Why FinAI never lets the language model pick the number

Ask a general-purpose chatbot what a balanced portfolio for a 30-year-old looks like and you will usually get a confident answer with percentages in it. The percentages read well. The question I kept asking while building FinAI was simpler: where did that number come from?

In most chat products the honest answer is that the model produced it, the same way it produces every other word. That is the problem FinAI is designed around.

A fluent sentence is not a calculation

Ji and colleagues (2023) surveyed the research on hallucination in natural language generation: output that is fluent and plausible but unsupported by the source or simply false. Their review covers summarisation, dialogue, question answering and more, and the pattern is consistent. A language model is optimised to produce likely text, not verified text, and it gives no reliable signal when it is guessing.

In most domains a plausible wrong sentence is an annoyance. In finance it is worse than no answer. "Your expected return is 7.2%" looks precise, and the precision is exactly what makes a reader trust it.

Let programs do the arithmetic

Research on tool use points to a practical fix. Gao and colleagues (2023) showed with program-aided language models (PAL) that letting the model write a program and handing the computation to an interpreter improves accuracy on arithmetic reasoning benchmarks over having the model compute in text. Schick and colleagues (2023) showed with Toolformer that models can learn when to call external tools such as a calculator. Lewis and colleagues (2020) showed with retrieval-augmented generation that grounding a model in retrieved documents makes its answers more factual on knowledge-intensive tasks.

The shared idea: the language model is good at deciding what to do and at explaining a result. It should not be the thing that produces the result.

Four beats

FinAI follows that split strictly. Every question goes through four steps, and only one of them is allowed to create a number.

Route. The question is classified before any sentence is written. "Is inflation bad for stocks?" is a macro question and goes to the macro engine. It does not get a portfolio template because the word "stocks" appeared.

Compute. Deterministic engines do the work: market data from providers, allocation shaped by risk, goal and horizon, Monte Carlo simulation, compound-interest and ROI calculation. The same inputs give the same output. The language model chooses which tool runs. It does not supply the number.

Narrate. The model explains the engine output in plain language. If the engine did not produce a price, a weight or a percentage, that figure does not appear in the explanation.

Present. The answer arrives as a structured card with the numbers and their context, not as a wall of chat. A Monte Carlo result is shown as a range of outcomes, not one guessed line, because a single line hides the uncertainty the simulation exists to show.

What this design does not solve

Education, not advice

FinAI is educational decision intelligence. It is not regulated investment advice and it is not a licensed advisor. Under MiFID II, investment advice means personal recommendations to a client about specific financial instruments, and FinAI is built to explain concepts and model scenarios, not to make that kind of recommendation. The EU AI Act adds transparency duties, including telling people when they are interacting with an AI system. FinAI is presented as exactly that: an AI system that explains engine output.

That boundary matters for the learning side too. FinBild lets people feel time value and diversification before they are named. FinAI is where someone takes that intuition to their own question, and gets numbers that came from a calculation rather than from a sentence. You can see the pipeline step by step on the FinAI page.

References