A recent Anthropic example highlights a problem in enterprise AI that is easy to underestimate.
Without the right context, Claude answered only around 21% of analytics questions correctly. With the right Skills, that number rose above 95%.
Sit with those two numbers, because the gap between them is the whole argument. Same model. Same questions. The only thing that changed was the context it was given. At 21% the model is worse than useless for analytics, it’s confidently wrong four times out of five. At 95% it’s a genuinely reliable tool. And nothing about the model’s raw intelligence moved between those two states. What moved was whether it understood the business it was answering questions about.
The key message is that Skills matter. But the bigger one is that AI performance depends heavily on context.
The tax nobody puts on the invoice
Every time an employee has to review, correct or verify an AI-generated output, the company is effectively paying the Hallucination Tax.
This is the cost that hides. When people talk about the price of AI they mean the subscription or the API bill. But there’s a second cost that never shows up on an invoice. It’s the analyst who has to double-check every figure the model produced before trusting it. It’s the manager who reads the AI summary and then reads the source anyway, because they can’t be sure. Every one of those review-and-correct cycles is human time spent cleaning up after the machine. That’s the Hallucination Tax. It’s real, it’s recurring, and because it’s paid in scattered minutes across the org rather than in one line item, most companies never add it up.
According to one survey, 55% of technology leaders said they sometimes have to personally correct AI outputs, while 89% believe AI analysis cannot be fully trusted until the underlying data is reliable and verified.
Look at who’s doing the correcting. Not junior staff. Technology leaders, 55% of them, personally fixing AI output. That’s some of the most expensive time in the building spent on cleanup. And 89% saying the analysis can’t be trusted until the underlying data is reliable points straight at the cause. The model isn’t the weak link. The data and context feeding it are.
The fix is a layer, not a bigger model
The core solution is to build the semantic and governance layer that gives those models a reliable understanding of the business.
This is the move most people skip because it’s less exciting than swapping in a newer model. A semantic and governance layer is the thing that tells the model what your data actually means. What “revenue” refers to in your systems and which of the six tables it lives in. Which numbers are authoritative and which are drafts. What the rules and definitions of your business are. Without it the model is guessing from raw fields and column names, which is how you get 21%. With it, the model reasons over a reliable, defined picture of the business, which is how you get 95%. The layer is the fix. Not a smarter model on top of messy data.
Why this gets worse, not better
And as AI agents move from answering questions to taking actions, the cost of unreliable AI becomes far more serious. A wrong answer may waste time. A wrong action can create real operational, financial and reputational consequences.
That escalation is the reason this isn’t optional. Today the failure mode is a bad answer, and the Hallucination Tax is a person catching it. But agents don’t just answer anymore, they act. They update records, move money, send messages, trigger workflows. When an agent acts on a wrong understanding, there’s often no human in the middle to catch it first. A wrong answer wastes an afternoon. A wrong action can move real money, break a real process, or damage a real relationship with a customer. The governance layer stops being a nice-to-have the moment the AI can pull the trigger itself.
In the future, competitive advantage won’t come from simply having access to the best LLM. It will come from the quality of the data, context and business knowledge organisations can give them.
The best models are becoming a commodity everyone can buy. What isn’t for sale is your data, your definitions, your business knowledge, cleaned up and structured so a model can actually use it. That’s the part nobody else has and nobody else can copy. That’s where the edge will be.
Image: “digital warning error message with glitch effect” by muhammad.abdullah / Magnific.com. Free for commercial use with attribution. Source: https://www.magnific.com/free-photo/digital-warning-error-message-with-glitch-effect_427035725.htm