There is a great deal of enthusiasm about putting artificial intelligence to work in wealth management, and a great deal of unease about it. Both are justified, and both point to the same question: not whether to use AI, but where to draw the line around it.
Large language models are genuinely extraordinary at one thing: turning unstructured human language into structure. They can read a rambling conversation about a family's hopes and worries and draft a coherent policy from it. They can tailor a question, summarise a position, and narrate a report in fluent prose. Used for that, they are a remarkable productivity tool.
They are also, by construction, probabilistic. A language model does not compute; it predicts the next plausible token. Ask it for a number and it will give you one that looks right, generated by the same machinery that occasionally invents a court case or a citation. For a portfolio recommendation that a family will live with for years, and that a regulator may examine long after the fact, plausible is not a foundation.
Where the line belongs
The sensible division of labour is clear once stated. The model handles language: it extracts a policy from prose, proposes at most a few clarifying questions, and later narrates a report around figures it is not allowed to alter. A separate, deterministic engine handles every number: expected return, volatility, stress, the allocation itself. The model writes the words around the maths; it never does the maths.
This is not a cosmetic distinction. A deterministic engine produces the same figures for the same inputs, every time. That single property is what makes a recommendation auditable. You can reproduce it, trace it back to the policy it came from, and defend it to a client or a regulator with evidence rather than assurance. A figure you cannot reproduce is a figure you cannot defend.
The model writes the words around the maths. It never does the maths. That boundary is what makes the whole thing defensible.
The regulatory dimension
The boundary is not only good engineering; it is increasingly good compliance. The EU AI Act, which entered into force in 2024 and applies in phases, takes a risk-based approach and imposes the heaviest obligations on high-risk systems and on opaque, fully automated decision-making. A transparent, reproducible calculation engine, with a human who reviews and approves before anything reaches a client, is far easier to explain and defend than a black box that produces an answer no one can reconstruct.
It aligns with MiFID II as well. Suitability rests on being able to show that advice fits the client. Reproducible figures and a complete record turn suitability from something a firm asserts into something it can evidence. The AI, kept to language, becomes decision-support; it is never the automated decision-maker, which is precisely the posture regulators are most comfortable with.
Human-in-the-loop, by architecture
The final piece is to make the human gate structural rather than aspirational. It is not enough to say an adviser reviews the output. The workflow itself should refuse to proceed until a person has approved the extracted policy. When the review gate is enforced by the system, not merely encouraged by the interface, the adviser's judgement is guaranteed to sit between the model and the client, every time, without depending on anyone remembering to look.
What the boundary looks like in practice
In a well-designed system the separation is visible at every step. When a family's answers are turned into a draft policy, the model's contribution is the structure and the wording; the constraints it proposes are rails, not allocations, and they are presented for a human to confirm. When scenarios are produced, the figures come from the engine and the model is given no opportunity to change them. When a report is written, the prose is generated around numbers the model may not alter, and a check re-verifies those numbers against the engine before the document is finalised.
Each of these is a small act of discipline, and together they add up to a system whose outputs can be trusted precisely because the machine was never allowed to improvise where it mattered. The model's inventiveness is confined to language, where it is welcome, and excluded from arithmetic, where it is not.
The cost of getting the boundary wrong is not hypothetical. A firm that lets a language model estimate a return, or quietly adjust a figure to fit a narrative, has not gained efficiency; it has manufactured a liability it cannot see. The error will be fluent, confident and wrong, and it may not surface until a client or a regulator asks the one question the firm cannot answer: where did this number come from? A clear boundary is the only honest answer to that question.
Drawn this way, AI stops being a threat to the adviser's authority and becomes an amplifier of it. The model does the reading and the writing that used to consume hours. The engine does the maths, reproducibly. And the adviser does the one thing neither can: exercise judgement, and stand behind it. The boundary is what lets all three play to their strengths.
For a family weighing whether to trust an AI-assisted process, the question to ask is therefore not how clever the model is, but where it is allowed to act. A firm that can draw the line clearly, language here, numbers there, judgement above both, is one that has understood the technology rather than merely adopted it. That clarity is worth more than any feature, because it is the thing that will still be reassuring on the day something goes wrong somewhere in the industry and every client starts asking harder questions.
Sources: EU AI Act (2024); MiFID II suitability requirements. This article is general information, not legal advice.