What an Advisor Actually Does, and Why Advice Is the Regulated Part

An advisory practice looks like one activity from the outside and is five inside. Separating them is the whole exercise, because they compress at completely different rates.

Gathering. Understanding a client's position: assets, liabilities, income, obligations, timeline, dependants, and the things they say sideways about what they actually want. Slow, relationship-dependent, and partly emotional labour.

Analysis. Modelling scenarios, testing whether a plan survives a market fall or a longer retirement, evaluating a portfolio against an objective. Structured, quantitative, and already heavily software-assisted.

Recommendation. Deciding what this client should do. This is the regulated act, and it is the subject of the next step.

Communication. Explaining the recommendation so the client understands it well enough to consent to it, then reporting on it periodically for years.

Administration. Meeting notes, suitability documentation, review packs, client correspondence, compliance records. Large, unloved, and the single biggest consumer of an advisor's non-client hours.

The shape of the opportunity follows immediately. Administration compresses heavily. Communication compresses substantially with review. Analysis was already partly automated and gains at the margin. Gathering compresses barely. And recommendation does not compress at all, because a regulator has assigned responsibility for it to a named person.

Most professions have a quality expectation attached to their advice. This one has a legal standard, and it is attached to a person.

In the United States, a registered investment adviser owes a fiduciary duty under the Investment Advisers Act of 1940, comprising a duty of care and a duty of loyalty, and the Securities and Exchange Commission has been explicit that it cannot be waived. A broker-dealer making a recommendation to a retail customer is subject to Regulation Best Interest, effective since June 2020, which requires acting in the customer's best interest without placing the firm's interest ahead of it.

In the European Union, MiFID II requires that a firm providing investment advice obtain the information necessary to assess suitability, and it may not recommend where it does not have that information.

What these have in common matters more than their differences. Each requires that a specific, identified party has assessed this client's circumstances and concluded that this recommendation is right for them. The duty attaches to a regulated person or firm, and there is no provision anywhere for delegating it to a system.

So a model can produce an analysis, draft the explanation, and prepare the documentation. What it cannot do is be the party that owes the duty.

This is not a limitation of current capability that better models will remove. It is a structural feature of how the regulation is built, and it is worth understanding as such rather than as a temporary obstacle.

The second structural feature, and the one advisors most often forget when adopting a new tool, is that this industry records its communications by rule.

In the United States, Advisers Act Rule 204-2 and Exchange Act Rule 17a-4 require firms to make and preserve extensive books and records, including communications relating to recommendations and advice, for defined periods. The SEC and FINRA have pursued substantial enforcement against firms over off-channel communications, where employees conducted business on messaging apps that were not captured.

The implication for AI adoption is direct and frequently missed. If an advisor uses a chat tool to work through a client's situation, that exchange may be a business communication relating to advice, sitting on a platform the firm does not capture, retain or supervise.

And FINRA has been explicit that the existing framework applies. Regulatory Notice 24-09, issued 27 June 2024, reminded member firms that FINRA rules are technology-neutral and continue to apply to generative AI tools, including the Rule 3110 supervision obligation, whether the firm builds the tool or uses a third party's, including AI features embedded in products the firm already uses.

That last clause is the one to sit with. A capability that appears inside an existing vendor product is still within scope, and firms have discovered AI features already switched on in tools they procured years earlier.

So the first governance question in this profession is not which model. It is where does the conversation live and is it captured.

Mapping the advisory process against what may and may not be delegated.

Gathering client information. A model can prepare the questions and structure the record. The conversation itself stays with the advisor, because much of what matters is what the client does not say directly.

Analysis and scenario modelling. Substantially delegable, since this is computation against stated assumptions, and the advisor's job is to own the assumptions rather than the arithmetic.

The recommendation. Not delegable. A named regulated person owes the duty of care and loyalty, or the best interest obligation, or the MiFID suitability assessment. There is no mechanism for a system to hold it.

Explanation and documentation. Delegable in draft, adopted by the advisor. The advisor is asserting whatever goes out.

Administration and records. Delegable, subject to the capture and supervision requirement, because the tool itself sits inside the regulated environment.

The line running through the middle is not about difficulty. Analysis is harder than recommendation in a computational sense. The line is about who is answerable.

If there is a single number worth knowing about this profession, it is how much of an advisor's week is not spent with clients.

Industry surveys consistently place client-facing time as a minority of the working week for most advisors, with the balance going to preparation, documentation, compliance and practice management. The specific figures vary by survey and by how the categories are drawn, so they should be treated as indicative rather than precise, but the direction is not in dispute among practitioners.

What fills that non-client time is largely writing things down. The suitability file. The meeting note. The annual review pack. The follow-up letter confirming what was discussed and agreed. The client correspondence that arrives between reviews.

That body of work has three properties that make it the ideal target. It is high volume. It is derived from information the advisor already has rather than requiring new judgement. And the quality bar is clarity and completeness rather than insight.

Which is exactly the profile of work that generation handles well.

The framing that follows is worth carrying through the rest of this cursus. The goal is not an advisor who advises faster. It is an advisor who spends a larger share of a fixed week in front of clients, because the documentation that used to consume the evening now takes a fraction of the time.

That is a less exciting claim than replacing the advice, and it is the one that survives contact with the regulation.

There is a failure mode specific to quantitative domains that deserves its own treatment, because it does not look like a failure.

A language model asked to compute a drawdown, project a portfolio value, or calculate a tax position will produce a number. The number will be formatted correctly, presented with appropriate confidence, and sit inside otherwise accurate prose. And it may be wrong in a way that no amount of reading will reveal, because there is nothing on the surface to distinguish a correct calculation from an incorrect one.

Compare this to a hallucinated fact, which a knowledgeable reader often catches. A wrong number in a projection is invisible unless independently computed.

The rule that follows is worth stating absolutely. Do not use a language model as a calculator for anything a client will see or rely on. Compute in the tool built for it: the planning software, the spreadsheet, the portfolio system. Then use the model to explain the result.

This division holds up well in practice because it plays to actual strengths. The planning software is excellent at Monte Carlo simulation and terrible at explaining the output to a nervous sixty-two-year-old. The model is the reverse.

A related caution on citations. Models will produce fund names, expense ratios, historical returns and regulatory citations that look right and are not. Anything factual about a specific product or rule gets verified against the source document, every time, without exception.

It is worth knowing the current state of AI-specific rulemaking for this sector, because the assumption that a comprehensive rule is imminent shapes how firms behave, and the assumption is currently wrong.

In August 2023 the SEC proposed rules on conflicts of interest associated with the use of predictive data analytics by broker-dealers and investment advisers. The proposal would have required firms to eliminate or neutralise conflicts arising from technologies used to interact with investors, and it was drawn broadly enough to reach a wide range of ordinary software.

It was heavily criticised in comment, and the SEC formally withdrew it on 12 June 2025, as one of fourteen proposals withdrawn together. Withdrawal means any future rulemaking on the subject starts again with a new proposal and a fresh comment period.

So the position for a firm today is that there is no AI-specific rule for investment advisers in the United States, and the applicable obligations are the ones that already existed: fiduciary duty, Regulation Best Interest, the Marketing Rule, the Compliance Rule, books and records, and supervision.

Two readings of that are available and the second is correct.

The first: nothing applies yet, so wait. The second: everything already applies, and there is no forthcoming rule that will tell you what to do, so the framework has to be built from the existing duties.

That is the same structure as the guardrails cursus on model risk in banking. The absence of a bespoke rule is not an absence of obligation.

An order that follows the regulatory weight rather than the excitement.

First, meeting documentation. Recording a client meeting with consent, transcribing it, and generating the note and the follow-up. This is the largest single time recovery available, it operates on your own record of your own conversation, and it improves the compliance position rather than threatening it, because contemporaneous notes are better evidence than reconstructed ones.

Second, meeting preparation. Assembling what you need to know before a review: what changed in the portfolio, what was agreed last time, what is outstanding, what life events were mentioned. Internal, high value, low risk.

Third, client communication drafting. Review letters, explanations of a recommendation you have already made, responses to questions. Reviewed and adopted before sending, and subject to the Marketing Rule where it becomes an advertisement, which lesson two covers.

Fourth, research digestion. Summarising fund documentation, market commentary and regulatory updates for your own use, with anything load-bearing verified at source.

And not at all without a governance answer first: anything that reaches a client without a person in between, and anything where the tool sits outside your firm's capture and supervision.

The ordering principle is the one from the first step. Administration first, because that is where the hours are and where the risk is lowest. The recommendation is not on the list at all, because it is not available.