The agent reads the portfolio, not the client's silence

In a private bank off St James’s, the alert lands on the banker’s screen at nine in the morning: next-best-action, a suggested rebalancing, a cross-sell window on a client who has not returned a call in five weeks. The system does not know the client has gone quiet, because silence is not something the system collects; it knows only that the portfolio has moved and that an optimal move exists against that movement. The banker, who does know, dismisses the alert and pours a coffee. In that small gesture, which no metric records and which will surface nowhere at quarter-end, sits everything the bank is automating without noticing it is the wrong thing.

Over the past eighteen months the word agent has entered every town hall in the industry, and it has entered wearing a reassuring promise: agentic AI will do the research, prepare the documentation, monitor compliance in real time, handing the banker back the cognitive time to spend on what actually matters, which is the relationship. A recent survey puts the share of banking firms already running models of this kind on the adviser’s desktop at around seventy per cent, and the figure will only climb; the direction is not in question. What goes unsaid in the town halls, partly because whoever is presenting the slides could rarely answer it, is an elementary question: what, exactly, does the agent see? And, more to the point, what does it not see?

The agent sees what is already structured as data, so it sees the asset allocation that follows from the benchmark, the position that drifts from the model, the market event that opens a window for action, the transaction history from which a propensity can be inferred. It is superb at all of this, faster and less distractible than any human being, and on that ground automation is simply the right thing to do; the trouble is that the value of a relationship with an ultra-high-net-worth client does not live on that ground. It lives elsewhere, in a region the system does not collect because it does not even know it exists: the weak signal, the sentence left unfinished over dinner, the wife who starts attending meetings she used to skip, the thirty-eight-year-old son who asks a technical question he could not have framed a year earlier. These are the cues that tell the seasoned banker a succession is stirring inside that wealth, months or years before it becomes a mandate, and not one of them is a data point.

I have watched more than one significant relationship close not through a decision but through a sequence of non-decisions, each invisible taken on its own. The client does not fall out with the bank, does not send a letter, does not slam a door; he simply stops replying, by degrees, moves a first tranche elsewhere on a plausible technical pretext, lets the relationship empty out by evaporation. From the outside, and above all from a dashboard, everything looks under control until the day it is not. The banker who had known that client for fifteen years could hear the withdrawal coming in the inflection of a voice long before the portfolio gave its first numerical signal, and it is precisely that sensibility, accumulated over years of proximity that cannot be delegated, which no agent can inherit.

This is the knot the industry would rather not look at: the knowledge that holds a complex wealth relationship together is local and tacit, distributed in the head of whoever spends time with that client, made of context that will not compress into a CRM field. It is not information a central system can aggregate and redistribute as a recommendation, because the moment you compress it into a data point you have already lost what made it valuable. The agent optimises the metric in front of it, the next action, the quarter’s engagement, the probability of conversion; it does not optimise, and cannot, the outcome of a relationship across the twenty years and two generational handovers on which that relationship is actually judged. Put another way, the agent has no personal stake in the long-run result: it performs the function it was given, and the function it was given is the one that is easy to measure.

The risk, then, is not that the agent gets the recommendations wrong; on the codifiable things it will be better than the median banker. The risk is second-order, and more insidious: the banker who hands the agent the readable layer of the work slowly stops training the muscle needed to read the unreadable one. He gets used to reacting to the alert rather than pre-empting it, loses the habit of the call placed for no apparent reason which was exactly how he picked up the unspoken, and when the event the model did not foresee arrives, because by definition models foresee the foreseen, he finds himself without the organ he needed to notice it in time. The industry is building advisers who are more efficient at everything that is not their differentiator, and, by the same move, blinder precisely where that differentiator lived.

And it is happening at the worst possible moment. The figures describe a wealth management business in which assets under management grow while revenues struggle, a growth that has become steadily more volume-driven and less value-driven, with the capacity to monetise those assets compressing quarter after quarter under competitive and regulatory pressure. Translated: the margin is defended less and less by the product and more and more by the quality of the relationship, which is to say by exactly the part of the craft that automation, applied blindly, tends to erode. There is a contradiction here that no technology roadmap in the sector seems to me to have brought into focus, because to admit it would mean slowing the very initiative being sold to the board as the priority.

None of this is an argument against agents, which we will all use, and soon, and which across a vast share of the work will render a real service. It is an argument against confusing the two layers. The bank that automates the research and the documentation, genuinely freeing time, and then spends that time sending the banker to the client more often rather than less, will have made the right move. The bank that automates the readable layer and then cuts, because the short-term numbers allow it, the hours of proximity that produced the tacit knowledge, will have optimised the quarter and hollowed out the decade.

That banker who at nine o’clock dismisses the alert and pours a coffee is not resisting the technology out of laziness or professional pride; he is doing, in that moment, the one part of the job that still justifies the margin the bank assigns him: he keeps to himself a piece of information the system does not have and will not have, and decides on the basis of it. The question the leadership should be asking is not how many agents to deploy by year-end, which is the comfortable question because it has a numerical answer. It is who, ten years from now and inside the next succession, will still be able to read the silence that no system collects.