Someone showed me their AI spend for the year recently. Twelve months of software licences, a consultant brought in for the integration, a handful of subscriptions. The total was not small. What they could not tell me was whether any of it had worked.
That gap, between money spent and outcome measured, is the most common problem in SME AI adoption right now. Founders are making real decisions about real budgets, but the measurement framework often is not there. Without it, every renewal conversation is a guess.
The case studies that exist point to a straightforward measurement model. The numbers are real and worth understanding before you sign another licence or shelve a pilot.
What does measuring AI ROI actually mean?
AI ROI measurement is the practice of tracking what changes in your business after an AI tool goes live, expressed in financial terms you can compare to what you spent. A credible model covers four things: hours saved per week costed at the fully loaded rate, revenue uplift from faster delivery or better conversion, direct cost avoidance, and payback period in months.
The percentage ROI figure gets the headlines, but it often obscures more than it reveals. The Forrester study commissioned by Microsoft, published in October 2024, is a useful illustration. Based on interviews and surveys with over 200 companies with up to 300 employees across sectors including retail and financial services, it projected a three-year ROI range of 132% to 353% for SMBs using Microsoft 365 Copilot. The underlying drivers were a 6% increase in net revenue, a 20% reduction in operating costs, and a 25% acceleration in onboarding new staff.
Those are significant numbers. They are also modelled projections from a commissioned study, not independently audited outcomes. Treat them as a planning benchmark rather than a guarantee and they are genuinely useful. Expect them to apply unchanged to your firm and you will likely be disappointed.
Why does it matter whether you measure this properly?
Without a measurement framework before you start, you are making a spending decision based on expectation rather than evidence. Without one after, you cannot tell whether to continue, scale, or stop. The OECD’s 2025 report on generative AI and the SME workforce found the technology in use in 31% of SMEs. Founders now comparing their AI spend against peers need the same measurement language those peers are using.
The problems that follow from vague measurement are predictable. You cannot spot waste, because you have no baseline to compare against. You cannot scale what is working, because you cannot isolate which part of the workflow delivered the gain. And you cannot have a credible conversation with your bank, your team, or a potential investor about where to go next.
The OECD also notes that generative AI helps compensate for skill shortages, which adds a second ROI driver many founders miss. If a tool is doing work that would otherwise require an additional hire, that capacity gain belongs in the model.
Where does the published evidence point?
The Forrester study commissioned by Microsoft gives the clearest published benchmark for SMB AI ROI. Based on over 200 companies with up to 300 employees across retail and financial services, it projected three-year returns of 132% to 353%. The spread is wide because results depend on which workflows are being automated and how disciplined the baseline measurement is.
OpenKit’s SME framework adds a useful second reference point. It cites an illustrative benchmark of 210% three-year ROI for customer support automation, alongside a 45 to 50% reduction in human-handled support tickets. These figures are examples for hypothesis testing rather than a standard you should expect to replicate. Their value is in giving you a shape to compare against: if your support pilot is showing a 20% reduction in ticket volume at three months, you can see where you are on the curve and what further gains look like.
What both sources confirm is that AI ROI is real and measurable when tied to a specific workflow with a clear metric attached. Vague deployments produce vague returns. That is the consistent pattern across the published evidence.
When is a payback period good enough?
For a small services firm, payback period is more useful than percentage ROI because it connects directly to cash flow. OpenKit’s SME framework treats 6 to 9 months as excellent, 12 to 18 months as acceptable if workflow improvement is already measurable, and beyond 24 months as high risk. At that horizon the technology is likely to have changed before the savings have arrived.
The payback calculation is simpler than the percentage ROI model. Take the total cost of the pilot, including licences, integration time, any training, and compliance work. Compare that to the monthly savings the tool is generating, measured at a point where the team has settled into using it. Divide cost by monthly saving and you have the payback period in months.
The discipline that makes this number meaningful is the baseline. If you do not know what the workflow cost before the tool went in, you cannot calculate what it costs now. Set the baseline before the pilot begins. A rough time-and-cost estimate for the manual process is better than nothing.
What gets left out of the model if you stop at hours saved?
Hours saved is the most visible benefit in an AI ROI calculation, but it is rarely the complete picture. If the tool processes personal data, UK GDPR applies and the ICO’s AI guidance is explicit that AI does not remove data protection obligations. A Data Protection Impact Assessment may be required before launch, and the cost of that compliance, including privacy controls, vendor due diligence, and any human review of automated decisions, belongs in the model from the outset.
Security costs matter too. The NCSC’s AI security guidance covers prompt injection, data leakage, and vendor misuse as real risks that can wipe out expected savings if controls are not in place from the start. Incident response costs can arrive before the payback period does.
For firms in or adjacent to regulated financial services, the FCA’s AI guidance sets expectations for model governance, auditability, and third-party oversight. Even if your firm is not directly regulated, those expectations often flow through from clients who are.
If your firm sells into the EU, or if your AI use falls into a higher-risk category under the EU AI Act, the compliance burden expands further. For many UK SMEs providing professional services that is not the primary concern, but a check at scoping stage takes little time and can save a great deal of remediation later.
The practical point: build the cost side of your ROI model to include governance, security, and compliance from the beginning. An ROI figure calculated without those costs is optimistic, not wrong.
The published evidence says AI ROI is achievable for SMEs. The Forrester benchmark sits between 132% and 353% over three years for SMBs with clear use cases and measurement in place. What separates the firms that see those returns from those that don’t is usually measurement discipline, not the choice of tool.
Start with one workflow. Set a baseline before the pilot begins. Agree a single before-and-after metric and a payback target before the contract is signed. Then build the cost side to include compliance and security. That is the complete model.
If you would like to work through what that looks like for your firm, book a conversation.
