Three and a half years after the release of ChatGPT, most institutions have worked through a broadly similar sequence of responses. Setting that sequence out plainly is worthwhile, to both frame the frustrating time to value, and to address it.
How AI adoption in financial services has evolved
The first response was restriction. Through late 2022 and much of 2023, the sanctioned position at most large institutions was to block access at the network perimeter. That has since been characterized as excessive caution, which understates the position considerably. No institution had adequate assurance on data retention, residency or client confidentiality, and the consequence of a disclosure breach in this industry is supervisory as well as reputational. The unintended effect was that the industry’s first formal act on the most consequential general-purpose technology in a generation was a prohibition, enforced at the firewall and comprehensively circumvented on personal devices.
The second response was licensing. Enterprise agreements, a sanctioned internal assistant, usually branded with the institution’s own initials. This generated real benefit in drafting, summarization and code review, but it generated that benefit at the level of the individual rather than the institution. Time returned to several thousand employees in half-hour increments does not appear in divisional P&L, because individual efficiency gains only aggregate once the underlying process changes shape. Firms bought productivity they had no mechanism to capture.
The third response was the pilot portfolio. Innovation councils, centers of excellence and a formal call for use cases across the business. Submissions that cleared governance review tended to be those that were easiest to approve, and the proposals easiest to approve were, almost by definition, those furthest from any process carrying commercial or regulatory consequence. The selection mechanism optimized for approvability and treated expected value as a secondary criterion. Two years later the industry appears genuinely surprised that a portfolio assembled on that basis has underdelivered.
The fourth response, where most institutions currently sit, is unstructured retrieval. Research archives, policy libraries, credit agreements, prospectuses and term sheets, indexed and queried in natural language. The work is useful, not to be dismissed.
None of this arrived at a uniform industry. Capital markets accommodate a wide spectrum of AI sophistication, ranging from desks that have run machine learning in production for well over a decade to businesses still intermediated by documents and telephone calls. The recent wave reached all of them simultaneously, with much the same tooling and much the same board-level urgency. That spectrum accounts for the dispersion in outcomes without accounting for the aggregate.
Each response altered the interface through which the institution reached a model. None of them altered the corpus the model was permitted to reason over. In most firms that corpus remains documentary, and documents are abundant, unstructured, largely unencumbered and subject to no lineage obligation whatsoever. The economically significant data sits elsewhere, in orders, executions, positions, quotes, financing balances and risk factors, all of it structured, transactional, entitled and reconciled. Your assistant has read every policy document in the firm, yet it has never seen a trade.
What is holding back AI in financial services?
The four main constraints are repeatability, entitlements, semantics, and data perimeter and alpha leakage. All four are matters of engineering and governance rather than model capability.
Repeatability and AI model risk
Ask a large language model an identical question twice and you may receive two materially different answers. In most commercial settings that variance is often desirable. In a regulated front office, it can be disqualifying, and understanding precisely why is essential to understanding where adoption has stalled. Constraining sampling parameters reduces the variance without eliminating it, which is a distinction a model validation function will not overlook.
Model risk management practice is formally supervised and rests on the assumption that a model is a stable function: something capable of being documented, independently validated, versioned, back-tested and subjected to effective challenge.
Reproducibility is not one requirement among many within that framework. It is the precondition for every other requirement in it. A function whose output varies under identical inputs cannot be independently validated, and a process that cannot be replayed cannot be effectively challenged.
Industry hesitancy is therefore neither cultural conservatism nor a deficit of imagination. The framework contains no category into which a non-deterministic component can be placed, and the individuals required to approve its use are personally accountable under the Senior Managers Certification Regime for outcomes they would be unable to reconstruct.
Entitlements and data access
AI systems in financial services must enforce data entitlements before retrieval, not after an answer has been generated.
An agent with access to firm-wide position data constitutes an information barrier incident in waiting. A trader posing an entirely reasonable question should not receive an answer quietly synthesized from a desk they have no permission to see.
The subtlety that catches firms out is that a permissioned answer is not the same thing as a permissioned data set. An aggregate, a ranking, or even a carefully worded refusal can each disclose the existence and direction of a position without exposing a single restricted record. Entitlement therefore must be enforced before retrieval rather than applied to output after the event, and it must carry the same client confidentiality and research separation obligations the firm already operates under. Very few internal deployments carried any concept of entitlement at all, which is a substantial part of why they were pointed at documents in the first place.
Semantic data and AI in capital markets
AI in capital markets needs a shared understanding of what financial data means before it can reliably reason across that data.
The term “benchmark” denotes one thing in trading, another in risk and another again on the prime desk, and each definition is correct within its own context. Practitioners reconcile these differences instinctively and know whom to consult when they cannot. Software does neither. What makes this expensive is the shape of the failure. An assistant reasoning across business lines rarely declines to answer. It returns something fluent, plausible and subtly wrong, which is considerably more dangerous than an obvious error because it survives review. Gartner expressed the point more bluntly earlier this year, observing that an absence of semantics produces inaccurate agents and wasted expenditure.
Data perimeter and alpha leakage
Keeping sensitive trading and investment data within the firm’s security perimeter is essential to deploying AI safely in capital markets.
Routing position and order data to a third-party inference endpoint is a conversation with compliance that few executives are willing to have more than once. Orders, executions, hedging patterns and position histories are the firm’s intellectual property.
They describe how a desk sizes, when it steps away and where it is prepared to take risk, and they retain commercial value long after the trade has settled. Every prompt and every unit of context crossing an organizational boundary is a description of strategy, and questions of retention, logging, sub-processing and downstream training use compound that exposure rather than resolve it. The perimeter question therefore resolves only through an architecture in which the data does not move and alpha cannot leak.
How can deterministic AI work with probabilistic models?
Reproducibility is a property of the system rather than a property of the model. A system can therefore be made deterministic even where one of its components is not.
This is the point at which the argument turns from diagnosis to design.
The architecture that achieves it separates two planes explicitly. The probabilistic plane reasons, interprets ambiguity, synthesizes and proposes; it is permitted to be inventive, and it is permitted to be wrong. The deterministic plane owns every calculation, every figure that will subsequently be quoted, the lineage back to source and the record itself. Output from the model carries the status of a proposal, and the record decides. Where a firm’s benchmark and calculation methodologies are encoded within that deterministic layer, an identical question returns an identical figure on every occasion, traceable to origin, irrespective of how the question happened to be phrased.
That separation is what converts an unauditable assistant into something a validation function can assess. The object under validation becomes the deterministic layer and the policy gates that govern it, both of which behave precisely as a model validator expects them to. The probabilistic component is then assessed on an entirely different basis, as a proposal generator constrained by what it is permitted to see and what it is permitted to initiate.
Two further design decisions carry comparable weight. Entitlement filtering must occur before the model receives data rather than being applied to the answer it has already produced, because an information barrier is a property of the data layer or it is not a property at all. And human approval must be an architectural feature rather than an operational undertaking, which means that anything carrying consequence requires an explicit human decision, and that the resulting instruction routes through the institution’s existing pre-trade controls without modification. The commercial significance of that last point is routinely underestimated. Adopting the capability does not require a firm to re-approve its control framework, because the control framework is left entirely intact.
Why does time to value matter for AI adoption?
For enterprise AI in financial services, the commercial value of a solution increasingly depends on how quickly it can move from implementation to a defensible production output.
Enterprise technology programs in this industry run to twelve or eighteen months for anything touching a core system, while the useful life of a frontier model is now measured in quarters. This is the first technology I can readily identify where the implementation cycle outlasts the shelf life of the thing being implemented. Institutions commission extended programs to deliver capability that will be two generations obsolete at the point of go-live and then conclude that the technology disappointed them.
Time to value is therefore the metric that decides the outcome, rather than benchmark scores or parameter counts. What matters is the number of days between the decision to proceed and the first output somebody is prepared to defend in front of a client or a supervisor.
We built the foundation before the question was widely asked, less through foresight than because clients required trading, risk, prime brokerage and wealth to run from a single multi-asset data spine. Governed definitions, deterministic calculation, lineage back to origin, entitlement enforced at the data layer, and the whole of it resident inside the client’s own perimeter.
With TSIQ, that foundation is now enabling us to bring intelligence into these workflows in days rather than quarters.
Frequently Asked Questions (FAQs)
What is preventing AI adoption in financial services?
The main barriers to AI adoption in financial services are repeatability, data entitlements, semantics, and data perimeter and alpha leakage. These are primarily engineering and governance challenges rather than limitations in AI model capability.
Why is structured data important for AI in capital markets?
Structured data contains much of the commercially significant information used across capital markets, including orders, executions, positions, quotes, financing balances and risk factors. Giving AI governed access to this data allows it to support workflows and decisions rather than being limited to searching documents and generating summaries.
Can AI be deterministic if large language models are probabilistic?
Yes. Reproducibility can be a property of the overall system rather than the underlying AI model. A probabilistic AI layer can interpret questions and propose actions, while a deterministic layer controls calculations, source data, lineage and governed outputs.
What is the role of AI governance in financial services?
AI governance ensures that AI operates within the institution’s existing requirements for data access, model risk, confidentiality, auditability and human oversight. For commercially or regulatorily consequential workflows, governance needs to be built into the architecture rather than added as an operational process afterward.
How can financial institutions prevent AI from accessing restricted data?
Data entitlements should be enforced before information reaches the AI model. This ensures that the model only receives data the user is authorized to access, rather than attempting to identify and remove restricted information after an answer has already been generated.
What is alpha leakage in AI?
Alpha leakage occurs when sensitive trading or investment information, such as orders, executions, position histories or hedging patterns, leaves the firm’s controlled environment. For financial institutions, preventing this requires an architecture in which sensitive data can remain within the firm’s own perimeter.
How does a semantic layer improve AI in capital markets?
A semantic layer gives AI a governed understanding of what financial data means in its specific business context. This is important because the same term, such as “benchmark”, can have different meanings across trading, risk and prime brokerage. Without that context, AI can produce answers that are fluent and plausible but materially wrong.
What is time to value for AI?
Time to value is the period between deciding to implement an AI capability and producing the first output that someone is prepared to defend in front of a client or supervisor. For financial institutions, reducing this period is increasingly important as AI models evolve faster than traditional enterprise technology programs can be delivered.
This article was originally posted on tsimagine.com.
