Enterprise RIAs have capable technology teams, proprietary data and a strong incentive to own more of their investment experience. As AI makes software development faster, building an internal personalization engine can look increasingly practical.
But there’s a difference between building an application that recommends an action and maintaining the infrastructure that safely executes that action across every account, custodian and market condition.
We asked Brian Ross, CEO of Flyer Financial Technologies, where the do-it-yourself approach can create problems and what firms should consider before turning an AI prototype into production infrastructure.
WSR: Why are enterprise RIAs considering building their own AI-powered investment workflows?
Ross: Because the initial use case can look deceptively achievable.
A firm may start with a narrow objective: identify tax-loss harvesting opportunities, interpret natural-language requests, recommend trades or create a personalized rebalance. With access to portfolio data and modern AI tools, a development team can produce an impressive prototype relatively quickly.
The problem is that the prototype usually demonstrates the validity of the intelligence layer, not the execution layer.
In 2024, Boston Consulting Group studied 1,000 senior executives across 59 countries and found that only 26% of companies had developed the capabilities needed to move beyond AI proofs of concept to produce tangible value. It also found that successful AI leaders focus 70% of their AI transformation efforts on people and process issues, and only 10% on algorithms.
I agree with that finding, because the model is rarely the hardest part of a wealth management workflow. The difficult work is permissioning, data quality, operating processes, exception management, governance and integration.
WSR: What might a successful internal pilot fail to reveal?
Ross: It may not reveal what happens when real accounts stop behaving like test accounts.
A pilot might use complete tax-lot data, clearly mapped households, one custodian and a limited number of securities. But production systems encounter missing records, delayed files, partially filled orders, inconsistent identifiers, account restrictions, custodian-specific instructions and market movements that change the appropriate action halfway through a workflow.
Imagine an internal AI system that identifies a tax-loss harvesting opportunity and prepares replacement trades. In the test environment, it works exactly as expected.
In production, the client’s spouse owns the replacement security in an account at another custodian. One account has incomplete tax lot information. Another custodian cannot accept the firm’s preferred closing instructions electronically. A partial fill changes the cash available for the remaining trades.
The original recommendation may have been reasonable, but the workflow is no longer safe to execute as designed.
Flyer’s point of view is that this is where an action layer becomes necessary. An agent’s proposed action must pass through an environment that determines whether the information is sufficient, whether the action is permitted and whether it remains appropriate as conditions change.
WSR: Isn’t this primarily a data engineering problem?
Ross: Data is one part of it, but “having the data” is not the same as having execution-ready information.
In 2024, Gartner surveyed data management leaders and found that 63% of organizations either lacked the right data management practices for AI or were unsure whether they possessed them. The firm predicted that through 2026, organizations would abandon 60% of AI projects that were not supported by AI-ready data.
I agree with the warning, but wealth management firms should take it one step further. Data can’t just be available to the AI; it has to be current, normalized, sufficiently complete and appropriate for the specific action being considered.
A portfolio snapshot that was accurate at the start of the day may not reflect a later trade, cash movement, model update or partial fill. A household relationship stored in one platform may not be visible to another. Tax lot instructions can differ by custodian and asset type.
Once AI is allowed to act, these are no longer ordinary integration inconveniences. They become decision inputs with direct financial and compliance consequences.
WSR: What is the biggest risk of building this internally?
Ross: The firm may underestimate the permanent operating responsibility it is accepting.
A homegrown execution system is not a one-time development project. Every custodian change, API revision, asset-class expansion, compliance rule, workflow variation and exception becomes the firm’s responsibility to interpret, test and maintain.
Flyer’s competitive analysis highlights some of the less visible infrastructure behind multi-custodial execution: normalized order flows, custodian-specific routing rules, file-based fallbacks where electronic connectivity has gaps, allocation handling, streaming position updates, webhooks, sandbox environments and testing for unsuccessful or unusual scenarios.
These capabilities are easy to overlook because they sit beneath the user experience and typically only become visible when something goes wrong.
A firm may believe it is building a personalized trading assistant. Over time, it can discover that it has also taken responsibility for maintaining a routing layer, an exception-management system, a testing environment, a compliance record and an operational support function.
WSR: Doesn’t building internally provide more control?
Ross: It can, but only when the firm distinguishes strategic control from technical ownership.
Enterprise RIAs should absolutely control their investment philosophy, personalization rules, client experience and approval policies. The harder question is whether they also need to own the work that happens underneath those decisions: pulling data from every custodian and system in play, normalizing it into a single trustworthy record, validating it before it drives a trade and orchestrating it all against a timeline that doesn’t pause for a late data feed or an incomplete file.
There’s a strong case for maintaining proprietary logic while relying on specialized infrastructure for functions such as order management, trade routing, custodial normalization, simulation, compliance checks and audit records.
This allows the firm to retain control over what differentiates it without rebuilding capabilities that are complex but largely invisible to the investor.
McKinsey recently argued that durable value in AI-enabled wealth management is likely to accrue to organizations that control the points connecting recommendations to permissioned data, compliant controls and execution rails, which are the technical and legal ability to securely turn a recommendation into a completed transaction. I agree. The strategic asset is not necessarily a completely proprietary stack. It is the firm’s ability to determine how its judgment is translated into governed action.
WSR: What happens when a system built for one use case begins to expand?
Ross: The architecture can become a constraint before the firm realizes it.
A tax-loss harvesting prototype may later be expected to support household rebalancing, cash management, direct indexing, UMA sleeves, fixed income or advisor-initiated natural-language requests.
Each expansion introduces new dependencies and edge cases. Logic that worked for equities at one custodian may not work for mutual funds or fixed income. A workflow designed for centralized trading may not accommodate Rep-as-PM permissions. A system built around account-level decisions may struggle when household, sleeve and firm-level rules conflict.
This is where an apparently inexpensive internal build can create substantial downstream costs. The firm is no longer adding features to an application. It is modifying execution infrastructure while that infrastructure is connected to live investor accounts.
WSR: Are you saying enterprise RIAs should never build?
Ross: No. The better question is where they should build.
Firms should invest aggressively in the capabilities that express their distinctive value: investment methodology, advisor experience, client preferences, tax policies, workflow design and proprietary decision logic.
They should be more cautious about recreating the underlying systems responsible for continuously moving, checking, routing, reconciling and documenting those decisions.
In 2024, Gartner projected that at least 30% of generative AI initiatives would be abandoned by the end of 2025 after proof of concept because of poor data quality, inadequate controls, rising costs or unclear value. I agree that those four issues provide a useful test for enterprise RIAs. Before approving an internal build, leadership should ask whether the firm has accounted for the full production environment, not merely the cost of producing the prototype.
WSR: What should firms ask before moving an internal prototype into production?
Ross: Ask who owns the action when the clean scenario ends.
Who confirms that the data is sufficient? Who maintains every custodial connection? Who tests unsuccessful workflows? Who decides when the agent should stop? Who updates the system when policies change? Who reconstructs the decision when compliance asks what happened six months later?
FINRA’s oversight guidance continues to place responsibility for technology management, supervision, books and records, third-party risk and applicable regulatory obligations on the firm. Using AI does not transfer that accountability to the model.
Enterprise RIAs have every reason to own their personalization strategy.
The trouble begins when owning the strategy is assumed to require rebuilding the execution infrastructure beneath it. The firms best positioned for AI-enabled personalization will know which capabilities create differentiation and which require specialized, continuously maintained infrastructure.
Building the demonstration is becoming easier.
Building a system that still works safely when the data is incomplete, the market has moved and thousands of accounts are waiting is a different undertaking.
Wealth Solutions Report can be reached at info@wealthsolutionsreport.com.
This article is published under WSR’s partner program. For more information on how to participate in the partner program, contact zack.drew@wealthsolutionsreport.com. Views expressed are the interviewee’s and do not necessarily reflect the views of WSR.
