Three weeks into an AI pilot, the demo still looks good. Behind it, someone is choosing the clean records by hand, reformatting dates and quietly dropping the rows that break things. Nobody mentions it. Then the pilot ends, that person moves on, and the firm tries to run the same process for real. It falls over straight away.
The model was fine. The data was never in a state a machine could use, and in professional and financial services firms that’s the most common reason AI projects stall.
The context problem
An AI model is only as useful as the information it’s given, however capable it is. This is sometimes called the context problem. Give a model clean, structured, consistently labelled data and it does well. Give it what most firms have and it produces unreliable output, misses important details or fills the gaps in the record with confident inventions.
The pilot almost never shows this, because someone always prepares the data for the demo. That preparation is invisible, so the dependency on it is invisible too. When it stops, the AI gets the blame for a data problem.
In an advice firm the set-up often looks like this: contact details in the CRM, correspondence in email, compliance notes on a shared drive, fee schedules in a spreadsheet, and meeting records in an old system nobody finished migrating away from. An experienced adviser finds their way through it by memory and habit. A model can’t. It needs a complete, consistent record in one place it can reach, or a pipeline that assembles that record reliably before the model sees anything.
What fragmented data costs before AI arrives
Fragmented data costs time long before anyone tries AI. Say a client review needs forty-five minutes of preparation: pulling records from four systems, reconciling the differences, checking the compliance notes match the CRM. Multiply that by every review, every week. Nobody adds it up, because it has always been that way.
Put AI on top of that and it can’t help until the records are consistent and reachable. Fix how the data moves first, and the AI part is often the smaller job. That data fix is usually integration work: connecting the systems you have so information reaches the right place without anyone assembling it by hand. It’s configuration rather than engineering. The AI layer, more often than not, is already inside tools the firm pays for. (Our guide to what “AI-ready” means for a small business covers the data side in more detail.)
Three questions that find the real blockage
Can you say exactly where a given client record lives? Not “in the CRM somewhere”. If the answer comes with a condition (“depends when they joined”, “the older ones are in another folder”), you have a fragmentation problem.
What does your team do in the twenty minutes before a client meeting or compliance review? The answer is usually the clearest description of your data gap you’ll get. Those manual preparation steps are exactly what a properly connected system would do for you.
Would your most important client process still produce the right result without your most experienced person? Senior people make up for gaps in the architecture with what they know. When they leave, or when you try to grow, that cover disappears and the gaps show.
If any of these produces an uncomfortable answer, start there rather than with the model.
Fix the flow before you evaluate the tool
AI tools are worth your time. What’s possible now is a long way past what was possible two years ago. But I don’t think the firms getting value from AI are the ones that trialled the most tools. They’re the ones that sorted out how their data moves before trying anything clever with it.
That rarely needs a big programme. Usually the most valuable step is finding the one or two manual preparation tasks that happen most often and building a simple integration to do them. That’s a project measured in weeks.
The model you need almost certainly exists already. A firm that buys a better model before fixing its data will get the same stalled pilot, at a higher price.