- A speaker on the AI Engineer show argued that an AI analyst cannot separate hard facts from estimates, so the surrounding system must enforce the labels.
- Their scheme: bank statements count as fact, chat remarks as interpretation, and the agent's own recall as assumption, each with a traceable source.
An AI financial analysis can go badly wrong when a guess is treated as fact. That was the argument on the AI Engineer show. The guests were Francesco Galletta and Felippe Felisola Caso. Both work at Monzo but did not speak for it. They argued that the software around the model, called the harness, must label each input as fact, interpretation or assumption.
What was said
The guests built a financial-analyst agent live. Their problem was paying for a wedding. A speaker gave a small example of the failure. The guest count was treated as a fact. But the fiancé had invited three more people without saying so. The whole budget model was then wrong.
The speaker said a wrong assumption can be fatal to any financial analysis if nobody knows where it came from. The core claim was blunt: "The model can't tell what is a hard truth and what is an estimation." So the harness has to enforce the difference.
The labels were simple. A bank statement is a fact. Unstructured thoughts typed into a chat count as interpretation. The agent's own reasoning, or anything it pulls from memory, counts as an assumption. Each kind of evidence gets its own level of importance.
The speaker also wanted a trail back to the source. They told of a shared master file. It held built-in assumptions about reasonable values. The file changed, and the user did not know. The inputs were the same, yet the results were wrong. The same blind spot, they said, applies to changes in a harness's own code.
Why it matters
Our reading: this moves trust from the model's answer to the plumbing around it. If you build such a tool, record the type and source of every input when it enters the system. If you buy one, ask the vendor how a number can be traced back to its origin.
A speaker on the show also said that even with 100 agents running, "you're still accountable for them." A tool that mixes guesses and facts without labels leaves a person answerable for numbers they cannot check.
The other side
By the speaker's own account, the wedding was a deliberately simple example. The excerpts do not show how well the labelling works on large, messy, real-world analyses. They also do not say who assigns a label when the line is unclear.
That line may be blurry. A speaker on the show also said that in accounting, "it's all about creativity and interpretation." A label shows where a number came from. It does not make the number right. People still have to judge it.
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)
