There’s a scene that repeats itself in countless small and mid-sized companies. Someone in management comes back from a conference convinced the business needs to “do something with AI”. A project gets signed, a pilot goes live, it gets demoed in a couple of meetings and everyone agrees it looks impressive. Six months later nobody uses it, nobody knows exactly what it cost, and nobody can say whether it saved a single pound.<\/p>\n\n
That isn’t bad luck or bad technology. It’s a project that started without a baseline. Working out the ROI of an AI project for an SMB<\/strong> before you sign anything is what separates an automation that sticks around for years from a pilot that quietly dies. And you don’t need a finance degree: you need four numbers and a spreadsheet.<\/p>\n\n\n Quick answer:<\/strong> Calculate AI project ROI by measuring what the process costs today (frequency \u00d7 time \u00d7 loaded hourly cost + cost of errors), subtracting the full 12-month project cost and dividing by that cost. If payback takes longer than 12-18 months, it isn’t worth it.<\/p>\n<\/div>\n\n\n The industry numbers are uncomfortable, but worth facing. Gartner forecasts that more than 40% of agentic AI projects will be cancelled before the end of 2027<\/strong>, driven mainly by escalating costs, unclear business value and inadequate risk controls (Gartner, 2025). And a study by MIT’s NANDA initiative covering 300 real deployments found that 95% of generative AI pilots produced no measurable impact on the P&L<\/strong> (MIT NANDA, 2025).<\/p>\n\n The revealing part of that second study is the reason why: the failure almost never sits in the model, it sits in the integration and in the choice of use case. In other words, in the decisions made before<\/em> anyone writes a line of code.<\/p>\n\n The failure pattern is remarkably consistent:<\/p>\n\n The formula is the same one you already know: ROI = (annual saving \u2212 annual cost) \/ annual cost \u00d7 100<\/strong>. The hard part isn’t the formula, it’s filling it in with honest numbers. Four steps:<\/p>\n\n An accountancy practice processes 500 purchase invoices a month. Each one takes roughly 4 minutes to receive, read and key into the accounting software: 33 hours a month. At a loaded hourly cost of \u00a322, that process costs \u00a38,800 a year<\/strong> in labour alone.<\/p>\n\n With automated capture and extraction, 85% of invoices go through untouched and the remaining 15% need a 2-minute review: 2.5 hours a month, about \u00a3660 a year. Adding \u00a31,680 a year in licences, consumption and maintenance, the process now costs \u00a32,340 a year<\/strong>.<\/p>\n\n Net saving: \u00a36,460 a year. If implementation cost \u00a33,500, payback lands just under 7 months<\/strong> and first-year ROI sits around 57%, climbing above 380% from year two onwards once the setup cost is behind you.<\/p>\n\n Notice one detail: this case works because it’s 500 invoices a month. At 40 invoices a month, the exact same project never pays for itself. Volume isn’t a minor input \u2014 it’s the input.<\/p>\n\n It’s worth stopping and rethinking when any of these show up:<\/p>\n\n The calculation isn’t paperwork to justify a purchase: it’s a prioritisation tool. Once you put five candidate processes in a table with their annual saving, their cost and their payback, the order to tackle them appears on its own. The winner is almost always a boring, repetitive process rather than the one that looked most innovative.<\/p>\n\n From there the advice stays the same: start with one case, measure it for three months and only scale what the numbers back<\/strong>. That is the exact opposite of buying a large platform and hunting for uses afterwards. This incremental approach is what keeps a hyperautomation<\/a> strategy from stalling halfway, and it fits with keeping human oversight<\/a> at the sensitive points \u2014 which also carries a cost that belongs in the sums.<\/p>\n\n A tightly scoped use case \u2014 document data extraction, an email workflow, an internal assistant \u2014 typically runs between \u00a33,000 and \u00a312,000 to implement, plus a monthly recurring cost for licences and consumption that usually sits between \u00a350 and \u00a3300 for an SMB. Projects far above that range are normally covering several processes at once.<\/p>\n\n For well-chosen cases, between 6 and 13 months. The IDC study cited above puts the average generative AI deployment at around 13 months. If your calculation shows more than 18 months, it’s worth checking whether you picked the right process.<\/p>\n\n The saving is rarely a redundancy: it’s freed capacity. You measure it in hours returned to higher-value work and, above all, in growth absorbed without hiring. “We went from 80 to 110 clients with the same headcount” is a perfectly valid \u2014 and far more honest \u2014 ROI metric.<\/p>\n\n Yes, provided the pilot has success criteria written in advance and a fixed end date. A pilot with no metric and no deadline isn’t a test, it’s an open-ended demo \u2014 and that’s precisely the format that inflates the abandoned-project statistics.<\/p>\n\n The MIT study found that projects backed by specialist vendors or partners succeeded considerably more often than purely internal builds. For an SMB without its own data team, the realistic route is to lean on someone who already has the workflow running and integrate it on top of the software you already use, with no migration.<\/p>\n\n Before investing in AI, spend two weeks measuring what doing it by hand already costs you. It’s the best-returning investment in the whole project: it tells you whether it’s worth doing and, if it is, it hands you the argument to defend it.<\/p>\n\n\n\nWhy so many AI projects get abandoned<\/h2>\n\n
\n
How to calculate AI project ROI for an SMB, step by step<\/h2>\n\n
\n
A worked example with real numbers<\/h2>\n\n
Signs a project will NOT pay off<\/h2>\n\n
\n
What to do with the number once you have it<\/h2>\n\n
Frequently asked questions<\/h2>\n\n
How much does an AI project cost for a small business?<\/h3>\n
How long does it take to recover an AI investment?<\/h3>\n
How do I measure savings if nobody leaves the company?<\/h3>\n
Is it worth running a pilot first?<\/h3>\n
Should we build it in-house or buy it?<\/h3>\n