In October 2025, University of Chicago researchers confirmed that the best AI can’t multiply two four-digit numbers, a problem fourth graders do with a pencil. The AI models they tested were wrong a whopping 99% of the time. Strange as it sounds, that failure has been very good for AtScale’s business.
Anyone who pays attention while they use AI has experienced this. AI writes and reasons like a genius, but it’s not built to do arithmetic, so they guess.
AI’s failing math grades manifest in many ways:
- Chicago’s research found that AI quickly answers a multiplication like 3,932 × 8,442 with a number that looks right and usually isn’t. It guesses the shape of an answer rather than doing the math.
- KTH Royal Institute of Technology (Sweden’s leading technical university) tested models on math and found even GPT-4 makes basic arithmetic slips, with accuracy sliding toward 20% on the harder sets.
- Anthropic ran its own analytics on Claude: right about 21% of the time on raw tables, and about 95% once grounded in a governed semantic layer.
That’s where AtScale comes in. Computing “revenue” is even harder than multiplying two numbers: you have to find the right database tables and pin down which “revenue” the question means: bookings, recognized revenue, new revenue, or ARR. That step alone requires guessing: is the number in the bookings_amt column, rev_recognized_usd, or net_new_arr?
For business analytics, automation, and cowork, accuracy is non-negotiable. The minute AI touches revenue, inventory, or profitability, “close enough” is worthless. Put one wrong figure in front of executives, and they stop trusting their entire investment in AI.
Better prompts or skills can’t fix it. One bank clocked its model guessing 21,000 times at numbers a single call could have returned. They started calling it their guessing tax. Then they tried building Claude skills to do the math. Then they learned that’s a hard problem. Then they bought AtScale.
AI needs a calculator for your business metrics. That’s what AtScale is, and why Snowflake studied this problem, invested in us, and embedded AtScale as the only outside product inside Snowflake Semantic Views. GigaOm named us a leader, and more big enterprises standardize on us every quarter. We just finished the best quarter in our history, and we’re hiring like crazy to keep up.
AtScale’s special sauce is the AI Computation Engine, or ACE. It knows what your business means by “revenue,” “pipeline,” or “inventory,” works out the answer, and checks it against a gold standard (we’ve measured it with the BIRD benchmark; results coming soon). Let AI be creative where it augments human creativity. When the number has to be exact, have ACE jump in. And more and more companies are buying an AI calculator.
For most software companies, the AI news has been grim. The 2026 SaaSpocalypse erased roughly a trillion dollars in software value as investors bet AI agents would gut the incumbents, and even Salesforce fell more than 25% on the year. AI is making semantic layers more important, not less.
In the end, it comes down to one line: guessing is not a strategy. AI can do remarkable things, and it still can’t be trusted to do the math a business runs on. We built the calculator that stops the guessing. That’s why we’re growing, and as long as AI can’t read your mind about how your enterprise data is structured, and can’t be trusted to do math on its own, we’ll keep growing.
Reviewed by: Dave Mariani
SHARE
Guide: How to Choose a Semantic Layer