In 2025 I was the technical lead of a WhatsApp assistant for woodworkers. The woodworker sent a voice message with the customer name, the material cost and the deadline. The assistant returned the quote as a PDF.
Understanding the audio worked. What blocked the product was the math.
The mistake
The same model that interpreted the speech also added up materials, applied the margin and calculated the total. Most of the time it got it right. In a quote, most of the time is not enough: a wrong total goes to the end customer with the woodworker’s name on it.
A language model writes the number that looks right. It does not calculate.
What I do today
I split the work in two.
- The model extracts. It reads the message and returns data in a fixed format: items, quantities, amounts and deadline. If a field is missing, it asks.
- The code calculates. A plain function, with tests, does the sum, margin, discount and installments. The model never touches the result.
The final answer joins both: the text comes from the model and the numbers come from the code.
How to tell if your agent has this problem
Look for any place where the agent reports a value it calculated itself: installments, shipping, discount, a deadline in days. If the number does not come from a function or a table, sooner or later it will come out wrong.