The AI stack is a correlation trade, and fees haven't moved

Sixty-three percent of the 131 asset managers in Mercer's 2026 survey reported using off-the-shelf AI tools, and that is the figure allocators should underwrite — ahead of the survey's louder numbers, in which 55 percent said they had integrated AI into at least one investment process and 91 percent expected to lean on it more heavily over the next year. Adoption at that breadth is a common input dressed as a capability. What an industry buys, and from how few suppliers, tells you how much of what it sells is actually its own.

An opinion piece published by WealthManagement.com on Sept. 11 argues the problem from the other end, and it deserves engagement rather than a shrug. By his own account, the author spent nearly a decade as a top-quartile CIO and portfolio manager without attending the undergraduate institutions or the MBA programs that supply most of the industry's investment ranks, and without ever sitting the CFA exams he paid his whole team to take. His thesis is that asset management has built an extraordinarily effective engine for agreement, beginning with the talent pipeline — a concentrated set of schools, a narrower set of graduate programs, one globally recognized CFA curriculum teaching thousands of professionals to analyze markets, value securities, and think about risk — and he grants those shared standards their value; the trouble begins when the machinery of common training starts returning common answers, managers who describe their strategies differently and end up holding similar securities, an industry efficient at spotting what already works and less reliably rewarded for seeing what hasn't started working yet.

The AI argument is the part that should travel: thousands of investment professionals, he writes, are pointing overlapping models trained on overlapping information at similar questions, and the answers should converge further. He offers that as a prediction rather than a measurement — his phrasing is that we should not be surprised if they do — but the survey's arithmetic sits underneath and runs the same direction, with 63 percent buying the same tools and 58 percent drawing on vendor-supplied data meaning the inputs are shared long before anyone can inspect the outputs.

For the allocators who hire managers — family offices, endowments, foundations, and the gatekeepers inside wealth platforms — this lands as a repricing problem, because the durable differentiation in a portfolio now sits in the parts of the process that cannot be licensed: data a firm owns instead of buys, a mandate that permits a position the consensus finds uncomfortable, an analyst given room to be early and wrong for a couple of quarters. The author offers his own résumé as the exhibit, crediting process and team for his performance but also a vantage point the shared curriculum never produced, and the point generalizes whether or not one accepts the self-assessment.

None of this argues for avoiding managers who use AI, which would be both futile and at odds with the survey; it argues for pricing them the way an allocator prices any shared input — as a table stake whose cost belongs in the fee negotiation rather than in the return assumption.

The portfolio-level version of the same problem is less comfortable to sit with because diversification in a multi-manager line-up is usually counted rather than tested: eight or ten managers with different names, different mandates, and different fee schedules read as eight or ten independent bets, but if the tools converge and the underlying data overlaps, the count stays the same while the bets move closer together. That argues for measuring overlap and correlation directly instead of trusting the org chart; the survey describes tooling and data use rather than returns, so this is inference, but it is the kind of inference an allocator is paid to make early, and it is cheaper to test a manager's overlap before the next quarterly letter than after it.

As this publication has argued, AI has already arrived in the fee-setting hour on the advice side of the business, where software is pressing on what a planning relationship can charge, and the reckoning is due one level up in the management fee, where a schedule built on the scarcity of a manager's judgment gets harder to defend when 63 percent of the managers in the survey are running vendor tooling. Pricing is the last line item in this business to reflect a change in the product, and it happens to be the one allocators control.

The RIA channel already ran a version of this experiment with capital instead of code: Fidelity's midyear data, which PWD covered in August, put private equity behind 89 percent of deals, the median target at $630 million, and a buyer pool getting shorter even as acquired assets nearly doubled. Convergence on one template produced bigger transactions and fewer distinct participants, and if manager tooling and datasets converge the same way, the likely result is fewer portfolios that differ in any way a client can see — more distinct branding in front of similar holdings.

The question for the next manager meeting has little to do with whether the firm uses AI, since 91 percent of the surveyed industry intends to use more of it and the answer is therefore yes; the question is what the model sees that no competitor can buy: a dataset, an origination channel, a counterparty relationship, a mandate with room to hold something unpopular for eighteen months. When Mercer runs the survey again, adoption will be near universal and therefore uninformative, and the number worth watching will be the share of managers who can name an input that isn't on a shelf.