The first time a junior analyst at a mid-size US asset manager pushed an AI-generated portfolio tilt into an investment committee meeting was probably in 2023. By 2025 it was routine. Mercer’s 2024 global survey of investment managers found that 91% of managers either already use or plan to use AI within their investment strategy or asset-class research — 54% already using it, another 37% planning to. That is not a pilot project anymore. It is the new default for how decisions get framed inside the buy side. The applied AI in finance market sat at $14.82 billion in 2025 according to Precedence Research, with North America capturing 39% of that spend and banking alone absorbing nearly half.
How AI moved from the research desk to the decision table
Inside most asset managers, quantitative research teams had been running machine learning models since the mid-2010s. What changed between 2022 and 2025 was not the existence of the techniques — it was who gets to use them, and for what. Traditional quant shops kept ML inside a walled garden of factor models and signal generation. The new wave of deployments pushes model output directly into the fundamental analyst’s workflow, the portfolio manager’s pre-trade checklist, and the CIO’s capital allocation review.
The shift happened in three waves. The first wave, which ran from roughly 2020 to 2022, was alternative data. Hedge funds and systematic managers stacked credit-card panels, satellite imagery, shipping data, and web-scraped text on top of their price-and-fundamental stacks. The second wave, starting in 2023, was generative AI. Language models made it cheap to read thousands of earnings transcripts, rating agency reports, and regulatory filings in hours instead of weeks. The third wave, visible through 2025, is decision support — systems that do not just surface information but propose specific portfolio actions with explainable reasoning, which a human decision-maker then accepts, modifies, or rejects.
Mercer’s survey captures the hand-off cleanly. Among firms with AI already integrated, more than half report that AI analysis informs rather than determines the final decision, while about a fifth report that AI proposes decisions that investment teams can override. That is the production-grade workflow most institutions have landed on: AI in the loop, human above the line.
What the AI-in-finance market looks like in 2025
| Metric | Value | Source |
|---|---|---|
| Applied AI in finance market, 2025 | $14.82 billion | Precedence Research |
| Projected market size, 2035 | $92.53 billion | Precedence Research |
| Forecast CAGR, 2026–2035 | 20.10% | Precedence Research |
| North America share, 2025 | 39% | Precedence Research |
| Banking share of end-use, 2025 | 48% | Precedence Research |
| Machine learning share of technology spend | 43% | Precedence Research |
| Fraud detection share of applications | 32% | Precedence Research |
| Asset managers using or planning AI in strategy | 91% | Mercer 2024 survey |
The Precedence data also shows that fraud detection remains the single largest application bucket at 32% of spend, with risk management, algorithmic trading, customer service, and credit scoring splitting the rest. The headline for decision-making, though, is not the fraud line — it is that algorithmic trading and risk management together represent a substantial and still-growing slice of the same pie, and that machine learning captures 43% of technology spend, more than NLP, robotic process automation, and computer vision combined.
Five decision jobs AI actually does inside institutions
Strip away the marketing and AI-for-decisions clusters into five concrete jobs inside banks, asset managers, and fintechs.
The first is capital allocation. Corporate treasury teams and institutional asset allocators use ML models to project cash flow, stress-test balance sheets, and allocate across asset classes under volatility scenarios. The models do not replace the CFO or the CIO — they produce the scenario analysis that used to take a week of analyst work in an afternoon.
The second is portfolio construction. This is where the Mercer numbers land hardest. AI-assisted factor construction, optimization under turnover constraints, and ML-driven rebalancing have become standard across systematic and multi-asset teams. The CFA Institute’s 2025 monograph on AI in asset management documents practitioners using ML for both alpha generation and risk decomposition, and the pattern that has emerged is hybrid portfolios where ML-derived signals sit alongside traditional factor models.
The third is credit underwriting and risk decisioning. This is where decision-making AI overlaps most directly with the machine learning systems US banks have deployed for credit-scoring and model-risk management. Middle-market lenders use gradient-boosted models on top of cash-flow data to underwrite small businesses in minutes instead of days. The decisions are still reviewed by humans for anything above a dollar threshold, but the first-pass rank-ordering of applicants is now almost entirely model-driven.
The fourth is trade execution. Algorithmic execution has existed for two decades, but the latest generation of AI-driven execution models optimizes across venues, order types, and time-of-day patterns in ways that reduce slippage on large institutional orders. Sell-side desks have invested heavily in proprietary execution algorithms because small slippage improvements compound into meaningful alpha at scale.
The fifth is alternative-data signal generation. This is the area that overlaps with the sentiment analysis systems US traders and fintechs use to turn text into tradeable signal. Processed sentiment, geolocation data, and transaction-level consumer data all feed decision-support dashboards that portfolio managers consult daily. The signals rarely dictate a trade on their own, but they shift confidence levels on theses that humans are already building.
The vendor and deployment map
The decision-making AI vendor map divides cleanly into three layers. At the platform layer, BlackRock’s Aladdin has embedded AI features across its risk analytics and portfolio construction workflows, and State Street’s Alpha platform has done the same for asset-servicing clients. Bloomberg and FactSet have added AI-driven screens, summarization, and factor analysis to their terminals. These incumbents are not going anywhere — the buying pattern is that asset managers already paying for the terminal or the platform prefer to add the AI layer there rather than buy a new point solution.
At the application layer, a handful of AI-native entrants have carved out specific workflows. Kensho (now part of S&P), Hebbia, Arkifi, and several newer research-automation vendors compete for the “summarize, extract, and explain” workflows that sit between raw data and a PM’s decision. Each has built defensibility around specific document types or buy-side client segments.
At the infrastructure layer, the same pattern visible across other AI-in-finance categories holds: AWS Bedrock, Azure OpenAI Service, and Google Vertex AI dominate cloud-hosted inference, while Meta’s Llama and Mistral are common for on-premises deployments where model weights need to stay inside the firm.
What the regulators are watching
Financial regulators treat AI-assisted decision-making as a category of model risk, and the supervisory stance has sharpened through 2024 and 2025. The Federal Reserve, the OCC, and the SEC have all emphasised that the existing SR 11-7 model risk management framework applies to AI-driven investment decisions, credit decisions, and capital allocation decisions without exception. The supervisory interest is loudest in three areas: how firms validate models before deployment, how they monitor drift once in production, and how they document the human-in-the-loop controls that prevent automated decisions from compounding systemic risk.
The overlap with anti-money-laundering compliance systems and the model-governance controls US fintechs have been building is direct. A decision-support model that cannot explain why it recommended a portfolio tilt or a loan denial is a model that cannot pass supervisory review. This is why the vendors selling into regulated buyers have all built explanation layers, feature-importance dashboards, and documentation tooling as first-class product capabilities rather than afterthoughts.
What it means for fintechs and operators
For founders, the decision-making AI opportunity sits in narrow, high-value workflows rather than in a horizontal chatbot. Credit-underwriting automation for niche asset classes, treasury-optimization tools for mid-market CFOs, and research-automation tools for specific buy-side segments are all categories where focused startups have beaten horizontal vendors on accuracy and speed. The horizontal plays — “let AI make investment decisions” — have mostly struggled to sell into institutional buyers, because the buyer’s risk team requires explainability and traceability that a black-box model cannot supply.
For operators at asset managers and banks, the governance workload is real. Model risk teams have had to expand their remit to cover generative AI components, alternative-data pipelines, and the human-review protocols that sit on top of automated decisions. The firms that treated this as an operational-excellence problem in 2024 are the ones shipping cleanly in 2026. The firms that assumed the existing model risk framework would absorb AI-driven decisions without changes are still catching up.
The bottom line
AI-assisted decision-making is no longer optional inside large asset managers and banks. A $14.82 billion market growing at 20% a year is the revenue proof. Mercer’s 91% adoption figure is the behavioural proof. The firms still extracting value from these deployments are the ones that treated AI as a decision-support layer under human oversight, not as an autonomous decision system. The ones that tried to skip the oversight layer are the ones re-writing their model governance frameworks now.
