost × 100
Two companion numbers make it far more useful:
- Payback period: upfront cost divided by monthly net benefit. It tells you how many months until you're back to zero. Owners understand this instantly.
- Cost per outcome: what it costs to produce one unit of work (a resolved ticket, a processed invoice, a qualified lead) before and after. This is the number that survives a board meeting.
If your project runs more than a couple of years, a finance team will want net present value too. For most small and mid-sized businesses, ROI plus payback is enough to make a decent decision.
The maths isn't the hard part. The hard part is deciding honestly what goes on each side of the line.
What costs belong in an AI ROI calculation?
All of them. A calculator that compares benefits only with the subscription fee leaves out much of the commitment. Include setup, team time and ongoing review as well.
| Cost line | What it covers | Often left out? |
|---|---|---|
| Licences and usage | Seats, API or token charges, hosting | Rarely |
| Build or setup | Consultant or agency fees, integration, configuration | Sometimes |
| Data preparation | Cleaning, structuring and connecting the data the AI needs | Often |
| Internal time | Your people in workshops, testing, training and process redesign | Almost always |
| Oversight | Someone checking outputs, handling exceptions, fixing mistakes | Almost always |
| Maintenance | Prompt and workflow updates, model changes, vendor price rises | Often |
| Governance | Policy, data protection review, security checks | Often |
Internal time is the big one. A tool that costs £400 a month can quietly absorb a manager's Tuesday afternoons for a year. If you want a sense of the external costs before you start, my guide to what AI implementation costs breaks down the market ranges.
How do you measure the return on AI investment?
Start with a baseline. Before anything goes live, measure the thing you expect to change for at least a few weeks: how long the task takes, how many get done, how many go wrong, what each one costs. Without a baseline you're comparing the new world with your memory of the old one, and memory is a generous accountant.
Then pick benefit metrics that match the use case:
- Process automation (invoices, data entry, document handling): cycle time, cost per item, error rate, volume handled without extra headcount.
- Customer-facing AI (chat, phone, email triage): response time, resolution rate, missed enquiries recovered, conversion. If you're weighing up an AI receptionist, missed calls recovered is the number to watch.
- Knowledge work (drafting, research, analysis): time to first draft, revision rounds, throughput per person. The hardest category to attribute, so be conservative.
- AI agents doing multi-step work: tasks completed end to end without human rescue, and the cost of the rescues. More on this in AI agents for small business.
Freed capacity and cash savings are different benefits. Redeployed time may let the same team handle more work, but it does not automatically reduce payroll. Report the capacity value separately from cash released. If you cannot show how freed time is used, leave it out of the quantified business case rather than inventing a saving.
So apply a realisation factor. Count only the share of freed time you can show went somewhere useful. If you can't show it, don't count it.
Soft benefits, like happier staff or better decisions, are real. Report them separately. Don't blend them into the ROI figure, because the first time a finance director pulls on that thread the whole number unravels.
What does an AI ROI calculation look like?
This is an illustrative example with made-up round numbers. It is not a client result. It's here to show the method, and how much the answer moves depending on what you count.
Imagine a 30-person business where five customer service staff spend much of the week answering emails. They bring in an AI tool that drafts replies for a human to check and send. Loaded staff cost is £30 an hour, and the team works 45 weeks a year.
The vendor version. The vendor says each person saves five hours a week. That's 5 people × 5 hours × 45 weeks × £30 = £33,750 a year. The software costs £4,800 a year. ROI: (£33,750 − £4,800) ÷ £4,800 = 603%. Lovely slide.
The honest version. Your baseline and three months of measurement show four hours saved per person, not five. Of those freed hours, you can show about 70% went into handling more volume without the extra hire you'd budgeted for. Fewer mistakes also mean fewer goodwill credits, worth about £1,200 a year.
| Year 1 (illustrative) | Amount |
|---|---|
| Capacity value: 5 × 4 hours × 45 weeks × £30 = £27,000, at 70% realised | £18,900 |
| Fewer goodwill credits | £1,200 |
| Total benefit | £20,100 |
| Setup and integration (one-off) | £10,000 |
| Software and usage (£400 a month) | £4,800 |
| Internal time: training and process redesign (one-off) | £3,000 |
| Oversight: 1.5 hours a week checking and fixing, at £30 | £2,025 |
| Total cost | £19,825 |
Year 1 ROI: (£20,100 − £19,825) ÷ £19,825 = about 1%. Break-even, roughly.
Year 2 looks different, because the one-off costs have gone. Software, oversight and £1,500 of maintenance come to £8,325. The example assumes the same £20,100 benefit. Maintenance is included in the setup fee in year one and budgeted separately in year two. Year 2 ROI: (£20,100 − £8,325) ÷ £8,325 = about 141%. Over two years: £40,200 of benefit against £28,150 of cost, about 43%.
Payback: monthly net benefit is about £1,106 (£1,675 of benefit minus £569 of running costs), against £13,000 of upfront cost. That's roughly 12 months, assuming benefits arrive steadily from month one, which in real life they won't.
Same project. 603% on the slide, 1% in year one, 43% over two years. The last two are the numbers you can defend.
For more on where these business cases go wrong, see why AI projects fail.
How long does AI take to pay back?
Longer than the sales deck says. Deloitte surveyed 1,854 executives across Europe and the Middle East, including the UK and UAE, in August and September 2025. Most said a typical AI use case reached satisfactory ROI in two to four years. That compares with the seven to 12 months they'd expect from a normal technology investment. Only 6% reported payback in under a year.
That's an enterprise sample, and enterprise projects are heavier. A small business switching on a well-scoped tool for one process can see payback much sooner, as the example above shows. But the direction is useful: plan for a slow first year.
The wider data is sobering too. IBM's 2025 survey of 2,000 CEOs found only 25% of AI initiatives had delivered the expected ROI over the previous few years. McKinsey's 2026 State of AI survey found 37% of respondents attribute any EBIT impact at all to AI, and only about 6% qualify as high performers, meeting both its significant-value criterion and at least 5% of EBIT attributed to AI. What separates them, according to McKinsey, is that they redesign workflows around AI rather than bolting it on to old ones.
A short, honest proof of concept is the cheapest way to find out which camp your idea is in.
My take on AI ROI
The common mistake is measuring after the fact. Teams go live, get busy, then try to reconstruct a baseline six months later. You can't. Take the baseline first, even if it's a crude spreadsheet.
Activity metrics are a trap. Messages answered, prompts run, documents generated. They go up and to the right and tell you nothing about money. Pick an outcome you'd be happy to defend to your bank manager.
Run costs move. In my own lab, Ascendancy IQ, where AI agents do real work under human direction, the usage bills and the human checking time are as much a part of the picture as the setup. A system that's cheap to build can still be expensive to supervise. Measure both.
Decide what failure looks like on day one. "If cost per ticket hasn't dropped 20% by month four, we stop." Write it down before launch. After launch, everyone involved has a reason to find the project a success.
How to measure AI ROI in six steps
- Name one outcome. Cost per invoice, response time, enquiries converted. One primary metric, owned by one person.
- Baseline it. Measure the current state for four to eight weeks before go-live, using the same method you'll use afterwards.
- List every cost. Use the table above. Include internal time and oversight, priced at loaded hourly rates.
- Set targets and kill criteria. What result by when, and what result makes you stop. Agree it with whoever signs the cheque.
- Measure realised benefit monthly. Apply a realisation factor to time saved. Keep soft benefits in a separate column.
- Report ROI and payback together, over the right horizon. Year one and a two or three-year view. Then decide: scale, fix or stop.
If you're doing this across several use cases, fold it into your AI roadmap so each project earns its place in the queue.
Frequently asked questions
What is a good ROI for AI?+
There's no universal benchmark worth trusting, because every survey defines return differently. A sensible test: does the project pay back within the period you'd accept for any other investment, using full costs and measured benefits? If yes, it's a good return.
Is there an AI ROI calculator I can use?+
Yes. Use the free calculator at the top of this page. Enter your own costs and expected benefits, refine the assumptions, and download the result. It separates capacity value from cash impact and includes a lower-benefit scenario.
How do you measure the ROI of generative AI tools like ChatGPT or Copilot?+
Measure outcomes for a defined task, such as time to first draft or proposals produced per week, against a baseline. Count the time saved only where it turns into more output or avoided cost, and include the time spent checking AI outputs.
How long does it take to see a return on AI investment?+
Deloitte's 2025 survey found most organisations reached satisfactory ROI on a typical use case in two to four years. Smaller, tightly scoped projects can pay back within a year, but plan for a slow start.
Should soft benefits count in AI ROI?+
Track them, but report them separately from the financial figure. Staff satisfaction and better decisions matter, but mixing them into the ROI percentage makes the whole case easier to dismiss.
Sources
- PwC 2026 Global CEO Survey, PwC press release, January 2026
- AI ROI: The paradox of rising investment and elusive returns, Deloitte, October 2025
- IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, IBM Institute for Business Value, May 2025
- The state of AI in 2026: On the road to ROI, McKinsey, August 2026
If you're building a business case for AI and want someone to stress-test the numbers before the board does, let's talk.
