- With money, the AI danger is a wrong answer that looks right. Two Monzo engineers say calculations should live in tested, versioned code.
- They also warn against agents writing new calculation tools on the fly, because the same question can then get different answers.
Where money is involved, the real AI risk is not a slow answer. It is a wrong answer that looks right. Francesco Galletta and Felippe Felisola Caso made this case on the AI Engineer show. They work at Monzo but did not speak for it. Their fix is to move arithmetic out of the model and into tested, versioned code. They also want to stop agents inventing calculation tools as they go.
What was said
The pair built a financial-analyst agent live, using the cost of a wedding as the problem. A person in the room described asking an agent for a formula. It came back neatly formatted, yet the result was plainly wrong.
A speaker on the show described a similar problem with a monthly budget. They let the agent sort their transactions and total them. The results differed slightly each time. The agent got it right three times, then failed on the fourth. So they turned the job into a script.
The rule they gave is to write the calculation once and run it every time. The model can choose which tool to use, but the harness, the code around the model, must stop it doing arithmetic itself. Without that, a speaker said, "Again, you're asking the stochastic machine to produce a consistent output, so you're likely not going to get it."
They also questioned the habit of agents building their own tools mid-task. By the hundredth run, a speaker said, the agent may write new code for a problem already solved. Their answer was tests, known-answer examples and version tracking, so the agent warns you when calculation code has changed. A speaker summed up the stakes: "When we're talking about money, 1% wrong sometimes is a lot of money."
Why it matters
Our reading: this moves the question for buyers and builders. Asking whether an AI tool is accurate is too vague. Ask where each number comes from. If it comes from a function with tests and a version history, you can audit it. If it comes from the model's own working, you cannot.
It also gives managers a simple check. Ask the same financial question twice and compare. If the answers differ, the arithmetic is probably not in code. A speaker also said that running many agents does not remove your accountability for what they produce.
The other side
The speakers raised the strongest objection themselves. Agents are improving and now call tools more often, so one view is to wait. They disagreed, but the excerpts give no data on how often models fail at arithmetic today.
The demo was also a wedding budget, not a production system. A speaker said they held back some detail because of intellectual property.
Code is not safe just because it is code. A speaker told of a shared master spreadsheet whose assumptions changed without the user knowing, and said the same blind spot can hit code if changes are not traced. A speaker also noted that accounting involves interpretation, so not every financial judgment reduces to a formula.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo (2026-10-17)
