Why can’t a large language model make regulated financial decisions?

A large language model can't make a regulated financial decision because probabilistic intelligence is not institutional authority. The model produces the most likely output given what it has seen. A regulated decision has to be one the institution has authorized, can explain and can reproduce when asked. A better model improves the output, but it doesn't turn that output into a decision the institution can stand behind.

Probability is fine inside the model, as long as the institution keeps firm control over what actually gets decided.

What is the difference between an answer and a decision?

An answer responds to a question. A decision is a position the institution takes and has to be able to defend later.

If a client asks whether they can afford a home, an answer is a plausible sentence. A decision is a determination made against the bank's liquidity thresholds, suitability requirements, product criteria and approval structure, and the bank has to stand behind it when a regulator, a supervisor or the client asks why.

When a system produces an answer and presents it as a decision, nothing has really been automated. The decision is simply missing the accountability it needs.

Isn't this solved by a better model?

No. The limit is authority, and a more accurate model doesn't change that.

A more capable model gives better answers more often. It still doesn't hold the institution's rules, and it can't be the party that decides what the institution permits. It makes a better estimate, while a regulated decision needs an outcome the institution controls.

Something real has changed, though. AI can now handle far more complexity without becoming the institution's authority. That is what makes this approach workable today, and it also sets the limit on what the model should be asked to do.

What about putting the model behind the institution's rules?

That instinct is right, and the details decide whether it works.

If the rules are handed to the model as context and the model is asked to apply them, the model is still the one making the decision, just with more information in front of it. The outcome stays probabilistic. Ask the same question twice and the bank can't promise the same result.

The alternative is to use generative AI to help build the institution's assets ahead of time. Rules, calculations, decision paths and approved guidance are drafted with AI assistance, then reviewed, approved and executed deterministically. AI does a lot of the work of creating the intelligence, and the policy is settled before a client ever asks a question.

What should a system do when it lacks the facts to answer?

It should stop and say what is missing.

Consider a homebuyer whose income, cash, investments, liabilities and liquidity needs are all verified, while the planned renovation spend is unknown. The right move is to ask for that number instead of estimating it. We call this step “not yet, ask this first”.

Treat a missing fact as a reason to pause. A system that always produces an answer can't flag the one case where it shouldn't, which is why every fact it holds should carry its source, its date, a confidence level and a note of what is still unknown.

Is an LLM wrapper enough to deliver financial advice?

No. An LLM wrapper in front of existing systems is a demo, and a regulated institution can't deploy a demo as its architecture.

The difference shows up in what the institution owns once the model has finished. With a wrapper, the institution is left with a set of responses. A proper architecture gives it governed institutional assets instead: rules, calculations, decision trees, required information, approved guidance and action pathways that it owns and can inspect, change and reuse.

What does a deployable architecture look like?

It has four layers, and governance applies inside each of them while the system runs:

The simplest test of the whole design is one question: why did you tell her that? A deployable system can rebuild the answer from the facts, the rule, the calculation, the path, the model's contribution, the approval and the client output. There is more on how this works in practice on our governance page.