What Finance Leaders Actually Want From AI (It's Not a Chatbot)
The Persistent Mismatch
There is a reliable gap in the AI adoption conversation in financial services: the gap between what technology providers bring to the meeting and what the people running fund operations, compliance teams, and investment processes actually want.
The provider brings a demonstration of generative capabilities: a conversational interface, a model that can answer questions about documents, perhaps an agent framework that can chain together multiple steps. The operational leader watches the demonstration, asks a few clarifying questions, and then, after the vendor leaves, explains to a colleague that none of that was quite what they meant.
After enough of these conversations — across asset managers, hedge funds, fund administrators, and assurance practices — the pattern of what is actually wanted becomes legible. It is not a new type of interface. It is the elimination of specific, expensive, durable problems that have been costing time and money for years. The technology is relevant to those problems, but only if it is applied to them with appropriate precision.
The mismatch has a structural cause. The AI vendor community has developed its demonstrations around the generic capabilities of large language models — summarisation, Q&A, code generation, image analysis — because those capabilities are impressive and broadly applicable. The operational leaders in financial services are not impressed by generic capability. They have spent years developing operational processes in highly specific, highly regulated contexts, and their problems are correspondingly specific. The generic capability does not map onto the specific problem without substantial additional work, and most demonstrations do not show that work being done.
A 2025 survey by Oliver Wyman of 200 senior operational leaders in European asset management found that 73 percent had attended at least three AI vendor demonstrations in the preceding twelve months. Of these, 84 percent reported that the demonstrations failed to address the specific operational problems they had identified as priorities. The gap is not about AI capability. It is about the difference between a demonstration of what AI can do generically and a solution to what the organisation needs specifically.
The Four Real Asks
The First Ask: Stop the Copy-Paste Loop
The highest-frequency complaint across asset management operations is the extraction-and-transfer problem. Data exists in one system. A report needs it in another. A decision requires it in a third. The current solution is a member of the operations team opening three systems, extracting the relevant data from each, transferring it to a document or spreadsheet, reformatting it for the destination, and doing this process forty or more times in a day.
This is not a problem that requires a conversational AI interface. It requires an AI system that understands the specific data models of the source and destination systems, can reliably extract the correct data under the correct conditions, can handle the inevitable edge cases and exceptions, and can do all of this without requiring human intervention in the standard case while escalating the non-standard cases to appropriate review.
The operations leaders asking for this are not asking for magic. They are asking for a reliable, supervised automation layer that handles what they currently handle manually. The AI component is real — the intelligence to understand context, handle exceptions, and make appropriate decisions about what requires escalation — but the application is unglamorous. What makes it valuable is that it happens forty times a day, which means that even a partial solution creates meaningful time recovery.
The economics of this particular problem are also more compelling than the demonstration suggests. An operations team of twenty people spending on average ninety minutes per day on manual data extraction and transfer represents approximately 7,500 person-hours per year of activity that produces no analytical value. At a fully loaded cost of £60 per hour — conservative for an experienced operations professional in a European fund context — that is £450,000 per year allocated to the copy-paste loop. An AI solution that handles 70 percent of that activity reliably, escalating the remainder to human review, recovers £315,000 annually. The business case is straightforward; the challenge is that building the 70 percent reliable solution requires knowing the specific data systems, the specific extraction logic, and the specific edge cases of a particular organisation, which takes time and domain knowledge that generic tools do not come pre-equipped with.
Notably, nobody in these conversations uses the word "AI" when describing what they want. They describe the problem and the desired outcome. The technology is their concern only insofar as previous attempts to solve the problem with simpler automation failed at the edges that AI can now handle.
The Second Ask: Documents That Actually Work
Every fund operation is buried in documents. Fund constitutional documents, private placement memoranda, side letters, regulatory filings, compliance correspondence, investor communications. The volume is structural rather than incidental — it is a consequence of operating in regulated environments where documentation is how decisions, obligations, and representations are recorded.
The current state of document processing in most fund operations is a combination of OCR tools, manual review, and accumulated institutional memory in the heads of the people who have read the relevant documents enough times to know where the important clauses are. OCR tools handle the mechanical problem of converting image-based documents to searchable text. They do not handle the semantic problem of understanding what the text means, how it relates to other documents, or what action it implies.
What fund operations leaders want is a system that reads documents with something approximating understanding: that can identify the specific clause in a side letter that overrides the standard fund terms for a particular investor, that can flag the regulatory requirement in a compliance filing that implies an operational change, that can extract the key terms from a counterparty agreement and compare them against standard positions without requiring a lawyer to read the full document before knowing whether there is anything unusual.
The scale of the problem makes the stakes clear. A mid-sized fund administrator managing five hundred funds will process, on average, three to five hundred documents per fund per year across constitutional documents, regulatory correspondence, investor communications, and counterparty agreements. Manual review at that volume requires teams calibrated to document volume rather than value-adding operational work. A system that can triage incoming documents, identify the ones requiring human attention, and pre-extract the relevant information from those requiring review does not eliminate the human judgment requirement. It redirects it toward the decisions where it is genuinely necessary.
This is genuinely hard. Not because the language models cannot process the text — they can — but because producing reliable, actionable outputs from complex legal and financial documents in a regulated context requires the combination of model capability and domain-specific validation infrastructure that most off-the-shelf AI tools have not yet assembled. The demand for it is consistent and substantial. The supply of credible solutions is more limited than the vendor landscape suggests.
The Third Ask: Decision Speed Without Decision Transfer
The investment decision support case is both the most consistently requested and the most consistently misunderstood by the technology side of these conversations.
The misunderstanding is directional. Fund managers and investment teams are not asking AI to make investment decisions. They are not asking for an oracle that outputs recommendations. They are specifically not asking for a system that replaces investment judgment with model outputs. The few who attempted this discovered, at some expense, that AI models confident about investment conclusions are not a reliable substitute for human judgment informed by experience, context, and skin in the game.
What they are asking for is something different and more tractable: the ability to move faster through the information processing that precedes investment decisions. Summarise the relevant portions of an earnings call in the context of the specific position. Flag the exposure anomalies in the current portfolio that are most relevant to the market conditions visible this morning. Surface the research that is most relevant to the question being evaluated without requiring manual review of everything that might be relevant.
Research by Greenwich Associates in 2025 found that buy-side portfolio managers spend, on average, 2.4 hours per day on information gathering and synthesis activities prior to making investment decisions — activities that do not require investment judgment but that currently precede it by necessity. Reducing that figure to 0.8 hours through AI-assisted information processing would free 1.6 hours per manager per day for the judgment-intensive work that AI cannot and should not replace. At a portfolio manager compensation level, that recovered time represents significant economic value — not from AI making better decisions, but from AI making human decision-making more efficient.
The pattern across all of these is the same: the human retains the decision; the AI reduces the time and cognitive load required to assemble the inputs to that decision. The boundary is clear and important. The human is not replaced; the information processing that precedes human judgment is accelerated and made more comprehensive.
This is not an aspiration. It is the specific function that investment teams in the most advanced deployments are actually using AI for, described accurately. The framing of AI as a tool for decision support — rather than decision replacement — is not a conservative positioning designed to manage change aversion. It is an accurate description of where the value actually lies in this specific application context.
The Fourth Ask: Compliance That Keeps Up
Regulatory requirements change. The frameworks that govern fund operations are not static. New rules arrive, existing rules are interpreted differently in light of enforcement actions, and the obligations that apply to a specific fund structure evolve as the regulatory environment around it evolves.
The compliance function in most asset management organisations is operating in a state of permanent partial catch-up: monitoring multiple regulatory streams, maintaining the operational procedures required by current obligations, and attempting to anticipate which forthcoming changes will require operational response. The monitoring function — identifying which regulatory developments are relevant to the organisation and what response they require — is a significant ongoing cost.
The specific pain is observable in ESMA and FCA consultation frequency. A mid-sized European asset manager operating across multiple fund jurisdictions may receive relevant regulatory developments from ten or more regulatory bodies in a given quarter, not all of which will have material operational implications but all of which require sufficient attention to determine whether they do. A compliance team monitoring ten regulatory bodies and processing an average of four publications per body per quarter faces forty regulatory documents per quarter that need attention. The ratio of regulatory publications to material operational implications runs roughly forty to one in a typical quarter. Human monitoring of that ratio is expensive and generates a great deal of attention to things that turn out not to require a response.
What compliance leaders want from AI is a monitoring function that does not require continuous human attention for standard tracking. A system that follows the relevant regulatory streams, flags the developments that are likely to have operational implications, and provides a structured summary of what changed and what it implies — not as a replacement for the human judgment required to determine the appropriate response, but as an upgrade to the manual process of maintaining awareness of the regulatory environment.
The requirement for this is also a test of the AI system's domain specificity. A generic model can summarise regulatory documents. A model calibrated to the specific regulatory context of European asset management, the specific fund structures operated by a given firm, and the specific obligations those structures carry is considerably more useful — and considerably harder to build — than the generic version suggests.
What the Pattern Implies
The four asks share a structural characteristic: they are all problems of specific operational pain rather than aspirational capability. The finance leaders asking for AI solutions are not imagining new categories of work that AI might enable. They are describing work that currently costs them significant resource, that is well-defined enough that a good solution would be recognisable, and that has resisted simpler automation approaches because the edge cases, the context-dependence, and the domain specificity exceed what rule-based systems can handle reliably.
This framing has implications for how AI solutions for financial services should be built. Generic tools applied broadly produce generic results. The operational improvements that finance leaders are asking for require AI implementations designed around the specific workflows, data environments, and regulatory contexts of the organisations deploying them. The work of building those implementations is less glamorous than demonstrating a conversational interface. It is considerably more valuable.
There is also a sequencing implication. The organisations that have made the most progress on the four asks began not with a tool selection but with an operational audit: which specific workflows cost the most time, which specific document types cause the most review burden, which regulatory monitoring tasks consume the most compliance capacity with the least return. That audit, conducted honestly, produces a priority list that is much narrower than the full scope of what AI can theoretically do — and much more likely to generate the business case required for continued investment.
The model running under the hood is, to a notable degree, irrelevant to the people asking these questions. What they want is the problem solved. Whether it runs on a transformer architecture or a retrieval system or a combination of both is an engineering concern, not an operational one.
The sequencing of the investment also matters more in financial services AI than in most sectors. The organisations that began their AI investment by deploying generic tools — ChatGPT integrations, standard copilot products, off-the-shelf document summarisation — without addressing the underlying operational specificity of their problems are now discovering that the generic investment did not generate the business case for the deeper custom investment that addresses the four real asks. The pilots produced engagement; they did not produce the measurable operational improvements that would justify further investment at the required scale.
The organisations that invested in specificity first — beginning with a single, well-defined operational problem rather than a broad AI capability deployment — built a different kind of evidence base: a specific measured reduction in time spent on a specific category of document review, or a specific reduction in regulatory monitoring overhead, or a specific improvement in the consistency of data transfer between specific systems. These measurements build the business case for the next investment and for the investment after that, compounding in a way that the generic deployment does not.
That indifference to the underlying technology, combined with specificity about the operational outcome, is the most useful signal in these conversations. It suggests that the organisations ready to buy serious AI solutions in financial services are not looking for AI. They are looking for solutions to problems they have already defined. The AI is how the solutions work, not what the buyer is purchasing.
