Attention Is the New Fiduciary Asset

Financial institutions have always competed for attention.

They send alerts, frame choices, recommend products, and decide which number appears first on a screen. Digital channels made those decisions measurable. Artificial intelligence makes them personal.

A system can learn when a customer is anxious, which wording produces action, what sequence reduces resistance, and which moment is most likely to result in a trade, a loan, or another deposit. It can generate and test variations for one person rather than for a segment.

This will be described as personalization. Sometimes it will be.

Sometimes it will be persuasion at a scale and precision the customer cannot see.

As financial AI becomes better at directing behavior, institutions will need to treat customer attention as an asset held in trust—not as inventory available for extraction.

I use “fiduciary asset” here as a design standard, not as a claim that every digital interaction creates a new legal fiduciary duty. The principle is simpler: when an institution understands how to influence a person's financial behavior, it inherits a responsibility not to use that knowledge against the person's objectives.

AI turns attention into a financial input

Traditional personalization used broad categories. A bank knew a customer had a mortgage, belonged to an age range, or had reached a balance threshold. The institution selected an offer from a limited set.

AI can work at a different resolution.

It can combine transaction patterns, browsing behavior, language, response history, channel preference, and recent life events. Generative systems can then adapt the content, tone, timing, and order of an interaction. The experience can change while the customer is using it.

This matters because attention is upstream of almost every financial decision. Before a person selects a product, changes a portfolio, increases a contribution, or takes credit, someone or something has framed what deserves notice.

The US Securities and Exchange Commission has asked what online finance platforms are optimizing when they use predictive analytics and behavioral prompts. If the objective includes the platform's revenue rather than solely the investor's benefit, a conflict can be embedded in the interaction itself.

The model does not need to give false information to create harm. It may present a true fact at the most persuasive moment, conceal an alternative through ordering, or repeatedly emphasize an action that benefits the firm.

That is why the next conduct risk is not only bad advice. It is industrialized influence.

Personalization and manipulation have different objectives

The boundary is not whether AI was used. It is whose objective the system is serving and whether the person retains meaningful choice.

Personalization can reduce noise. It can explain a complex fee in familiar language, identify information relevant to a life event, or help a client focus on a long-term objective during market volatility.

Manipulation uses the same capabilities to exploit bias, vulnerability, or urgency.

The OECD defines dark commercial patterns as digital practices that subvert consumer decision-making through the way choices are presented. It notes harms including financial loss, privacy breaches, and psychological effects, and warns that machine learning may enable these tactics to be targeted at individuals.

AI makes the distinction harder to observe from outside. Two customers can see different journeys. A regulator, auditor, or product owner may inspect a compliant default experience while a live system continuously adapts the real one.

An institution therefore cannot govern this risk by reviewing screenshots. It needs to understand the objective function, training signals, experimentation logic, and behavioral outcomes.

Disclosure cannot carry the whole burden

Financial services often responds to a conflict with disclosure.

Disclosure matters. Customers should know when they are dealing with AI, what role it plays, and when commercial incentives affect a recommendation. But information does not neutralize a system engineered to overcome reflection.

A person cannot meaningfully consent to thousands of invisible experiments on timing, language, and emotional framing. Nor can a sentence at the bottom of a screen restore the options that were never made salient.

The OECD's 2026 work on AI and personal finance identifies opportunities in access and personalization alongside risks from bias, hallucination, commercial influence, privacy, and exclusion. It also notes that the effect on long-term financial well-being remains uncertain.

That last point should change product measurement. A journey that lifts conversion this week may weaken the customer's resilience, trust, or outcomes over years.

This is particularly important in wealth management. As I have written, a plausible recommendation can be more dangerous than an obvious hallucination. Persuasive delivery can make that recommendation harder to question.

What an attention duty would require

Treating attention as a protected asset can be made operational. I would begin with six rules.

1. Optimize for the customer's recorded objective

Every adaptive experience should have an explicit outcome hierarchy.

If the customer is trying to build an emergency reserve, the system should not optimize primarily for product uptake. If the client has a long horizon and a stable plan, engagement should not be increased by manufacturing urgency around normal volatility.

Commercial goals do not disappear. They must be visible and subordinate where duties or commitments require it.

2. Make friction symmetrical

Many interfaces make buying easy and reconsidering difficult. AI can intensify that asymmetry.

If a product can be accepted in one conversational flow, its material risks, alternatives, cancellation path, and cooling-off conditions should be equally accessible. A system should not personalize away the friction that protects a customer while preserving every obstacle that protects the firm.

3. Separate education, advice, and promotion

A customer should be able to tell whether the system is explaining, recommending, or selling.

The language model may be the same, but the mandate and controls are not. Each mode needs its own approved sources, permitted claims, conflict rules, and escalation path. Blurring them creates an interface that feels like trusted advice while behaving like optimized distribution.

That is one reason an AI agent needs a real job description, including what it may not do.

4. Protect vulnerability signals

AI may infer financial stress, cognitive difficulty, bereavement, compulsive behavior, or unusual urgency. Those signals can help an institution offer support.

They can also become a list of exploitable moments.

Policy should prohibit using vulnerability indicators to increase pressure, price discrimination, or cross-selling. Access to the signals should be restricted, their use logged, and their protective purpose tested.

5. Preserve quiet

Not every insight deserves a notification.

The system should have budgets for frequency, repetition, and urgency. It should distinguish a material alert from an opportunity to generate activity. Customers should be able to set preferences that the optimization system cannot silently route around.

Sometimes the most respectful recommendation is no message at all.

6. Give people a way to contest the frame

Customers need more than the ability to reject an offer. They should be able to ask why it appeared, correct the assumptions behind it, request another option, and reach a person when the interaction is consequential.

The UK's FCA Consumer Duty requires firms to act in good faith, avoid foreseeable harm, and support customers in pursuing financial objectives. Those principles provide a useful test for adaptive interfaces: does the system help the customer decide, or merely help the institution convert?

Measure agency, not only engagement

Most digital teams can report clicks, completion, frequency, and revenue. An attention duty requires a second dashboard.

Measure rapid reversals, cancellations after high-pressure journeys, complaints, repeated prompts before acceptance, and the difference between immediate conversion and sustained outcomes. Track whether customers understand material terms, whether they can find alternatives, and whether vulnerable groups experience different pressure.

The G20/OECD principles for financial consumer protection place financial well-being and resilience alongside digitalization as cross-cutting themes. The CFPB's Financial Well-Being Scale offers another useful idea: financial success includes security and freedom of choice, not only balances or transactions.

That suggests a better north star. Do customers have more control after using the system?

The brand advantage will belong to institutions that restrain themselves

AI influence will become cheaper and more effective. Competitors will use it. Quarterly metrics will reward it. The temptation to optimize every surface will be strong.

But a financial institution is not an entertainment feed. The customer cannot simply recover from every bad decision by closing the app. Credit, pensions, protection, and investment choices can shape a decade.

There is a spiritual dimension to attention because attention determines what enters a person's field of choice. To respect attention is to recognize that a customer is more than a predictable collection of responses.

The firms that understand this will not abandon personalization. They will use it with restraint. They will make the helpful option clearer without making the profitable option irresistible. They will design moments of reflection into decisions that deserve them, just as intelligent financial AI should know when to wait.

Trust will not come from promising that the model is aligned. It will come from evidence that the institution declined to exploit what the model knew.

In an age of infinite personalized persuasion, protecting a customer's attention may become one of the clearest demonstrations of fiduciary character.