The messy layer underneath AI

When someone says "we want to use AI", they almost always mean: we want an AI that can do something useful on top of what we already have.

The problem is rarely the AI layer. It's the layer underneath.

In 80 percent of the audits I do, the first real observation isn't "here are the AI opportunities". It's "here are five different systems that don't talk to each other, and an Excel sheet gluing them together".

A typical example: a company has Salesforce as a CRM, Visma for accounting, Tripletex for some clients migrated in 2023, Microsoft 365 for documents, and a shared OneDrive folder where every important PDF ends up. Sales updates Salesforce. Accounting reads Visma and Tripletex. Nobody reads each other's systems. When an invoice comes in, someone has to find the customer in CRM, check payment status in Visma, and update a shared report in Excel.

The AI layer — the auto-generated summary, the smart search, the agent that "takes over customer follow-up" — can't function on top of this.

You can run Copilot against SharePoint, but if half the data lives in Visma and the rest in Salesforce, Copilot doesn't see what Visma sees. It will give confident answers based on half the picture.

That means the first job in almost every AI implementation in the SMB segment isn't AI. It's cleanup.

It's boring. It's also where most of the value lives.

Concretely: if you move two systems onto one platform, eliminate the Excel middle layer, and define a single source of truth for customer data, you've done something that pays for itself regardless of whether the AI layer ever materialises. And when the AI layer arrives, it will actually work.

People who have been in enterprise IT for twenty years know this. People who have been in AI for twenty months don't. It's the most common bridge-the-gap conversation I have — and the one that actually shifts a managing director's priorities.

If the messy layer is too big to clean up in one operation, there's still a rule that helps: build your AI around the one data source that is least messy. Start there. Let the rest wait.

Not because the mess doesn't matter. But because you need a win first to convince the organisation that cleanup is worth the trouble.