What Agentic AI in Wealth Management Actually Looks Like When You're the One Building It

Deloitte put a number on it this year. Agentic AI could add somewhere between $10 trillion and $35 trillion of capacity to wealth management by 2032, and lift adviser productivity by 30% to 100%. That’s the headline doing the rounds.

I’ve spent the last few weeks scoping a real build for a private wealth firm. So here’s the same story from the other side of the desk.

The $35 trillion, up close, is a portfolio manager sitting down before the market opens with about ten linked Excel workbooks and roughly 300,000 formulas. He runs the exports. He copies the numbers across. If a target weight changes at the last minute, he re-exports and re-pastes, and that costs him 30 to 45 minutes he doesn’t have.

This piece is about that gap: what making agentic AI work for one real firm actually takes.

The number everyone’s quoting

The Deloitte prediction is real and it’s worth reading. Advisers spend around 70% of their time on back-office work rather than talking to clients. Move a decent chunk of that to software and you free up 25% to 50% of their week. Do that across the industry and the capacity maths gets you to those trillion-dollar figures.

Here’s the part that gets left off the slide. The same report says the gains arrive in stages, and most firms are stuck at the first one. Only about 6% of advisory firms use agentic tools at all today. Around 5% have joined their systems together well enough to run anything across them. The value, in Deloitte’s own framing, gets capped by the weakest of three things: how advisers actually use it, whether the firm is set up to support it, and whether the technology underneath can join up the data.

So the $35 trillion is real. It’s also gated behind the least glamorous work in the building.

What it actually looks like on a Tuesday morning

When people say “back office”, this is what they mean in a wealth firm.

Positions come out of one system. Cash and account data out of another. Prices and analytics from a market-data provider. Each one exports in its own format, and someone stitches them together by hand every morning.

When an integration between two of those systems breaks, trade review drops back to being a spreadsheet job. Fees get worked out by hand. Opening a new account means re-keying the same client details into different systems, because the know-your-client checks were never unified across the jurisdictions the firm operates in. And there’s usually one expensive market-data feed that everyone quietly resents paying for.

All of that is plumbing. Unglamorous, and it’s the thing that actually decides whether an AI project works. The machinery that matters here runs underneath the adviser. It ingests the data, does the calculations the same way every time, and has the answer ready before the market opens.

The 90/10 rule that’s not in the pitch deck

Sit with the person doing the rebalancing and a pattern shows up fast. Around 90% of the sell decisions follow rules you can write down. If an account is heading into debit, here’s the ranked order you sell in. Here are the tax rules. Here are the concentration limits. That 90% is codifiable, and it’s exactly the kind of work AI handles well.

The other 10% is the portfolio manager’s judgement. A feel for a particular client, or a reason not to sell the obvious thing. You don’t automate that. You build the system so that override stays easy, and so that every time someone uses it, it’s logged.

This is the Klarna lesson applied to finance. Klarna replaced 700 support agents with AI, made a lot of noise about it, then quietly hired humans back. The AI did the volume. It couldn’t do the last mile. In wealth management the last mile is that 10%, and pretending the software eats all of it is how you lose the trust of the people who have to sign the trades off.

Why the compliance bit is the actual product

A wealth firm usually answers to more than one regulator. That single fact reshapes the whole build.

Every weight change, every override, every generated trade ticket has to be logged, attributable, and kept. The data has to stay inside the firm’s own environment, not wander off to some third party’s servers. Anything the system does that carries consequences has a human signing it off, with the trail to prove it.

Deloitte names these barriers too, and none of them is about a cleverer model. Fragmented systems. Unclear rights over who’s allowed to touch which data. Weak testing and compliance setups. That’s where most of these projects quietly die.

For a regulated firm, the audit log is the reason the thing is allowed to exist at all. You build it first, not last.

Why the pilots that skip discovery fail

You’ve probably seen the other number doing the rounds: roughly 95% of corporate AI pilots don’t deliver a return. Wealth management won’t be the exception, and the reason is the one Deloitte keeps pointing at. Firms buy the tool before they’ve understood the work — the same failure pattern I’ve written about in Plug-and-Play AI Is a Myth.

You can’t order this off a shelf. The build starts by sitting with the person who runs the morning ritual and writing down what actually happens. Which exports. Which formulas. Which decisions are rules and which are judgement. Then you turn 300,000 spreadsheet formulas into a few hundred deterministic functions you can test, and you draw a hard line between what ships first and what waits.

That scoping work is unglamorous, and it’s the whole game. It’s the difference between the 5% who get near the capacity Deloitte is promising and the 95% who buy a licence and go back to Excel.

If this is your Tuesday morning

If you run a wealth firm and any of this sounds like your operation, start small. Spend a couple of weeks mapping what actually happens on the desk before anyone talks about platforms. Get honest about what’s a rule and what’s judgement.

That’s the work we do at Exponential Partners. If you’re not sure whether your setup is ready for it, message me and I’ll tell you straight.

References

1. Deloitte Center for Financial Services — “The agentic AI productivity wave is heading for wealth management” (2026 Financial Services Industry Predictions). deloitte.com
2. MIT — The State of AI in Business: ~95% of corporate AI pilots fail to deliver measurable ROI, 2025.
3. Klarna AI customer-service reversal, widely reported 2024–2025.