Financial services firms spent over $20 billion on AI in 2025. Most of them have little to show for it.
The models aren't the problem. GPT-5 and Claude Opus are more than capable of handling the research, analysis, and synthesis tasks that define investment management. The capital isn't the problem either, firms have committed it. PwC's most recent global CEO survey found that 56% of executives report zero financial return from their AI investments.
The failure patterns are consistent enough across hedge funds, asset managers, and private equity firms that they're now predictable. MIT's Project NANDA found that roughly 95% of enterprise generative AI pilots produce no measurable profit-and-loss impact. FinTellect AI puts the number even harder: 80% of AI projects in financial services never reach production, and of those that do, 70% fail to deliver measurable business value.
The firms that beat those numbers don't have better technology, but they diagnose the actual problem before they build anything.
This piece covers the three root causes of AI adoption failure in financial services, and what separates the firms that get it right.
In 2017, major frontier model releases came roughly every 15 months. By 2025, that interval had compressed to six weeks. Whatever a firm needs models to do, the capability is there.
Finance hasn't kept up. MIT's data show that AI-related structural change in financial services has lagged nearly every other sector. Tech firms and professional services firms absorbed the technology and restructured around it. Finance, despite being among the heaviest AI spenders, largely hasn't. McKinsey's 2025 global survey found that only 7% of organizations report AI as fully deployed and integrated enterprise-wide.
The problem isn't what the models can do. It's what happens between the model and the analyst's desk.
A frontier model doesn't know your IC memo format. It doesn't know which version of a deal model to reference, which broker research subscriptions your analysts actually use, or how your investment committee frames a risk recommendation. Deploying a generic AI tool into a firm-specific workflow and expecting adoption is like handing a new hire an IQ test and asking them to produce a credit memo on day one. The test score is irrelevant because the context isn't there.
92% of financial firms using vertical AI solutions already have access to GPT or Claude Enterprise. The model isn't what's missing.
The most common failure in enterprise AI adoption is misdiagnosis. Firms build a solution before they understand the problem.
RAND Corporation's 2024 analysis of 65 experienced AI practitioners found that misunderstood problem definition is the top root cause of AI project failure, where stakeholders never agree on what the AI is actually supposed to solve. In financial services, the pattern is consistently predictable.
Leadership gets told to figure out AI. Someone decides to deploy a broad, general-purpose tool, a ChatGPT Enterprise license, a third-party research platform, or an off-the-shelf summarization tool across the organization. The logic is reasonable, to give analysts a general capability and let them find uses for it.
But analysts aren't looking for a general capability. They're looking for help with a specific task they already do every day and find slow or painful like drafting an investment memo, synthesizing an earnings call, running portfolio attribution. When a deployed tool can't do those things on day one, employees don't use it. ROI is impossible to measure and the tool gets shelved.
S&P Global's 2025 survey of over 1,000 enterprises found 42% of companies abandoned most AI initiatives during the year, up from 17% in 2024. The average organization scrapped 46% of proofs-of-concept before production. Gartner estimates more than 40% of agentic AI projects will be cancelled by 2027, largely because organizations chased the technology rather than a specific business problem. MIT adds that enterprises report the lowest pilot-to-production conversion rates of any organization type and scale considerably slower than mid-market peers when they do move.
Providing a license is not the same as achieving adoption. Adoption means analysts complete their most important workflows faster, produce better output, and know it within the first week.
A large asset manager deploys a general research assistant to 200 analysts. After 90 days, 15 are using it consistently. The other 185 found it couldn't handle their query types with sector-specific language, internal shorthand, proprietary frameworks so they went back to doing things the way they always had.
The second root cause is structural, and in some ways harder to fix than the first.
Investment analysts think in terms of theses, alpha, IC formats, and coverage universes. Engineers think in terms of APIs, retrieval pipelines, prompt engineering, and latency. When these two groups collaborate on an AI build, the result is never what either expected. The engineer delivers exactly what was specified. The analyst discovers the specification missed 80% of what actually matters about how they work.
Neither is doing anything inherently wrong. Firms failing at adoption consistently cite the gap between technical teams and investment professionals as a key obstacle. As one firm said, "If you forget about the human element and don't take people on the data and AI journey, you will never have adoption."
The pattern is predictable. The AI project gets owned by the technology function. The investment team shows up for an initial requirements meeting and again for the demo. The actual build happens in between, in isolation from the workflow it's supposed to serve. When the tool ships, it solves the workflow the engineer imagined, not the one the analyst runs.
Firms reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting modeling techniques. They start with the work, not the technology.
In investment management, the divide is worse than in most industries. The vocabulary is specialized. The output formats like IC memos, attribution tables, or sector screens are firm-specific. An AI system built without someone who actually understands buy-side research will not perform at the level analysts need, no matter how capable the underlying model is.
This is the one firms consistently underestimate and the one that does the most damage, because firms spend months or years on this only to realize they failed after pouring millions into it.
Financial institutions run on data that is spread across incompatible systems, owned by different teams, and sourced from a dozen vendors. A single research workflow at an asset manager might pull from internal investment memos spanning a decade, earnings transcripts from multiple vendors, broker research from 15 sell-side relationships, Bloomberg terminal data, FactSet estimates, proprietary deal models in Excel, and email threads where the actual thesis evolved. That's before touching the deal room.
Generic AI tools handle all of this the same way. They do vector search across a flat document store. For institutional investment work, that's not close to sufficient.
A working AI system for investment research has to know which version of an internal model is current. It has to treat a broker research note differently than an internal IC memo. It has to know when to pull live market data versus archived filings, and flag when a 2021 deal memo's assumptions no longer hold against current sector comps. None of that comes standard. It has to be built, deliberately, for how the firm actually works.
The data problem shows up in the numbers consistently. Informatica's CDO Insights 2025 survey found data quality and readiness was the top barrier to AI success, cited by 43% of enterprise data leaders. A Forrester study commissioned by Capital One found 73% of enterprise data leaders ranked data quality above model accuracy, computing costs, and talent shortages as their primary obstacle. Fivetran's 2026 benchmark found that 53% of engineering time at most enterprises goes to pipeline maintenance, and 97% report disruptions to AI initiatives from data infrastructure issues.
In financial services, the problem is compounded by security. Institutional data cannot move through systems with standard data retention policies. Zero-data-retention architecture isn't optional. Any serious deployment requires it from day one, not as an afterthought once a pilot has already run on the wrong infrastructure.
Firms that fail assume that their existing data infrastructure was good enough to build on. It rarely is. AI built on fragmented, inconsistently labeled data produces unreliable output. Analysts catch the errors quickly, stop trusting the tool, and stop using it. The adoption problem and the data problem are the same problem.
The firms that get AI right in financial services approached the problem differently from the start. Three things separate them.
Rather than deploying a general AI capability firm-wide and hoping usage emerges, successful firms identify the single workflow with the clearest ROI and build that first.
In investment management, it's usually the investment memo. Writing an IC memo is high-value, high-frequency, and standardized enough within a firm that AI can get it right quickly. An analyst who sees a draft IC memo produced in 10 minutes (one that matches the firm's format, cites the right data sources, and flags the right risk factors, all in the firms voice) understands immediately what this is worth. Adoption follows from that. It doesn't come from mandates or training sessions.
Build the thing that makes users' most important workflow faster on day one. Prove it works and let adoption scale from there.
Terminal X starts every deployment this way, identifying the single highest-ROI workflow before anything gets built, so analysts see value on day one.
Successful firms deploy former investment professionals alongside AI engineers (people who can sit with a research desk, learn how analysts actually work, identify where AI creates real leverage, and build for exactly that).
It doesn't need to be a long engagement. Firms that compress discovery and build, working on-site, learning the workflow, producing a functional firm-specific deployment in one to two weeks, get higher adoption than firms that run six-month implementation programs. The faster analysts see something real, the faster they build their process around it.
Terminal X deploys ex-Wall Street professionals alongside AI engineers on-site at client firms, typically delivering a working, firm-specific deployment within two weeks.
Successful firms treat data architecture as a prerequisite. That means getting deal documents and internal research indexed and queryable before the retrieval layer goes in. It means pulling third-party sources (Bloomberg, FactSet, broker research) directly into the AI context layer so analysts aren't manually reconciling outputs from four different tabs. Zero-data-retention architecture goes in from day one, not retrofitted after a pilot has already run on the wrong infrastructure.
Everything should be built on-premise, inside your firm's cloud infrastructure so only you have access to your internal, proprietary data.
Terminal X handles the full data layer by fusing proprietary firm data with third-party vendors, all in your cloud, at your firm, and builds how institutional research actually works.
AI adoption failure in financial services is a solved problem, just not by most firms. The constraint isn't models or capital but structural issues that generic tools and standard implementation approaches don't touch. Misdiagnosis, the context-gap between analysts and engineers, and data infrastructure that was never built for this.
The firms getting ROI start with a specific workflow, put the right people on-site, and build data infrastructure that reflects how institutional research actually works.
Terminal X deploys AI agents for institutional investment teams, built on your firm's data and designed around your workflow by professionals who have worked both sides of the desk. Request a demo