Most firms are asking too little of their AI tools

Last week we wrote about the person in the loop, and whether someone checking an AI tool’s output really amounts to having the tool under control. That was a close-up question about a single letter and a single approval. This post steps back to a wider one, which is whether firms are asking enough of these tools in the first place.

In most of the wealth managers and family offices we work with, AI is used to tidy an email, summarise a meeting or produce a first draft of a review letter. That is useful, and it saves time. It is also a small fraction of what the tools can now do, and we think the larger risk for a firm of ten or twenty people is not that its AI does too much but that it is never asked to do very much at all.

Where the idea comes from

Felix Rieseberg, who leads engineering on several of Anthropic’s Claude products, published a piece this week called How to be more ambitious: a tutorial for humans with models. It is written for a general audience, and his examples come from software and corporate reporting, but the framework travels well.

He describes four kinds of ambition. The first is speed, which means doing the same work faster, and it is where almost everyone already is. The other three are less familiar, and each of them has a fairly direct equivalent in a wealth management firm.

Three ways to ask for more

Ask for the outcome, not the task. Rieseberg calls this altitude. He describes moving from asking a model to fix individual faults in his software to asking it to keep the software free of faults altogether. In a wealth firm, the equivalent is moving from “draft this client’s review letter” to “make sure no client goes into a review with a risk profile, a contact record or a set of objectives that no longer reflects their life”. The first is a task that ends when the letter is written. The second is the thing the letter was always meant to achieve.

Ask for the work an expert would do, not a better-sounding version of yours. He calls this depth. A model asked to act like a specialist will often adopt the vocabulary without changing its method, so the useful instruction is to have it work from the raw material a specialist would start from. Most firms carry beliefs about their own business that have never been checked, such as why clients leave, which introducers bring the families who stay, or where the time goes in onboarding. A tool given the underlying records and asked to test one of those beliefs will sometimes find that it does not hold.

Ask what other disciplines would see. This is his idea of range. One technique he borrows is to describe the work as a verb rather than a noun. A quarterly client report is not really “a pack”. It is helping a family understand what happened to their money, and feel settled about what happens next. Described that way, it becomes obvious that a good editor, a behavioural specialist and the family’s own lawyer would each look at it differently. He is careful to say that the point is not to have the model pretend to be these people, but to have it do what they actually do.

Why this matters more to a small firm than a large one

A firm of twelve people cannot employ an analyst to test its assumptions, an editor to rework its client reporting and a sceptic to pick holes in its investment notes. Larger institutions can, and part of what they buy with that headcount is the ability to look at their own work from several directions. For a smaller firm, range of this kind was simply out of reach.

That is the part of Rieseberg’s argument we find most relevant. He closes by writing about what psychologists call self-efficacy, the belief that you can do something, and about how often people are held back less by ability than by an assumption about what is possible for someone in their position. His example is Michelangelo, who resisted painting the Sistine Chapel because he thought of himself as a sculptor. Small firms tend to make a similar assumption about themselves, which is that rigorous analysis of their own business is something banks do.

How this fits with having a person in the loop

There is an obvious tension between this and what we wrote last week. Rieseberg notes that working at a higher level is often where people find ways to take the human out of the loop altogether, and that was exactly the situation our last post was cautious about.

We do not think the two arguments pull against each other. The four things we said a reviewer needs, which are seeing how an answer was reached, being able to stop it, having someone check that reviews are working and bringing knowledge the model lacks, are what make it safe to ask for more. A firm without them is right to keep its AI on small tasks. A firm with them can afford to be considerably more ambitious, because it will notice when something goes wrong.

Rieseberg makes a related point of his own. He warns against letting a model judge its own work and argues for an outside check wherever possible. That is the same principle as an independent review in any regulated firm, applied to a new kind of colleague.

There is also a practical condition that sits underneath all of this. Asking a tool for an outcome usually means giving it access to the records that outcome depends on, and Rieseberg describes having to rework his own setup before his model could do what he asked. For a wealth firm, that is a question of where client data sits, who and what can reach it, and whether the answer satisfies the residency and confidentiality obligations the firm already has. Those questions need answering before a tool is given more to do, not after.

A question to take into the week

Pick one thing your firm already uses AI for, and ask what that task is ultimately for. If the honest answer is something larger than the task itself, such as clients who are well understood, reviews that catch changes early, or onboarding that does not lose a family’s goodwill in the first month, it is worth asking what it would take for the tool to work towards that instead.

The answer may well be that the firm is not ready yet, because the data is in the wrong place or nobody is in a position to check the results. That is useful to know, and it is a more productive conclusion than assuming these tools are only good for tidying emails.