Opinions expressed by Entrepreneur contributors are their own.
Key Takeaways
- AI models predict numbers; they don’t calculate them. Any figure a customer will act on — like a price quote, shipping estimate or tax figure — should come from a deterministic tool, not the model itself.
- If running the same request 100 times must return the same answer, the model shouldn’t produce that answer. Let the model read the question and pick the right tool. A deterministic engine should then work out the value.
- Sounding right isn’t the same as being right. Even strong models make calculation, rounding and formula errors, and they can flip correct answers to wrong ones under pushback.
As many as 77% of small and midsize businesses in the U.S. use AI regularly, and customer service is one of the top three uses.
The trouble starts when that answer is a number the customer acts on, because a large language model (LLM) predicts numbers rather than working them out.
The number is your liability
When a model drafts an email or ad copy, there’s usually no single correct answer because it’s a creative task. But when a product returns a number the customer treats as a settled fact, that’s where the hazard lies. Examples of such numbers include:
- Price quote
- Shipping estimate
- Tax figure
- Dosage
- Unit conversion
- Financing payment
A wrong number can read like a right one, which can erode trust in the company when discovered. A Toronto BMW dealership learned this lesson the hard way.
A man who wanted to sell his car sent an inquiry to the BMW dealership he’d bought it from, and got a text back from “Quinn” offering 27,162.79 Canadian dollars, about $19,000, to buy it back. Quinn was an AI chatbot, and that figure was not a valuation, but the balance the man still owed on his loan, handed to the bot by mistake and passed along as the price. A salesperson later called to revoke the offer, and the dealership only reinstated it after CBC News asked for comment.
The test is reproducibility
Before we build anything at Omni Calculator, where the model produces a value our users act on, we ask ourselves if the output must be reproducible. If running the same request 100 times must return the same answer, that answer must come from a deterministic tool the model calls, rather than from the model itself.
That test is the premise behind Omni Calculator Builder (now in public beta). You describe the calculator you want in plain language, the model turns that description into calculator logic, and the logic runs on our deterministic math engine. The model designs the tool, but the number itself comes from the engine.
That’s because the model predicts text one token at a time, so it doesn’t hold the value of a number the way a calculator does. Our research backs this up:
Earlier this year, we ran the third iteration of the ORCA Benchmark (Omni Research on Calculation in AI), a study of how free-tier AI models handle math, among others. We found that accuracy ranged from 48.4% for ChatGPT 5.3 to 70.4% for Grok 4.20, with Claude Sonnet 4.6 between them at 53.2%.
The recurring failures were due to raw calculation errors, rounding issues and broken formulas. What’s even more interesting is that Claude and ChatGPT turned a correct answer into a wrong one 60-65% of the time when a user pushed back with “Are you sure?”
In other words, if the answer has to be right every time, the model won’t cut it.
The fix is to keep the model off the math
Since the model won’t get the math right every time, split the work so it never produces the final number itself.
Let the model read the question and pick the right tool. A deterministic engine (e.g., a script, a calculation API) should then work out the value, leaving the model to write the reply around that number.
Picture a mortgage lender whose website uses a model like Claude or ChatGPT to help visitors work out what they can afford. Someone enters their income, debts, down payment and rate, then asks what the monthly payment would be. Rather than answering from the model, the site should pass those inputs to something that runs the amortization formula (the same way, every time) and returns an exact figure.
One way to implement it is to connect the model to Wolfram Alpha as a tool it can call. When the payment question comes in, the model pulls the numbers out, hands Wolfram the amortization formula with those values and uses the figure Wolfram returns. It writes the reply around that number, so the visitor reads a normal paragraph whose key figure was worked out by a deterministic tool before the model wrote a word.
Where this leaves you
Models are getting smarter and can now answer virtually any question, but sounding right and being right are different things, and your customer won’t always be able to tell them apart (at least, not at first).
So before you give them a number to act on, route it to a deterministic tool the model calls, and keep the model on the language.
Key Takeaways
- AI models predict numbers; they don’t calculate them. Any figure a customer will act on — like a price quote, shipping estimate or tax figure — should come from a deterministic tool, not the model itself.
- If running the same request 100 times must return the same answer, the model shouldn’t produce that answer. Let the model read the question and pick the right tool. A deterministic engine should then work out the value.
- Sounding right isn’t the same as being right. Even strong models make calculation, rounding and formula errors, and they can flip correct answers to wrong ones under pushback.
As many as 77% of small and midsize businesses in the U.S. use AI regularly, and customer service is one of the top three uses.
The trouble starts when that answer is a number the customer acts on, because a large language model (LLM) predicts numbers rather than working them out.
The number is your liability
When a model drafts an email or ad copy, there’s usually no single correct answer because it’s a creative task. But when a product returns a number the customer treats as a settled fact, that’s where the hazard lies. Examples of such numbers include:
This story originally appeared on Entrepreneur
