Picture a director I’ll call Anita. Sitting through a vendor pitch where the deck reads “industry benchmarks show 3 to 8x ROI.” The vendor treats the slide as a baseline expectation. Anita knows her peers in the same sector have not seen those numbers; the conversations she has had at industry events suggest something closer to 1.5x. She does not yet have the language to push back without sounding sceptical of the whole engagement. She also knows that signing on the basis of a number she does not believe is a slow way to lose trust with her board.
The headline number is not made up, and it is also not what most SMEs see. Decomposed honestly, the gap between the slide and her board’s experience has a structural explanation.
Where does the 3.5x to 8x figure actually come from?
The headline range originates from IDC, Microsoft, and Forrester research aggregating case studies and customer database analyses. The methodology behind these aggregations is rarely transparent, but the structure can be inferred. The figures aggregate across multiple use cases, industries, and customer sizes, with the higher end of the range reflecting cases where AI is applied to high-leverage processes and the lower end reflecting more modest applications.
Several caveats apply. The figures are often based on customer self-report or case-study data rather than independent audit. Different studies use different definitions of ROI; some include improved decision quality and reduced risk, which are hard to monetise. Some focus narrowly on labour cost savings. Others include revenue uplift from new capabilities. The timeframe matters substantially: first-year ROI is often cited but typically excludes the transition and learning costs that fall in months one through three. The customer set is often biased toward early adopters and technology-sympathetic firms.
What this means in practice: the 3.5x to 8x range is an aspirational upper bound. It represents what is achievable in best-case scenarios with excellent execution. Treating it as a representative estimate is what produces disappointment at year-end.
What does the decomposition look like?
The arithmetic is straightforward. Assume an AI deployment costs £30,000 in Year 1 (software, implementation, training). Assume it saves £25,000 in Year 1 through labour hours saved, reduced errors, and improved process efficiency. That is 0.8x first-year ROI, with break-even around month fourteen.
Year 2, with no additional software cost and reduced implementation cost, assume the same £25,000 in annual benefit. That is 1x ROI in Year 2 alone. Cumulated over three years, assuming consistent benefit, the cumulative benefit is £75,000 against a £30,000 investment, or 2.5x cumulative ROI.
Add a 1.5x multiplier for secondary benefits (improved client satisfaction, reduced compliance risk, better decision quality), and the cumulative figure climbs to roughly £112,500, or 3.75x ROI. That gets within the 3.5x to 8x range, conditional on the secondary-benefit assumptions holding up.
The arithmetic shows the headline is achievable. It also shows what has to be true for it to land. Best-case software cost. Best-case adoption that produces the full £25,000 saving. Generous secondary-benefit assumptions. Three-year time horizon. Move any of these toward typical, and the cumulative figure drops sharply.
What does the failure data say?
The MIT NANDA failure analysis provides the counter-evidence. Across multiple studies of technology projects, approximately 60 to 70 percent fail to deliver expected returns, where “failure” means delivering less than 70 percent of projection. The distribution of project outcomes is heavily skewed to the left. A minority of projects deliver exceptional returns, 2x to 3x or higher. A plurality deliver moderate returns, 1.2x to 1.5x. A substantial minority deliver poor returns, 0.5x to 1x.
The Standish Group CHAOS Report 2024-2025, which tracks technology project outcomes across thousands of projects annually, reports that approximately 35 percent of technology projects deliver expected benefits, 50 percent deliver substantially less, and 15 percent are essentially failures. The median overrun on projected benefits is approximately 30 percent. A proposal projecting 2x ROI typically delivers 1.4x in the originally-stated timeframe.
Both data sets converge on the same point. The 3.5x to 8x cases exist; they sit in the upper tail of the distribution, drawn from firms that executed well, picked the right use cases, and made the surrounding investments the climb requires. The median case is materially lower.
Why don’t most SMEs hit the headline?
Six structural reasons. The first is adoption: actual rates are typically 50 to 70 percent of eligible users, not the 80 to 90 percent assumed in the headline cases. Second is reallocation: freed-up time is rarely deployed to revenue-generating activity unless the firm has pre-planned it. Third is change resistance: workflow misalignment reduces productivity gain below theoretical estimates.
Fourth is informal measurement: firms do not detect underperformance until it is too late to correct. Fifth is use-case mismatch: the AI is applied to work where it produces lower-quality output requiring rework. Sixth is wrong tool: a generic AI assistant where a domain-specific tool would have produced better results.
Any one of these is enough to drop ROI from headline to median. Most SMEs encounter at least two of them. The cumulative effect is the gap between vendor projections and lived experience.
What is the realistic benchmark?
For a competent SME deployment, with reasonable adoption, deliberate reallocation, manageable change resistance, and honest measurement, year-one ROI is typically 1.2x to 1.8x. Cumulative two-year ROI is typically 1.5x to 2.5x. Some firms exceed this with high-leverage use cases and excellent execution. Some fall short, particularly when implementation is poor or use cases are unsuitable.
This is genuinely valuable return. A 2x cumulative ROI on a £30,000 investment is £30,000 of net benefit; not the headline number but real money, particularly compounded across multiple deployments over multiple years. The honest framing is that AI is worth doing at the realistic benchmark, and selling it on the aspirational benchmark damages the case more than helps it.
The takeaway question for any vendor or consultant claim is concrete. Across all your customers, not just your case studies, what is the median ROI at twelve months and at twenty-four months? What is the lower quartile? A vendor who cannot answer those questions, or whose answer diverges from their case-study figures, has told you what you need to know.
If you are weighing an AI proposal that quotes the headline range and want help working out the realistic figure for your specific use case, book a conversation.
