Don't ask what AI can do for you. Ask where you're losing value, then write the problem down: what happens today, what gets in the way, what it costs, and what good would look like in a year. If you can't fill that in, you're not ready to buy anything.
Most organisations don’t struggle to come up with AI ideas. They struggle to pick the right problem, and then they solve the wrong one very well.
This isn’t new, and it isn’t unique to AI. When Thomas Wedell-Wedellsborg surveyed 106 senior executives from 91 companies for Harvard Business Review, 85% agreed that their organisations were bad at diagnosing problems. And 87% said that weakness carried significant costs. AI just makes the cost arrive faster, because it’s now cheap and quick to build a working answer to a question nobody should have asked.
The wrong question
“What can AI do for us?” sounds sensible. It isn’t. It starts with the tool and goes looking for somewhere to put it. You end up with a long list of pilots, each one technically interesting, few of them connected to anything the finance director cares about.
That’s a vendor conversation. So is “what are our competitors doing?”, which is usually anxiety dressed up as strategy.
A better question comes from Nadine Soyez, who works with organisations on AI adoption: where are we losing value or underperforming, and could AI stop it?
That question points you at the business, not the technology. It also gives you a way to judge the result, because you’ve said what’s going wrong before you’ve said what you’ll buy.
Where to look
Four places tend to hide the most value. Work that gets stuck, where things queue because a person has to find information before they can act. Decisions that wait, where someone senior signs off something they barely have time to read. Quality that varies depending on who does the work. And money that leaks, quietly, because nobody owns the gap between two teams.
The last one is easy to miss. In a debt recovery process I know well, everyone’s attention sat at the end: court, collection agencies, write-offs. That’s where the money visibly disappeared. But the loss was really happening in the middle, in the checks people had to do before anyone could safely chase a debt. Who is the customer? Is the property actually occupied? Are we chasing the right company? Each answer meant someone searching several systems by hand. When those checks were slow or incomplete, the wrong party got chased, claims got withdrawn and cash sat uncollected.
Nobody had described that as an AI problem. It was a problem with evidence. That’s exactly why it was a good place to start.
Write the problem down before you touch a platform
A useful problem statement is short. What happens today? What gets in the way? What does it cost, in time, money, complaints or risk? What would you change? And what would good look like in twelve months?
If you can’t fill that in with reasonable confidence, you’re not ready to buy anything. You might be ready to go and measure something, which is often the most valuable step anyway.
Then ask one more question, again borrowed from Soyez: what would convince a sceptic that this worked? If nobody can answer it, go back to the diagnosis. You’ll save yourself a pilot that ends with a nice demo and no decision.
What this looks like in practice
I’ve seen capable tools quietly shelved because nobody framed the problem honestly at the start. I’ve also seen simple automation deliver real value because someone took the time to understand where the work actually got stuck.
The second group rarely had the most impressive technology. They had a clear problem, an owner who cared about it, and a number they could watch move.
Start there. Choosing the tool gets a lot easier once you know what it’s for.
Got a list of AI ideas and no idea which one matters?
Tell me what’s getting stuck in your business and we’ll work out whether it’s a problem worth solving. First coffee’s on me.