Access isn't the hard part. Getting the right answer is.

Anthropic recently published how their own data team uses Claude to answer questions about their business data (the full post is here). If you run a business on your data, it's worth your time. It explains something most AI tools skip over. Connecting AI to your data is easy. Getting it to give you the right answer is the hard part.

Here's the bit that stopped us.

Anthropic's data team gave Claude direct access to thousands of files of prior queries and analysis, a record of every question their team had already answered correctly. They checked it actually read them before answering. Accuracy barely moved. Less than a single percentage point.

The right answers were in there, most of the time. The AI could see them. It still couldn't use them.

Their conclusion was simple. Access was never the problem. Structure was.

So why does AI get business data wrong?

Anthropic point to three reasons, and they'll be familiar to anyone who's tried this.

First, the AI can't always tell which bit of data you mean. Ask for "revenue" and there might be dozens of fields that could count. Pick the wrong one and the number's wrong, but it still comes back looking confident. Anthropic put it well, warning that pointing AI at raw data can create "a false sense of precision."

Second, data changes. Fields get renamed, definitions shift, and the AI's understanding quietly goes out of date.

Third, there's simply too much to search. The right answer might be in there, but the AI can't find it in the noise.

Connecting AI to your data is the easy bit

This is where "my system already connects to Claude" runs out of road.

A single connection can show AI one system's raw data. That's the easy 20%. And it's exactly the kind of setup Anthropic warns about: access without structure. Raw tables, cryptic field names, no steer on which to trust.

The hard 80% is everything that turns raw data into a right answer. And it's the same job for a global AI lab as it is for a plumbing firm running Xero and Simpro. You need a structured, governed layer in between.

In Anthropic's own testing, adding structured guidance on what their data meant and how to use it took accuracy from under 21% to over 95%.

What actually makes the answers right

That layer is what we've spent years building. Here's what's in it.

We bring your data into one structured place and give every field its meaning back. Cryptic IDs become business names. Deleted and duplicate records are filtered out. And when Claude connects, we point it at the clean, governed version of your data rather than the raw tables underneath.

Every system we connect ships with hand-built, tested descriptions of what its data means and how it fits together. So Claude isn't guessing its way around an API it's never seen.

The AI always reads the current shape of your data, so its understanding doesn't drift as things change.

Your data stays fresh, structured and ready to query. That's the right setup for reporting and analysis. A stable, up-to-date picture, not a moving target.

And the same governed data answers both your Power BI reports and your AI questions, so the number's the same wherever you ask. One source of the truth.

It's all permission-scoped and audited too. Self-service, without losing control.

One question, every system

Here's the part even Anthropic's write-up doesn't cover, because they're working within one warehouse.

Real businesses don't run on one system. They run on a job system, an accounts package, a CRM, and more. And the questions that matter most span all of them. "Which jobs went over budget, and what's that done to this month's profit?" That's your job system and your accounts, answered in one go.

One group we work with runs a dozen-plus Xero entities across multiple Simpro builds. No single connection can report across that lot. Bring it into one structured place and you can just ask, across the lot.

That's the real difference. It isn't connecting AI to a system. It's connecting it to your whole business.

Where this goes

We're not done. Not even close. The deepest layer of all is your own business context, the things only you know about how your data works, and that's a big part of where our focus is now.

But the foundations are live today. Anthropic just described the stack that makes AI analytics work. We've built it into a platform, so you don't have to.

Curious what AI Insights could tell you about your data? Take a look at AI Insights. Book a demo, or start a 10-day free trial.