The International Institute of Finance (IIF) and consultancy EY published their eighth joint global survey in 2026 on AI adoption in financial services. Covering 65 institutions across eight regions, the study finds that AI has shifted from pilot projects to live production across the sector.
Widespread live deployment and rapid adoption
According to the survey, 98 percent of respondents are using AI in live environments; the remaining 2 percent plan to deploy it in the near term. All globally systemically important banks (G‑SIBs), all insurers, capital markets participants and payment services providers reported live AI use.
The study highlights a sharp acceleration in adoption: while traditional predictive machine‑learning solutions took more than 15 years to become widespread in finance, generative AI has reached broad use in roughly four years, and agent‑based AI in about two years.
Generative AI surpasses predictive AI; agent systems grow fastest
For the first time in the survey’s history, generative AI is more prevalent in production than predictive AI: 97 percent of respondents use generative AI in production, compared with 94 percent for predictive AI. Agent‑based AI is the fastest‑growing area—the share using it in production rose from 23 percent to 70 percent in a single year.
Sectoral breakdowns show the trend clearly: all G‑SIBs deploy generative AI in production, and 77 percent of them also use agent‑based systems. Insurers reported 100 percent usage of both predictive and generative AI. Every surveyed capital markets participant is actively experimenting, with pilots heavily focused on generative and agent‑based approaches (95 and 92 percent respectively).
Investment, measurable returns and scaling
Most firms increased AI budgets between 2024 and 2026, and none plan cuts. Forty‑three percent expect budget increases of 0–25 percent; only 5 percent expect no change. AI is treated as a strategic priority broadly—North American and European firms view it as a long‑term growth factor, and 60 percent of Japanese institutions classify it similarly.
Measured returns vary by technology: 86 percent of respondents report realized financial benefits from predictive AI. For generative AI, 81 percent report measurable returns today, while expectations reach 92 percent generally and 100 percent in North America. Firms running more than 200 use cases rose from 2 percent in 2024 to 23 percent in 2026; Japan stands out with 60 percent of institutions reporting more than 200 use cases. For agent‑based AI, 32 percent currently report measurable financial results, but expectations push that to 75 percent within a year.
The share of institutions running 50–100 use cases more than doubled between 2025 and 2026.
Main use cases
Predictive AI is mainly used for risk management (66 percent), fraud prevention (58 percent) and compliance including anti‑money‑laundering (55 percent). Among institutions using generative AI in production, 98 percent apply it to internal knowledge management and 97 percent to process automation. On customer‑facing solutions, payments firms lead in chatbots; within 12 months, 84–85 percent of G‑SIBs and non‑systemic banks plan similar deployments.
Deployment barriers and tokenomics
Top deployment obstacles remain data quality issues, talent shortages and limited data access; half of respondents also cite implementation costs as a primary barrier. Although token prices have fallen on some metrics, per‑iteration token usage has increased—especially for the most advanced models. That dynamic is boosting interest in open‑source models (57 percent) and small language models (SLMs, 49 percent).
Different solution types face different hurdles: agent‑based AI struggles most with integration into legacy systems and establishing governance frameworks, while open‑source models raise data privacy and cybersecurity concerns.
Growing reliance on third parties and transparency gaps
Use of external AI infrastructure and platforms in production rose steeply from 59 percent in 2023 to 88 percent in 2026. The share of firms limited to pilots fell from 16 percent to 6 percent. External models and platforms are primarily used for data preparation, model training and performance testing, yet governance capabilities lag: about half of firms lack the information needed for effective model validation.
Among institutions using external models, 83 percent report lack of access to model architecture and technical details, 65 percent cite opaque model update and retraining processes, and 55 percent note difficulty accessing training data. This affects large players too: 69 percent of G‑SIBs said they do not have sufficient information to validate external models. On the positive side, oversight of embedded AI services has improved—firms with no governance practices fell from 13 percent to 3 percent.
Ninety‑one percent of respondents prefer a shared responsibility model when working with suppliers.
Risk management, accountability and governance
All institutions running live AI systems supervise them within formal risk‑management frameworks: 56 percent built frameworks specifically for new AI applications, 19 percent rely on existing model risk management (MRM) or enterprise risk management (ERM) systems, and 21 percent are extending current frameworks to address agent‑based AI risks.
Governance responsibility is shifting from single accountable owners toward multi‑executive, shared governance models—the share using such models rose from 9 percent in 2025 to 45 percent in 2026.
Risk profiles differ by technology: generative AI’s main concerns are data privacy and hallucinations (65 percent), while agent‑based AI raises model drift and autonomy risks (60 percent), followed by cybersecurity (52 percent) and accountability issues (45 percent). Cybersecurity risk for autonomous agents is 12 percentage points higher, reflecting the larger attack surface from unsupervised operation.
AI‑based oversight and feedback loops
AI‑based governance tools (AI‑in‑the‑loop) are spreading quickly: use of AI to automate model monitoring rose from 41 percent in 2025 to 81 percent in 2026. Deployment of feedback mechanisms for error detection and correction also increased—from 65 percent in 2024 to 81 percent in 2026—with another 14 percent currently developing such systems. Overall, 95 percent of the market is actively addressing feedback and monitoring.
In response to agent‑based AI, 62 percent of institutions are updating governance frameworks. The most common approaches are risk‑based tiering (89 percent) and case‑by‑case assessments (86 percent).
Regulation and supervisory engagement
The legal environment is becoming more complex: 71 percent of respondents said existing non‑AI‑specific rules (for example, data protection and cybersecurity) already affect their operations. Fifty‑five percent comply with sector‑specific rules, and 48 percent fall under comprehensive AI‑specific legislation. In the euro area, 78 percent of institutions are subject to comprehensive AI regulation; in North America 100 percent and in Japan 80 percent of firms reported being subject to some AI‑relevant requirements.
Regulators across jurisdictions focus on explainability, bias and ethical issues. Supervisory priority for AI model complexity rose from eighth to fifth place in a single year. While regulators’ knowledge is improving—evidenced by high rates for AI governance (88 percent), data frameworks (83 percent), predictive AI (76 percent) and cloud technology (73 percent)—firms still believe many supervisors do not fully grasp the technical details of AI lifecycles.
Engagement with authorities has intensified: the share of firms consulting financial supervisors rose to 68 percent, and contact with non‑financial authorities (e.g., data protection and consumer agencies) rose from 3 percent to 14 percent. Firms not planning any regulatory dialogue fell from 17 percent to 10 percent.
Conclusion
The 2026 IIF–EY survey shows AI in financial services has matured rapidly: predictive, generative and agent‑based systems are being deployed in production at scale, investments are increasing, and returns are becoming measurable across more use cases. The sector’s immediate challenge is no longer adoption itself but the ability to scale AI responsibly—ensuring robust controls, transparent supplier relationships, strong data and risk management frameworks, and effective regulatory dialogue. The rapid spread of agent‑based AI underlines that AI has evolved from an adjunct innovation into a foundational operational technology for finance.
Related professional event
Portfolio’s Future of Finance 2026 conference takes place on September 23; the event includes a dedicated session on AI maturity in the financial sector.
