How much does AI really return? 95% of pilots don't earn back — here's where the return is (and how long it takes)

There's one number no AI vendor will ever put on a slide, and it's the most important of all: according to the MIT study "The GenAI Divide" (2025), roughly 95% of enterprise generative-AI pilots produce no measurable impact on the bottom line. Only one project in twenty reaches a concrete acceleration in revenue. This isn't a statistic meant to discourage you — it's the honest starting point from which you build a real return, instead of chasing an imagined one.

Because the MIT reading is almost counter-intuitive: that 95% is not a problem of model quality. The tools work. MIT calls it the "learning gap": companies fail to graft AI into their workflows, their organisation and their habits. The model is good; it's the implementation around it that's missing. And that's exactly the piece that decides whether you're in the 5% or the 95%.

The honest ceiling: most companies don't earn the spend back

The MIT finding isn't an outlier. The PwC 2026 CEO Survey confirms it from the side of the people who sign off: 56% of chief executives report zero increase in revenue or reduction in costs from AI so far, and only 12% report both. But the detail that matters is a different one: the CEOs who do see a return are two to three times more likely to have deployed AI broadly across operations, rather than trying it in a single department. The return doesn't come from the isolated pilot: it comes from adoption that gets into the processes.

Holding these two numbers together — 95% of pilots at zero impact, 56% of the top brass with no return — serves one purpose: calibrating expectations before you sign. Anyone who promises you a fast, guaranteed return is selling you the exception as if it were the rule. The rule is that AI only pays off where someone closes the implementation gap. The rest of this article is about where the return really concentrates, and how long it honestly takes to arrive.

Where ROI really concentrates — and it isn't where the budget goes

MIT also measured where the return shows up, and the answer overturns the common intuition: more than half of generative-AI budgets go to sales and marketing tools, but the highest measured ROI sits in back-office automation — eliminating outsourcing spend, cutting external agency costs, streamlining admin and operations workflows. The money chases the most spectacular use case; the return sits in the most boring one.

It's a direct argument against the "let's start with a chatbot for sales" reflex. For an SME the practical reading is: look first where the work is repetitive, measurable and low-risk — accounts payable, reconciliation, document triage — not where it's most visible. It's the admin and finance and operations use cases that deliver the sharpest before/after, not the shop window.

On the same axis sits the buy-versus-build choice: buying from a specialised vendor succeeds around 67% of the time, while projects built in-house succeed roughly a third of the time (~33%) — half the success rate. For an SME with no data-science team, buying (or having a ready-made stack grafted in) isn't the lazy choice: it's the empirically safer one. It's the theme of the buy or build article, confirmed here by the numbers.

Payback times are longer than they're sold to you

The second expectation to correct is the when. The market sells returns within the year; the evidence says otherwise, and it agrees across multiple sources:

• Gartner (782 infrastructure and operations leaders, late 2025): only 28% of AI initiatives fully meet their ROI expectations, and 20% fail outright. Among those reporting setbacks, 38% cite gaps in team skills and 38% cite poor data quality or accessibility — both solvable by an implementation partner, not by better software.
• Deloitte (1,854 executives, Europe and the Middle East): only 6% of organisations see a payback in under a year, and even among the best projects just 13% break even within 12 months. The majority expect two to four years for a satisfactory return — against the seven to twelve months typical of an ordinary technology investment.
• BCG ("Closing the AI Impact Gap", 2025): realised ROI settles at around 10%, against a target of ~20% that most companies had set for themselves — and daily frontline use is stuck at 51% even where executives and managers use AI several times a week (over 75%). Again: the gap is depth of adoption, not the quality of the tool.

The practical consequence for an SME is sharp: promising a payback under twelve months isn't credible against this evidence. The honest frame is two to four years, with back-office and ops/finance cases as the fastest realistic wins. A vendor guaranteeing you more is ignoring the very data they ought to know — and the difference between the 28% that earn back and the 20% that fail isn't the software, it's the skills and the data: exactly what a well-run implementation works on.

The layoffs trap: ROI doesn't come from cutting heads

There's one last number that works as a positioning compass, not just a statistic. Again Gartner (350 executives at companies above a billion in revenue): 80% of companies that experimented with AI and automation report headcount reductions — but the reduction rate is almost identical between high-ROI and low-ROI companies. In other words: no correlation between cutting jobs and getting better returns. As Gartner vice-president Helen Poitevin puts it, "chasing value through workforce reduction alone leads most organisations to limited returns".

The best companies used AI to augment their people — more productivity from the same headcount — not to replace them. For an Italian SME it's at once the cleaner story ethically and the one the data actually backs: the return isn't in laying people off, it's in freeing up hours and redeploying them onto higher-value work. And it's also how the team accepts the change instead of resisting it: a tool that augments its user gets adopted; one that threatens the job gets sabotaged, and an unused tool has zero ROI no matter how good it is.

What it means for your company

Tie the numbers together and what comes out is a method, not pessimism. Four rules to stay on the side of the 5% that earns back rather than the 95% that doesn't:

• Start from the back-office, not the shop window. The highest measured ROI is in automating admin, finance and operations — not the sales chatbot, where half the budget goes instead.
• Buy, don't build. For an SME with no ML team, adopting a ready-made stack succeeds twice as often as building everything in-house.
• Budget for two to four years, not twelve months. It's the horizon that holds up in front of whoever keeps the books — and here credibility is worth more than an inflated promise.
• Aim to augment, not to replace. It's what the data correlates with return, and it's what makes adoption take root in the team.

All four rules share a common thread, and it's the gap MIT named: the return isn't decided by the model, it's decided by the implementation. The 38% citing missing skills, the 38% citing data that isn't ready, the pilot that never enters the workflows — it's all learning gap, and it's all the work that sits between "we bought a tool" and "the company earns more". It's why Innesti's line isn't "we sell you AI" but "we graft it in": the part that decides the ROI is the part almost everyone skips.

Before estimating any return, though, you need to know where you're starting from. We've turned that first step into a free, self-serve assessment: a few questions and an indication of where AI can pay off in your company, with what controls around it and how to measure it — the method for measuring ROI is this article's sibling. Take the AI-readiness assessment — then, if it makes sense, let's talk.

This article is for orientation. The figures on return, success rate and payback times cited come from industry studies and surveys — MIT NANDA "The GenAI Divide: State of AI in Business 2025"; PwC 2026 CEO Survey; Gartner (surveys of I&O leaders in late 2025 and of executives at large companies); Deloitte (2025 survey of 1,854 executives in Europe and the Middle East); BCG "Closing the AI Impact Gap" (2025) — and should be read as indications of direction, not as guarantees of results: samples, definitions and horizons vary across the sources, and every estimate must be verified against the context of the individual company. Every tool choice and every AI investment must be assessed against the data, the controls and the goals of your company, with the oversight of whoever answers for it on the books.