Companies are using AI to analyze data, answer business questions, and generate reports. But most leaders, if they're honest, will admit they're not totally sure they can trust what comes back.

A number shows up in a slide and it looks right. But was it pulled from the right source? Did the AI understand what you were actually asking? Would it give the same answer tomorrow?

Last week, Anthropic published a piece that gets at why this problem is bigger than most companies realize.

Companies assume if you give AI access to good data, you get good answers. That assumption is wrong.

When your CFO asks how revenue is tracking, she doesn't mean any definition of revenue. She means a specific definition used by your org, calculated in a specific way, and excluding specific edge cases that  already got resolved. 

None of that is in your database, because it’s institutional knowledge. It lives in specific tools, in a few spreadsheets, and the heads of the people who've been around long enough to know it.

Your AI has access to the database, but it doesn’t have access to your institutional knowledge.

So it does what any intelligent system does when it lacks information: it makes a reasonable inference. And a reasonable inference, applied to business-specific data with company-specific definitions, is wrong a lot of the time. 

Anthropic ran an experiment to find out why AI keeps messing up

Anthropic ran a series of experiments to understand what made AI accurate or inaccurate when answering data questions.

When they gave Claude access to all their data but nothing else, it answered business questions correctly 21% of the time. That’s 1 in 5 questions wrong, from the company that builds the AI.

Their instinct was to throw more at it. So they gave the AI access to all their previous work: years of dashboards, reports, and analysis. Surely if Claude could see how every question had been answered before, it would do better. When they tried that, accuracy moved by less than 1%. 

Claude could see the right answer, sitting in previous work, and still couldn't use it correctly. Access to information is not the same as understanding it.

What actually worked was building a knowledge layer: a structured, curated translation of their data into business meaning. It meant making institutional knowledge that previously lived only in expert heads legible for the machine. The team needed Claude to understand: 

• What each metric actually means at Anthropic.

• Which definitions are canonical. 

• What context is required to interpret each number correctly. 

With this knowledge layer, accuracy jumped to 95%.

But when the Anthropic team stopped manually maintaining the knowledge layer for a month, accuracy fell to 65%.

Most companies don’t have large engineering teams like Anthropic 

Anthropic's solution works. A manually built, carefully maintained knowledge layer, colocated with your data code, enforced through code review, updated as part of every schema change, supported by evaluation pipelines that measure accuracy over time. If you build and maintain that, you get 95% accuracy.

The problem is what it takes. Anthropic has a dedicated team of data scientists and engineers who treat knowledge maintenance as a first-class engineering problem. They have CI pipelines that flag when a data model changes without a corresponding knowledge update. They have systems that harvest corrections from analyst feedback and turn them into documentation updates. This is a significant ongoing investment even for a company whose entire business is AI.

Most companies can’t do that because they have finite resources. They need accurate AI without the overhead. 

The knowledge layer has to be self-maintaining to be useful 

The lesson from Anthropic's experiment is that companies need a knowledge layer, and one they can easily maintain. 

The knowledge layer must be constantly updated. Otherwise, accuracy degrades, and quickly. 

Unfortunately, the maintenance burden is huge.

Sphinx is built to solve that. It creates the knowledge layer and keeps the layer alive, detecting when definitions have gone stale, pulling in corrections from real usage, and updating itself as your business evolves. It’s self-learning and self-correcting.

With Sphinx, your AI tools actually understand your business well enough to answer questions correctly, without requiring an engineering team dedicated to building and maintaining the knowledge infrastructure by hand.

Every enterprise using AI needs a knowledge layer. The only question is whether you want to absorb the costs of building it and maintaining it yourself, or whether you want it to run automatically.

See what your AI is missing

Sphinx automatically builds the knowledge layer your AI is missing, keeps it current as your business changes, and gives you visibility into where your AI is getting things right and where it’s not.

Book a demo and we will show you exactly how it works.