The same AI scored 21%. Then 95%. Nothing about the AI changed.

Anthropic, the company behind Claude, recently published something unusually candid: a detailed account of how they got their own AI to answer their own business questions. Revenue, usage, customers. The same questions every leadership team asks.

The headline number is startling. When they pointed Claude at their data cold, it answered correctly 21% of the time. After they finished the work described below, the same model scored 95% and higher. Nothing about the AI changed. Everything around it did.

Why the smartest AI gets your numbers wrong

The failures weren't exotic. They came down to three very human problems.

• Which “revenue”?A real company's data has dozens of fields that could plausibly be revenue. Booked or collected? Before or after refunds? The AI doesn't know your answer, so it guesses. Confidently.
• Stale knowledge. Businesses change constantly. New products, renamed fields, revised definitions. Documentation written last quarter quietly becomes wrong.
• Needle in a haystack. The right answer usually exists somewhere in the data. Finding it among millions of fields is where the AI got lost.

Notice what's missing from that list: intelligence. The model was never the problem.

What actually took them from 21% to 95%

Four layers, stacked in order.

• One clean foundation. A single, governed home for the data. One canonical version of each concept instead of forty lookalikes.
• Definitions owned by humans.What “revenue” and “active customer” mean is decided by people, and written down where the AI is required to look first.
• Context that lives with the data. Short reference notes: which tables to use, how they join, the gotchas. They sit in the same place as the data itself, so changing one updates the other.
• Continuous checking. Every answer traceable to its source. Accuracy tested against questions with known answers. Every human correction folded back in.

One more finding deserves its own paragraph. When they paused the maintenance, accuracy rotted from 95% to 65% in about a month. The truth layer isn't a project you finish. It's a system you keep alive, because your business keeps changing.

What this means if you run a company

Every tool you own is about to offer you an AI that answers questions about your business. The pitch is irresistible, and the demo always works. Anthropic's research is the honest fine print: the AI is only as trustworthy as the foundation underneath it. Point it at scattered, inconsistent data and you get fast, confident, wrong answers. Those are the most expensive kind, because they look right.

So the unlock isn't buying AI. It's the unglamorous part: one clean foundation, definitions your team owns, context that stays current, and a loop that keeps checking. That's the part we build. It's the entire reason Strata exists. If you want to see what that foundation would look like for your business, the free roadmap maps it in about ten minutes. And the original write-up is worth your time: How Anthropic enables self-service data analytics with Claude.