How to Maximise the ROI on Your AI Project (and Land in the 5% That Works)

Most AI projects don’t pay off. Here’s what the few that do have in common.

Here’s a number worth sitting with. In its 2025 “State of AI in Business” study, MIT found that around 95% of enterprise generative AI pilots delivered no measurable impact on the bottom line. Only about 5% took off. That isn’t a story about AI being overhyped. It’s a story about how the projects were run, because the 5% weren’t using better models than the 95%. They made better decisions before anyone built anything.

That’s the good news. ROI on an AI project is mostly decided up front, in a handful of choices that have nothing to do with which model you pick. Get them right and you give yourself a real shot at being in the 5%. Here are the four that matter most.

1. Pick a use case with a number attached

The single biggest predictor of return is choosing a problem whose value you can actually measure. It sounds obvious, and almost nobody does it.

Notice where the returns actually hide. MIT found the biggest ROI wasn’t in the flashy customer-facing tools that soak up most AI budgets, but in unglamorous back-office and process automation: the repetitive internal work that quietly eats hours. So start there. Find a process that costs you measurable time or money every week, and name the metric before you begin: hours given back, cost removed, a quote turned around in one day instead of three. If you can’t put a number on the problem, you won’t be able to prove a return on the solution.

The old management adage, what gets measured gets managed, might as well have been written for AI projects. Pick the task you can count.

2. Get your data ready, because that’s where the return leaks out

When an AI project fails, the model is almost never the culprit. The cause is usually the data and the workflow around it. Gartner expects 60% of AI projects that lack what it calls “AI-ready” data to be abandoned through 2026.

This is the part the demos hide. Most of the real work, and most of the real cost, in any AI project is getting your information into a state where the result actually means something: connected, cleaned, and reliable enough to trust. Budget for it deliberately rather than discovering it halfway through. A project that skips this step doesn’t save money, it just moves the bill to the end and adds disappointment on top.

3. Don’t build from scratch what you could buy or partner for

This is the MIT finding that should change how most businesses approach AI. Solutions bought from specialist vendors or built with a partner succeeded around 67% of the time. Internal, build-it-all-yourself projects succeeded only about a third as often.

Building in-house feels cheaper and more like “ours.” In practice it’s where a lot of ROI goes to die: longer timelines, the ongoing burden of running it, and a team reinventing plumbing that already exists. That doesn’t mean buying is always right, and it doesn’t mean never building. It means the default of “we’ll build it ourselves” is statistically the riskiest route to a return. The same logic runs through our pieces on no-code tools versus a consultancy and running your own AI infrastructure: control is worth paying for, but only when you actually need it.

4. Measure from day one, and let people actually use it

Define what success looks like before the build starts, then track it from the first week, not eight months later when someone asks where the value went. MIT’s diagnosis of the failures was blunt: the systems that stalled were the ones that didn’t get woven into how people work.

Two habits protect the return here. Start narrow, prove the number on one workflow, and only then widen the remit. And keep a person at the decision point, so the output is trustworthy enough that your team actually adopts it. Adoption is not a soft metric. An accurate tool that nobody uses returns exactly nothing.

What an ROI-first project actually looks like

Put those four together and you get a very different shape of project from the open-ended “AI transformation” that tends to stall. You get something narrow, measured and time-boxed.

That’s deliberately how we run an AI Launchpad: a fixed fee agreed before we start, six weeks to working software, a demo every week so you see progress rather than promises, and production cost projections attached at the end so you know what the running bill looks like. It also includes the answer most providers won’t give you: an honest “don’t proceed” if the numbers don’t work. For AWS customers, there’s often a neat trick to funding it, because a cost and security review frequently finds enough savings to pay for the pilot outright.

None of this requires you to be technical, or to bet the business on a single model. It requires picking the right problem, getting the data honest, being clever about build versus buy, and refusing to start without a number you’re aiming at. That’s the whole difference between the 95% and the 5%.

Pro Tip: If you only do one thing differently, do this: write down the metric you expect to move, and the figure you’d call a success, before you approve any AI work. Our free AI Adoption Playbook walks through choosing that first measurable use case and getting your data ready for it.

Want to know what return your use case could actually deliver?

A free 30-minute call with an engineer, not a salesperson. Bring the process that costs you the most time, and we’ll talk through whether AI can pay for itself on it, honestly.

Book a free call or see how an AI Launchpad works.