In early 2025, two companies we know set out to "do something with AI." They had roughly the same budget, the same enthusiasm, and the same board asking pointed questions about falling behind.
The first one moved fast. They bought a well-known AI platform, ran a flashy internal launch, and told everyone the future had arrived. Twelve months later, the tool sat mostly unused. A handful of people had tried it, been underwhelmed, and drifted back to the spreadsheets they trusted. The project was quietly filed under "a learning experience" — which is what organizations say when something expensive didn't work.
The second company did something far less exciting. They picked one painful, boring task — matching supplier payments against a mess of bank transactions — and pointed AI at just that. No launch event. No press release. By year's end, a job that used to eat two full days a month took a couple of hours, and the finance lead had stopped dreading the close.
Same hype. Same money. Wildly different outcomes. And the difference had almost nothing to do with the technology, and everything to do with where they chose to start.
Start with a problem, not a model
The most common way to waste money on AI is to begin with the tool and go looking for a use. It feels productive — you're "adopting AI" — but you're really solving for the technology instead of the business.
The companies that get value do the opposite. They start with a question that has nothing to do with AI at all: where does our business quietly lose time, money, or accuracy every single day?
That question leads you somewhere unglamorous, and that's exactly the point. The highest-return AI lives in the repetitive corners of your operation — the places where people copy data between systems, re-key invoices, chase the same approvals, or answer the same customer questions on a loop. Nobody puts these tasks in a strategy deck. But add them up across a year and they're enormous.
The three-part test
When you're weighing up a candidate task, hold it against three traits. The best opportunities have all three:
- High volume — it happens hundreds or thousands of times a month. Volume is what turns a small per-task saving into a big annual one.
- Rules-based — a knowledgeable person could write down, more or less, how each decision gets made. If the logic can be described, it can usually be assisted.
- Costly when wrong — mistakes create rework, delays, penalties, or unhappy customers. That's what makes the improvement worth paying for.
A task that ticks all three is almost always worth pursuing. One that ticks none — however futuristic it sounds — usually isn't, no matter how good the demo looks.
Three questions before you build anything
Once you've found a candidate, resist the urge to jump straight into building. Ask three questions first, because the answers decide whether the project succeeds long before any code is written.
What decision or output are we trying to improve? Be ruthlessly specific. "Faster customer responses" is a wish. "Cut first-response time on support emails from six hours to under one" is a target you can actually hit and measure.
Do we have the data to support it? AI is only ever as good as the information it can reach. If the data it needs is scattered across five systems, half of it trapped in PDFs and the rest living in someone's inbox, fix that first. More AI projects die here than anywhere else.
How will we know it worked? Agree the number you expect to move — hours saved, errors reduced, response time cut — before you begin. Teams that skip this step end up with a tool nobody can prove is helping, which is how tools quietly get abandoned.
If it can't be measured, it can't be improved — or defended when someone asks what the money bought.
Prove it small, then let it compound
Here's the part that separated our two companies. The one that succeeded didn't try to transform everything at once. They picked a single, well-scoped problem, shipped a governed solution, measured it honestly against the target they'd set — and only then moved to the next one.
That approach does something a big-bang rollout never can: it builds trust. When people see one real task get genuinely easier — when the finance lead actually gets her evenings back at month-end — they stop being sceptical and start bringing you the next problem themselves. Momentum becomes internal, and internal momentum is the thing money can't buy.
The company that failed had the opposite experience. Because they led with the tool and no clear win, every team formed its own opinion — most of them unfavourable — and the initiative never recovered from the first impression.
The quiet version wins
AI doesn't have to arrive with fanfare. In fact, the versions that last rarely do. They show up as a task that used to hurt and now doesn't — one workflow at a time, each one measured, each one earning the right to the next.
So before you evaluate a single platform, go find your version of that boring, painful, high-volume task. Start there, keep a human in the loop, measure what changes, and let the wins add up.
Not sure where AI would genuinely move the needle in your business? That's exactly the conversation we like to have. Explore our AI services or book a consultation, and we'll help you find the right place to start.
