When a business tells me their AI implementation isn't working, the first thing I do is look at the data. Not the tool. Not the prompts. Not the workflow design. The data.
In the majority of cases, that's where the problem is.
This isn't a popular answer. AI tools are new and still unfamiliar, so when results are disappointing, it feels natural to blame the technology. But the technology is usually doing exactly what it was asked to do. The problem is that it was asked to act on data that can't be trusted.
What I saw in financial services
Earlier in my career I worked in financial services. The business had grown — through acquisitions, through the natural accumulation of different systems over time — into a situation where client data was spread across multiple CRM platforms. Different parts of the business used different systems. Nobody had a complete, single picture of a client.
That might sound like a technology problem, but it was fundamentally a data problem. Each system held a partial truth. One had the right contact details. Another had the correct account ownership. A third had the relationship history. None of them, on their own, was sufficient to act on.
When the business wanted to start using that client data more systematically — for client outreach, for marketing, eventually for automating parts of the relationship management process — it ran into a wall. Which phone number was current? Which email was the right one? Who was actually responsible for the client relationship? There was no authoritative answer.
Before any of that automation could happen, we had to do something much more fundamental: establish which data was right. We profiled thousands of phone numbers and email addresses against external reference data — Experian, in this case — checking for accuracy, currentness, and consistency. It was painstaking, expensive work. And it had to happen before a single automated process could be trusted to run.
That was a large firm with significant resources. The lesson translated to every scale.
The same problem in a 20-person business
The dynamics are different at SMB scale, but the underlying issue is identical: without reliable, consistent data, AI has nothing solid to stand on.
What tends to happen in smaller businesses is slightly different. It's rarely multiple enterprise CRMs in conflict. It's more often a CRM that's only partially kept up to date, alongside a spreadsheet that someone maintains separately, alongside a folder of email threads that are the actual record of what clients have agreed to. Three partial truths, none of them complete.
When I start working with a new client, one of the first things I ask is: if you needed to contact every active customer with the same message right now, how would you do it? Which system would you pull that list from? How confident would you be that it was complete and current?
The answers to those questions tell me almost everything about the state of the data — and whether you're ready for something like data-driven AI workflows.
Why AI makes this worse, not better
The important thing to understand about AI systems is that they don't know when the data they're reading is wrong. A human looking at a client record with conflicting information will pause, check, ask. An AI system will read what it's given and act on it with complete confidence.
That confidence is the danger. A marketing automation workflow that sends a renewal reminder to an outdated email address, or drafts a personalised message using a client name that's slightly wrong, or misses a customer because their record is duplicated in the system — these aren't AI failures. They're data failures that the AI has efficiently scaled up.
This is why, when I'm scoping an AI implementation with a client, one of the first questions is always: where does this workflow get its data from? And how confident are we in that data? If you don't have a single authoritative source, that's the conversation to have first.
Not every field needs to be perfect — that's an unrealistic standard and not actually what matters. What matters is that the fields the AI will act on are clean and current. If you're automating invoice reminders, the fields that matter are the customer's name, their email, and the outstanding amount. Those need to be right. The fact that their industry sector is recorded inconsistently across your CRM is irrelevant for that specific workflow.
The deeper point about AI and data
There's a tempting assumption in a lot of AI conversations right now, which is that the technology is clever enough to work around imperfect data. That it will infer, interpret, and fill in the gaps. Sometimes it can, up to a point. But acting on inferred data at scale is a risk that most businesses aren't aware they're taking.
The most robust AI implementations I've seen — and helped to build — share a common characteristic: before the AI was anywhere near the data, someone sat down and asked whether that data was trustworthy. Not perfect. Trustworthy, for the specific purpose at hand.
That's a question that benefits from a CA's instinct, not just a technologist's. I'm looking at data the same way I'd look at a set of management accounts: is this telling me something I can act on, or is it a plausible-looking story that might not hold up under scrutiny? If you're not sure, start with tracing where your data actually comes from.
Where to start
If you're thinking about AI and worried about whether your data is ready, the right starting point is a specific question, not a general audit. Pick the one workflow you'd most like to automate — the task that eats the most time. Then work backwards: what data does that workflow need? Where does that data currently live? How accurate is it?
That conversation — which might take a few hours with the right person — will tell you more than months of vague data quality improvement work. It's also the step most businesses skip before implementing AI — I wrote about that in mapping your business systems first.
If you'd like a second pair of eyes on that question, book a free 30-minute call. No pitch. Just a straight answer on whether your data is ready, and what to do if it isn't.