AI ROI: a board-ready framework for funding AI that pays back
Enterprises spent $30–40B on GenAI pilots and 95% saw no P&L return. Here's how to fund the 5% that pay back.
AMDIM · Strategy
June 10, 2026
Here's a number that should ruin a CFO's afternoon.
Last year, companies spent somewhere north of $30 billion building generative-AI pilots. MIT went looking for the return on all that money and found that 95% of those pilots delivered nothing measurable to the P&L. Not "underwhelmed the board." No measurable return at all.
We keep seeing the same scene play out. The demo dazzles. The pilot ships. And a few months later, someone in finance asks the only question that ever really matters — what did it actually save us? — and the room goes quiet.
Here's the strange part: the technology usually works. The model does roughly what it promised. The thing that breaks is almost never the AI. It's the way the investment was framed, sequenced, and measured — or rather, the way it wasn't.
So this isn't a piece about whether to fund AI. The 5% who do get returns are quietly pulling away from everyone else, and standing still isn't a strategy. It's about how to fund it the way they do: with a baseline, an honest cost, and a number you'll be held to.
The gist, if you're skimming: AI ROI is a discipline, not a technology. Baseline the process, tie it to one lever on your P&L, count the whole cost — not just the licence — and measure payback in months. That's most of what separates the 5% from the 95%.
It's not a technology problem — which is good news
Good news, because approach is something you control. When we look at the companies stuck in the 95%, the same four habits show up again and again.
They never set a baseline. Nobody wrote down what the process cost before the AI — the hours, the error rate, the cost per transaction. So when the model improves things, there's no before-state to measure against, and the win stays invisible even when it's real.
They bank on soft benefits. "Productivity." "Better experience." Both real, both genuinely valuable — and both almost impossible to put in front of a board. Boards fund hard, cash-traceable savings. Soft gains belong in the upside case, not the headline.
They leave the workflow untouched. This is the big one. McKinsey found roughly 80% of organisations bolt AI onto their existing processes instead of redesigning them — and workflow redesign turned out to be the single strongest predictor of whether EBIT actually moved. A faster version of a broken process is still a broken process.
And they underestimate the cost. The licence is the part everyone sees. The data work, integration, evaluation, governance and human oversight underneath it are the part that quietly turns a six-month payback into eighteen.
Notice that none of these is an engineering failure. They're all funding-and-governance choices — which means a leadership team can fix every one of them.
So how do you actually measure AI ROI?
The ROI sum itself is boring: value over cost. The discipline is in how you fill it in. We use four moves, in order — and skipping any one is usually how the number stops being believable.
exhibit
First, baseline the process. Before a line of code, measure what the target process costs today: volume, hours, error rate, cycle time, cost per unit. "We process 40,000 invoices a month, nine minutes each, with a 6% exception rate" — that's a baseline. It's also the most politically awkward step, because it commits someone to a number. Which is exactly why the 95% skip it.
Then map it to a value driver. Take your P&L apart into the levers that actually move it — conversion, churn, days sales outstanding, claims cycle time, cost-to-serve — and tie the AI to one of them. Suddenly "we built a chatbot" becomes "we cut cost-to-serve by taking three minutes off average handle time." One is a project. The other is a business case.
Then price the real cost. Not the API bill — the whole thing (more on that next). Gartner keeps finding that organisations underestimate this, which is precisely how a project that looked profitable quietly isn't.
Finally, express it as a payback period. Most boards trust "recoups in seven months, then £1.4m a year" far more than "320% ROI." A payback measured in months, against the true cost, is harder to game and much easier to govern at each funding gate.
The cost you're not counting
The reason so many "profitable" projects disappoint is that the business case only ever counted the visible costs. The licence sits at the top of an iceberg, and the payback lives or dies on everything below the waterline.
exhibit
None of this is an argument against investing. It's an argument for investing with your eyes open — so the payback you present survives contact with reality, and so nobody quietly kills the project in month nine when the "surprise" costs land. Pricing the full iceberg up front is what protects your credibility the second time you ask the board for money.
Which bets pay back first
Not all AI value arrives on the same schedule, and pretending it does is how portfolios stall. Two questions sort the work: how cash-traceable is the benefit, and how often does the process run?
High-frequency, high-traceability work pays back first. That's exactly where MIT found the winning 5% concentrating — not on the glossy customer-facing showcase, but on dull, repetitive back-office automation where the savings are obvious and the cycle repeats thousands of times a month.
As a rough rule of thumb from how these projects actually behave: high-frequency operational use cases tend to recoup in 3–9 months, mid-complexity ones in 6–12, and big transformations in 12–24+. IDC's data backs the shape of it — value realised in about 13 months on average across adopters. So sequence the portfolio deliberately: let the fast, hard-savings wins land first and fund the slower, higher-ceiling work. Don't bet the programme on the moonshot.
A quick word on what counts as a "win," because this trips people up. Lead with hard savings — cost avoided, work deflected, throughput gained at flat headcount — because they show up in the ledger and boards fund them. Treat productivity and experience gains as upside, not the headline. Model revenue impact only where you can attribute it cleanly. And don't forget risk reduction — fewer errors, lower compliance exposure — which is a legitimate, board-relevant lever in its own right.
Make it a business case, not a pitch
By the time this reaches the board, it should read like any other capital request — not a technology demo. In our experience, five things make it fundable, and the absence of any one is usually why it isn't:
the metric it will move (one KPI, named and owned); the baseline cost of that metric today; the full cost, all the way down the iceberg; the payback in months, with a conservative and an expected case; and the governance — who owns the number, how human oversight works, how risk is controlled.
That last one matters more than it looks. McKinsey found high performers were far likelier to have a defined human-in-the-loop process — 65% of them, versus 23% of everyone else. Boards fund what they can govern.
If you can't answer the value question — which number moves, by how much, by when — don't launch yet. That single rule is most of the difference between a portfolio that compounds and one that stalls.
There's a quieter advantage here, too. Wharton's 2025 research found 72% of leaders now formally measure GenAI ROI — and the ones who measure are the ones realising value. You can't defend at next year's budget what you never instrumented this year.
Where this goes wrong
A framework isn't a guarantee. Three things can still sink a disciplined process, and they're worth naming out loud.
Baseline theatre. A baseline assembled to justify a decision you've already made is worse than no baseline at all. It has to be measured, ideally by someone who doesn't have a stake in the answer.
Attribution inflation. The moment revenue ticks up, everything wants the credit. Be ruthless about what the AI actually caused versus what rode a good quarter. Over-claim once and you'll be disbelieved forever.
The pilot that can't scale. A use case that pays back in a tidy pilot can fall apart when it meets real data volumes, real integration and real governance. The return only counts if it survives the move to production — which is its own discipline, and a common graveyard. (We've written separately on why most AI stalls between pilot and production.)
Naming these isn't pessimism. It's what makes the optimistic case believable.
So, on Monday
If you take one thing from this, let it be this: stop asking for "the ROI of AI" in the abstract, and start demanding it per use case, in the language of your own P&L.
Pick one high-frequency, high-cost process. Baseline it this month. Tie it to a single value driver and a single KPI. Price the full cost, not the licence. Commit to a payback in months, with an owner. And put human oversight and measurement in from day one.
Do that, and you're funding like the 5%, not the 95%. The technology was never the hard part. The discipline around it is — and that part has always been yours to control.
Frequently asked questions
What is a good ROI for an AI project? There's no universal figure — it depends on the use case and how honestly you've drawn the costs. A more useful target is payback period: high-frequency operational use cases commonly recoup in 3–9 months. Across adopters, IDC reports an average of about $3.70 of value per $1 invested in generative AI, and roughly 13 months to value — but those returns concentrate among organisations that measure and redesign workflows.
Why do most AI projects fail to deliver ROI? Rarely the technology. MIT's 2025 research pins the gap on "approach": no documented baseline, benefits that are soft rather than cash-traceable, processes left unchanged, and underestimated total cost. McKinsey found ~80% of firms bolt AI onto existing workflows instead of redesigning them — the strongest single predictor of whether EBIT actually moves.
How do you calculate the total cost of ownership for AI? Add to the licence or API cost: data preparation and integration, evaluation and testing, governance and compliance, ongoing monitoring and retraining, human oversight, and change management. These below-the-waterline costs are routinely underestimated, and they're what turn an apparent short payback into a long one.
Should we measure AI ROI in revenue or cost savings? Lead with hard cost savings — they're cash-traceable and boards fund them. Treat productivity and experience gains as upside, and model revenue impact only with clean attribution. Risk reduction — fewer errors, lower compliance exposure — is a legitimate, board-relevant lever in its own right.
Want to know which of your use cases will pay back first — and what's quietly blocking the rest? That's the work we do as an AI strategy and advisory partner: baseline the process, build the costed business case, and sequence a roadmap your CFO can actually fund. A good place to start is our AI Maturity Scorecard — about ten minutes, and it'll show you where you stand today.
/ go deeper
Put this to work on your numbers.
/ keep reading
Strategy · 10 min read
Pilot to production: why most AI stalls — and the operating model that ships it
Only ~48% of AI projects reach production. The blocker isn't the model — it's the operating model. Here's the playbook that ships.
Read the articleEngineering · 11 min read
Agentic AI in production: architecture, guardrails, and evaluation
Most teams reaching for agents don't need one. A field guide to agent architecture, guardrails and evals — and the production maths that decide if it ships.
Read the article