Wealth management firms are committing larger budgets to artificial intelligence, but most remain unable to demonstrate what those investments are producing, exposing a widening gap between the industry’s technology ambitions and its financial controls.

Research published by wealth management consultancy F2 Strategy found that AI spending has increased rapidly over the past three years as registered investment advisers, broker-dealers, banks and trust companies move beyond isolated experiments. Yet most organizations surveyed have not established a formal process for measuring the performance of their AI initiatives, making it difficult to connect rising technology costs with improvements in profitability, productivity, client retention or growth.

The May 2026 survey covered 40 leading RIAs, wealth managers and broker-dealers representing $8.6 trillion in assets. Those respondents were drawn from F2 Strategy’s broader national cohort of 116 firms overseeing approximately $31 trillion. The scale of the participating institutions indicates that the measurement problem is not limited to small advisory practices with limited technology resources. It extends to some of the industry’s largest and most sophisticated organizations.

F2 said most wealth management firms in the survey had no formal method for evaluating AI projects. None of the participating bank and trust organizations had established one. The absence of consistent measurement means firms may know that an application saves time or generates positive employee feedback without knowing whether the benefit exceeds the combined cost of software licenses, computing consumption, integration, data preparation, governance and organizational change.

The findings arrive as AI moves from discretionary experimentation into recurring operating expenditure. Wealth firms are deploying generative AI assistants, meeting-note applications, document-processing systems, workflow automation, compliance tools and increasingly sophisticated agent-based software. Some projects are purchased from established wealth technology vendors, while others are built through combinations of commercial language models, proprietary data and custom software.

F2’s research indicates that firms are concentrating primarily on efficiency rather than immediate revenue creation. Common objectives include reducing administrative work, increasing advisor capacity, accelerating client onboarding and allowing existing employees to support a larger number of households. That approach reflects the economics of wealth management, where advisor time is constrained and growth can require repeated hiring across operations, compliance and service functions.

AI could allow a firm to expand assets and accounts without increasing headcount at the same rate. However, a headcount-neutral growth strategy only creates measurable value when management can identify the tasks being automated, quantify the hours released and determine how that capacity is redeployed. Time saved by an advisor or operations employee does not automatically become financial return if the employee’s workload, client responsibilities or revenue-producing activity remain unchanged.

The survey’s most immediate infrastructure warning concerns data. Sixty-four percent of wealth management respondents said they did not have a unified data layer capable of making their AI projects work effectively. Among bank and trust respondents, the proportion rose to 83%.

Wealth firms frequently store client information across separate custodial platforms, customer relationship management systems, planning software, portfolio accounting tools, document repositories, communications archives and compliance systems. When those records use different formats or contain conflicting information, an AI application may be unable to produce a reliable view of a household, portfolio or workflow without extensive reconciliation.

Fragmented systems tend to restrict firms to broad, generic uses of large language models, such as summarizing documents, drafting internal material or transcribing meetings. More advanced applications—including personalized portfolio analysis, automated service recommendations and multi-step operational agents—require controlled access to accurate, current and consistently labeled information.

The data problem also affects risk management. Wealth managers handle sensitive financial, tax, estate-planning and personal information, and recommendations may be subject to fiduciary, supervisory and recordkeeping requirements. Firms therefore need to know which information an AI system can access, how outputs are generated, where human review is required and whether a decision can be reconstructed after the fact.

A unified data layer does not eliminate those obligations, but it can make controls easier to apply. Consistent permissions, data lineage and monitoring can help a firm determine which records informed an output and whether the system relied on incomplete or outdated information. Without that foundation, expanding AI across an organization can magnify existing weaknesses in data quality and operational governance.

The minority of surveyed firms that measured their AI initiatives reported materially stronger evidence of value. Among those organizations, 68% said they achieved at least 25% more efficiency in the targeted workflows. The result suggests that measurable returns are possible when projects are limited to defined processes and evaluated against an operating baseline.

Targeted workflow measurement is substantially different from attempting to calculate a single return for an enterprise-wide AI strategy. A firm can compare the time required to process an account-opening package before and after automation, track the number of documents requiring correction or measure changes in average handling time for service requests. It is harder to assign a value to broad claims that AI makes an organization more innovative or better prepared for the future.

F2 identified a growing separation between AI leaders and firms it described as middle-stage adopters or laggards. Leading organizations are assembling technology stacks that combine general productivity tools, specialized financial applications and custom agents. The consultancy estimates that those firms may hold a 12- to 24-month advantage over competitors that have not completed the necessary data and operating-model work.

That lead may be important in an industry where technology capabilities are increasingly part of recruiting, acquisition and client-service strategies. Large advisory consolidators can distribute the cost of technology across many offices, while private-equity-backed wealth platforms are under pressure to increase margins and demonstrate operating leverage. Smaller RIAs may depend more heavily on third-party platforms but can sometimes move faster because they have fewer legacy systems and organizational layers.

Ownership and governance pressures are helping drive spending. F2 said some investments are being encouraged by private equity sponsors, boards and shareholders that want firms to remain competitive as AI develops. In those cases, management may view adoption itself as strategically necessary, even when a traditional return calculation is not yet available.

That reasoning can be defensible for foundational projects, but it also increases the risk of indiscriminate purchasing. Providing every employee with an AI license or adding a meeting assistant across an organization may demonstrate activity without establishing whether the tools solve a high-priority business problem. Firms can also accumulate overlapping products that perform similar functions while creating additional integration, security and supervision requirements.

The expanding use of consumption-based pricing introduces another measurement challenge. In addition to fixed software fees, firms may pay according to the number of prompts, tokens, automated tasks or computing resources used. Costs can therefore rise as adoption grows, particularly when autonomous systems perform repeated actions or analyze large volumes of material.

F2 expects such “token economics” to become a more significant issue during the 2027 budgeting cycle. A meaningful evaluation must capture not only vendor contracts but also variable model usage, implementation work, data engineering, employee training, cybersecurity review, legal oversight and the cost of maintaining parallel manual processes while a system is being tested.

The return side of the equation also needs greater precision. Depending on the use case, a firm could measure fewer processing errors, shorter onboarding times, lower service costs, increased advisor capacity, improved response times or higher client satisfaction. Revenue-oriented applications could eventually be evaluated through referral conversion, asset consolidation, household retention or the ability to deliver profitable advice to smaller accounts.

Executives must also distinguish between theoretical capacity and realized value. An application that saves an advisor five hours each week has created operational capacity, but its economic value depends on what happens next. The firm may need to redirect that time toward prospecting, financial planning, client reviews or relationship management before the efficiency appears in revenue or retention results.

Bank of America’s recent expansion of EricaAssist illustrates how large financial institutions are beginning to communicate AI benefits through specific operating metrics. The bank said more than 18,000 customer service representatives use the system, which provides contextual guidance during client calls in less than three seconds and has reduced average call time by nearly one minute. Bank of America spends $14 billion annually on technology, including more than $4 billion for new initiatives that include AI.

Although the bank’s program extends beyond wealth management, its reported metrics show the type of operational baseline firms can use to evaluate an AI investment. Usage levels, response speed and handling-time reductions can be monitored before management attempts to translate them into broader financial outcomes.

For wealth managers, the most credible near-term returns are likely to come from narrowly defined administrative and service processes rather than fully automated investment or planning decisions. Document validation, meeting preparation, call summarization, information retrieval and workflow routing can be measured more easily and generally preserve a clear role for employee review.

Client-facing applications create a higher standard. Affluent and high-net-worth households frequently have complex tax, estate, liquidity and concentrated-stock issues that do not fit standardized recommendations. An incorrect summary or unexplained portfolio suggestion can damage trust even when the technology performs accurately in most cases. Firms therefore need to evaluate output quality and client impact alongside labor savings.

Change management represents another source of hidden cost. Advisors and employees may avoid tools they do not trust, maintain duplicate records or continue manual procedures after an application has been introduced. A technically successful installation can consequently fail to generate value because it was not integrated into job responsibilities, performance expectations and supervisory processes.

F2’s report argues that every AI project should be accompanied by a change-management strategy. That requires decisions about which tasks will change, which employees retain approval authority and how released capacity will be used. It may also require training experienced employees for more client-facing or judgment-intensive roles rather than treating automation exclusively as a workforce-reduction tool.

The research does not suggest that wealth managers will reduce AI investment. F2 expects budget commitment to continue and does not anticipate a slowdown in 2027. Instead, the findings indicate that the basis for approving projects is likely to become more demanding as variable costs become clearer and boards request evidence that multiple pilots are producing durable improvements.

The next competitive phase will therefore be shaped by measurement discipline as much as technological capability. Firms that define a problem, document current performance, establish cost controls and assign responsibility for capturing the benefit will be better positioned to decide which applications deserve wider deployment.

Organizations that lack those practices may continue spending because AI is viewed as strategically unavoidable. But without unified data, clearly defined objectives and credible performance measures, they will have limited ability to distinguish genuine transformation from an increasingly expensive collection of technology experiments.