Why Financial Advisors Should Use AI | Alyra

The internal return is already proven. The client-facing return is larger and needs one thing most advisory firms have never built.

Around forty percent of advisers already use AI internally. Four percent have it anywhere near a client. The first number is the case for starting. The second is where the value still sits.

The 2025 Investment Management Compliance Testing survey, which polled 577 advisory firms, put internal-only AI use among investment advisers at forty percent, and found four percent had introduced AI for even straightforward client-facing interactions. Those are survey figures rather than regulatory filings, so treat them as directionally right rather than precise. The shape they describe is unmistakable either way.

The usual explanation is that client-facing AI is harder to build. It is not, particularly. Drafting a client email is a simpler engineering problem than reconciling a portfolio across four custodians. The reason one happens and the other does not is that they carry different costs when they are wrong.

Start where the return is already proven

The internal uses are unglamorous and they work. Reconciling a portfolio across four custodians. Turning a recorded meeting into notes a paraplanner can act on. Pulling the three relevant paragraphs out of a two hundred page document. Drafting the first version of a review pack so an adviser edits rather than starts.

None of that touches a client directly, which is exactly why it is the right place to begin. The work is real, the hours are real, and when the output is wrong the person who asked for it notices in the next minute.

Two different jobs wearing the same name

Internal use has one requirement. The output has to be useful enough that the person using it keeps using it. When it is wrong, the person who asked notices, corrects it, and moves on. The loop is short and the cost of an error is a few minutes.

Client-facing use has a second requirement that nobody writes down. You have to be able to stand behind the output after it has left the building. Not privately believe it was right. Be able to show why it was right, to a client, to a compliance officer, and possibly to an examiner, months later.

Those are different products. Most firms built the first one and then discovered, at the moment of sending, that they had never built the second.

What standing behind it actually takes

Three things, and none of them is model choice. The output has to name its source, so that a claim about a client's drift can be traced to a statement, a date and a line. The system has to behave the same way next quarter, which generative systems do not do by default, so you need a way to detect that quality has moved rather than assuming it has not. And a named person has to own the verdict. Not the tool, and not the vendor.

Firms did not stall on client-facing AI because the technology was not ready. They stalled because sending it meant owning it.

What it looked like when we built it

On a system we built over a large corpus of regulated documents, the standing requirement is that ninety-eight percent of answers name the document they came from. The number matters less than where the check lives. It runs automatically every time the system changes, and a change that makes answers harder to trace does not ship.

That was never a compliance feature bolted on at the end. It is what made the output usable at all, because the people relying on it had to defend their conclusions to an auditor who had no reason to be generous.

What changes when you build it in that order

Start internally and you get the citation habit, the evaluation harness and the named reviewer working on something that stays inside the building, where being wrong costs minutes. Those three things are the whole gate. Once they exist, client-facing work stops being a leap and becomes the next step.

Firms that do this in the wrong order pick the client-facing use case first because the return is obvious, meet the verification problem at the worst possible moment, and conclude that AI is not ready for their business. The technology was ready. The scaffolding underneath it was never built.

The four percent figure is not a measure of how good the tools have become. It is a measure of how many firms have built something they are willing to put their name on. That is the opportunity, and it is still open.

Start a conversation