Why Asset Managers Can’t Measure Their AI ROI

Mercer surveyed 131 asset managers globally in February 2026 and found that 55% have integrated AI into at least one of their strategy’s investment processes. Another 27% are running pilots or proof-of-concept work. Only 18% report no AI integration at all, and 91% say they plan to expand their use of the technology within the next twelve months.

Those are the numbers that made the headlines. But they’re not the interesting numbers.

The interesting number is 5%. That is the share of managers who have given AI autonomous or semi-autonomous authority over an investment decision. Everything else sits upstream of the decision: 73% use AI for operational efficiency and automating routine work, 68% treat it as an analytical partner that surfaces insights for a human to act on. Beverley Sharp, Mercer’s Global Manager Research Leader, summed it up in a line that deserves more attention than it got: “AI is delivering measurable efficiency and insight for asset managers today, but the technology is largely a partner rather than a decision-maker.”

The gap between the 55% headline and the 5% reality is where due diligence should be concentrating. It is also where most of the marketing is happening. We have made a version of this argument before in a different context, which is that a basket of AI recommendations is not advice until somebody with a fiduciary obligation signs their name to it. The same logic applies to an investment process. A model that produces a ranked list is not managing money. A person who acts on that list is.

Efficiency Is Real. Alpha Is Not Showing Up Yet.

Look at what managers say they are getting. Enhanced operational efficiency: 69%. Faster or higher-quality insights: 55%. Measurable improvement in investment returns: 8%. Reduction in portfolio volatility: 8%.

Read that last pair again. After two years of the most intense technology adoption cycle the industry has seen since electronic trading, fewer than one in ten managers can point to a return or risk outcome they attribute to AI.

This is not a problem unique to asset management. KPMG’s Global AI Pulse survey of more than 1,400 leaders across industries found 76% saying AI already delivers meaningful business value, up 12 points in a single quarter, while just 7% could point to established ROI. That is the same shape as the Mercer split, which suggests the gap is not something peculiar to investment processes. It is what happens anywhere adoption runs ahead of measurement, and it is the reason the burden of proof belongs on the manager making the claim.

That is not necessarily damning. Efficiency gains are real economic value, and a research analyst who can process forty earnings transcripts instead of eight has genuinely expanded their coverage universe. We have seen the same pattern on the wealth management side, where AI did not break anyone’s operations so much as it made the existing cracks impossible to ignore. Speed exposes process debt before it creates alpha. But it does mean that when a manager tells an asset owner that AI is “driving our investment process,” the honest translation in most cases is “AI is helping our analysts read faster.” Those are different claims with different implications for fee justification, capacity, key-person risk and repeatability.

The 8% figure also raises an attribution question that very few managers can answer cleanly. If AI improved your returns, show the counterfactual. What did the process produce before the model was introduced, and what is the mechanism by which the model changed the output? Firms that cannot answer that are not necessarily lying. They just have not built the measurement infrastructure to know whether the thing they bought is working.

What Separates Real Decision Support From a Slide

Ezra Group works with firms trying to move AI from pilot programs into production use inside the investment process, and the pattern that separates the two is consistent enough to be diagnostic.

• Production AI has a named owner and a named input. In a real implementation, the model output is a specific, versioned input into a documented process. It has a name. It appears in an investment committee memo or a portfolio construction workflow the same way a factor score or a broker estimate does. Pilot theater looks different: several analysts each have a subscription to a chat tool, they use it in ways nobody has documented, and the firm counts that as AI adoption because seat licenses are easy to count.
• Production AI has kill criteria. Every real deployment we have seen started with a stated hypothesis and a threshold at which the thing gets turned off. Mercer’s own framework makes the same point, recommending that asset owners assess whether a manager’s use of AI is purposeful with a clear hypothesis, credible and governed, empirically validated, delivering measurable value, and sustainable at scale. Pilots that have run for eighteen months without either graduating or dying are a governance finding, not an innovation program.
• Production AI accounts for model drift. This one gets missed constantly. If your research process depends on a third-party foundation model accessed through an API, the vendor can change the model underneath you. The version you validated in March is not the version answering questions in September. Firms with real controls pin model versions, log which version produced which output, and re-validate on upgrade. Firms without them have an investment process that silently changes whenever a vendor ships a release, and no record of when it happened.
• Production AI measures its own override rate. Human-in-the-loop is the phrase every manager uses to describe oversight, and it is almost never defined. This is the governance gap most firms are still ignoring: a policy document that says a human reviews the output, with no evidence anywhere that review is happening. There is a simple metric that makes it real: how often does the human disagree with the model, and what happens then? If the portfolio manager overrides the output 2% of the time, that is not oversight, that is a rubber stamp with a compliance story attached. If the PM overrides 80% of the time, the model is not adding anything and somebody is paying for it anyway. The useful range is in between, and a firm that has never measured it has not thought seriously about the control.

It Helps to Know What These Agents Really Do

One reason these conversations go badly is that neither side has a shared vocabulary for what an AI agent in this business does. The manager says “we deployed an agent.” The asset owner pictures something autonomous. The reality is usually a workflow with a language model in the middle of it.

We built the Ezra Group AI Agents for Advisors Directory partly to fix that. It now catalogs close to 50 agents, and the organizing decision was to list individual agents rather than vendors, because most firms in this space now ship several specialized agents that solve unrelated problems. Each listing covers the business function, the intended user, what the agent can and cannot do, and where it sits in the development cycle.

Several of the listings touch the investment side directly, including portfolio management, research summarization and investment commentary drafting. Reading through them is a fast education in exactly how narrow most current agents are, and that narrowness is the point. It is much easier to evaluate a manager’s claim about AI in portfolio construction once you have seen what the available tools genuinely do at this stage.

The Data Problem Is the Actual Constraint

Sixty-nine percent of managers in the Mercer survey identified data quality or access as a significant barrier to further AI adoption. That is the single largest obstacle in the study, ahead of regulatory and compliance concerns at 59%. Only 28% have developed AI models internally.

Anyone who has spent time inside an asset manager’s technology stack knows why. The problems are unglamorous and they are structural, and they are the same problems we keep running into on the advisory side, where the dirty secret of AI implementation is that data discipline determines the ceiling long before model selection does.

Most firms do not have one security master, they have three, and the reconciliation between front, middle and back office is a person rather than a process. Research notes live in email, PDFs and a document management system with no entity resolution back to the instrument, so a model asked about a position cannot reliably find what the firm already knows about it. Point-in-time data capture is inconsistent, which means backtests of AI-generated signals leak future information in ways that are hard to detect and flattering to the results. Licensing terms on broker research and market data frequently prohibit using the content to train or fine-tune models, and compliance often discovers this after the data science team has already done it.

Then there is the permissioning problem, which is the one that should worry asset owners most and gets discussed least. Information barriers exist for a reason. A firm that indexes its entire document corpus into a retrieval system so analysts can ask questions across it has, unless it was very careful, built a machine that will cheerfully carry material nonpublic information across a wall that took the compliance department years to construct. Document-level access controls that mirror the firm’s actual information barriers are hard to build and easy to skip during a pilot. Ask whether they were built.

None of this is a reason not to use AI. It is the reason the 5% number is 5%. You cannot responsibly hand decision authority to a model when the data underneath it is not clean, not point-in-time, not permissioned and not lineage-tracked. The managers who understand this are working on plumbing. The ones who do not are working on slides.

The Regulator Has Already Drawn the Line

The SEC settled charges against two investment advisers, Delphia and Global Predictions, for AI washing. Delphia paid $225,000 after claiming it used machine learning on collected client data that it had never used in its algorithms. Global Predictions paid $175,000 for promoting “AI-driven forecasts” and calling itself the first regulated AI financial advisor without documentation to support either claim. The standard the SEC articulated is simple: say what you are doing and do what you are saying.

The 2026 exam priorities extend this well beyond firms that market themselves as AI shops. Examiners are looking at whether marketing materials, Form ADV disclosures and client communications accurately describe the extent, nature and limitations of AI usage. They are assessing whether firms evaluate tools before deployment, monitor outputs for accuracy, maintain human oversight of material AI-driven decisions and address algorithmic bias. And they are specifically checking whether claimed AI-driven portfolio management genuinely influences decisions or plays a supplementary role.

That last item is the regulatory version of the same question this article opened with. The gap between the 55% and the 5% is now an examination topic.

Questions Asset Owners Should Be Asking

Mercer’s framework points at four core questions: what specific investment problem does AI solve, where in the workflow is it applied, how is the capability operated and governed, and what counts as outperformance. Those are the right starting points. Here is how to make them harder to deflect.

Show me where in the process the output lands. Not a diagram of the architecture. The actual artifact. The IC memo, the screen, the constraint in the optimizer. If they cannot produce the artifact, the model is not in the process.

What did this replace, and what happened to the thing it replaced? Real adoption displaces something. If nothing was retired and no headcount was reallocated, you are looking at an addition to the research budget, not a change to the investment process.

Who owns the model inventory, and can I see the count? A firm with real governance can tell you how many models are in production, who validated each one, when they were last reviewed and which ones were retired in the last year. The retirement number is the tell.

How do you version and pin third-party models? Covered above. The answer “we use the latest version” is a finding.

What is your override rate and how has it moved? Also covered above. Expect blank stares. Note who does not give you one.

How are your information barriers enforced inside the retrieval layer? Document-level permissions inherited from the source systems, or a single index everyone queries. There is no third answer.

Which of your data licenses permit model training, and who checked? If legal has not reviewed the vendor agreements for AI usage rights, this is a live exposure sitting inside the research process.

What are the kill criteria for your current pilots, and what have you killed? A firm that has never killed a pilot is not running experiments, it is running demonstrations.

If your AI lead leaves tomorrow, what breaks? The 28% who built internally have concentrated key-person risk. The 72% who did not have concentrated vendor risk. Both are answerable. Neither should be a surprise.

What would you have to see to hand a model actual decision authority? This is the best single question on the list, because the answer reveals whether the firm has a theory of the case or a technology budget. Managers who have thought hard about this describe specific validation thresholds, specific data prerequisites and a specific governance sign-off. Managers who have not will tell you about their innovation culture.

The Honest Position

The Mercer data describes an industry in a reasonable place. AI is producing genuine efficiency gains, it is expanding what research teams can cover, and the profession has been appropriately conservative about handing it the keys. Ninety-one percent plan to do more of it within a year, and they should.

The risk is not that asset managers are moving too slowly. It is that the distance between what firms are doing and what firms are saying keeps widening, and asset owners who accept adoption headlines at face value will end up paying active fees for a research process they cannot describe. The managers worth backing are the ones who will tell you plainly that AI is a partner and not a decision-maker in their shop, then show you exactly where the partner sits and what would have to change before it gets promoted.

Ask for the artifact. Ask what got retired. Ask what they killed. The answers separate the two groups quickly.

Ezra Group provides technology, operations and market research consulting to wealth management firms and wealthtech vendors, including AI governance and implementation work for firms moving from pilot to production.