# What 10,000 verified claims found _September 18, 2026 · Jeff, CEO_ > A look at the error rates of enterprise AI data analysis in production. The AI industry talks about hallucination in terms of stray facts and bad creative writing. But when you deploy frontier models to analyze enterprise data, the failure mode changes. The AI doesn’t start making up imaginary products. Instead, it gets the math wrong, misreads the source API, and silently passes the error to a decision-maker. Emet has verified more than 10,000 claims in AI-generated analyses on live client data environments. Here is what the numbers actually say about the state of enterprise AI data analysis. ## The baseline error rate is 20% In our production measurement, 1 in 5 claims in AI-generated analyses carried an error. This isn't on legacy models. These are frontier models running real business queries against structured data sources like Stripe, Meta Ads, and GA4. The errors fall into three main categories: 1. **Unsupported numbers:** The AI claims a metric that cannot be found in or derived from the source data. 2. **Bad math:** The AI pulls the correct raw inputs but fails the calculation step (e.g., miscalculating MoM growth). 3. **Misattribution:** The AI pulls a real number from the data but assigns it to the wrong entity, time period, or dimension. ## Errors compound silently When a human analyst makes a mistake, they often leave a trail of confusion or caveat their confidence. When an AI makes a mistake, it does so with absolute certainty. If an agent queries one week instead of the whole quarter in step two of a twenty-step analysis, every subsequent derivation inherits the error. The final recommendation reads perfectly well, but the foundational math is flawed. The failure mode is silent. ## The solution is deterministic verification You cannot prompt engineer your way out of a 20% error rate. You cannot rely on the model to grade its own homework. The only way to trust AI with data analysis is to re-derive every figure deterministically in code against the source data. That is what Emet does. Emet reads the trace and re-derives the math before the decision is made. Every number needs a receipt. --- **[Book a Demo →](https://emet.so/contact)** Emet · Truth infrastructure for AI data systems · https://emet.so