Agentic AI is only beginning to surface in U.S. finance firms' annual reports—0.4% of firms in 2024 and 1.6% in 2025—but those that mention agentic systems tend to surround them with denser governance and safety language, implying deployments are currently concentrated where governance is more mature.

Main Finding

Agentic AI references in U.S. financial firms’ annual filings were effectively absent through 2023, emerged in 2024 (0.4% of firm‑years) and rose in 2025 (1.6% of firm‑years). When firms do mention agentic AI, explicit evidence of autonomy is rare but tends to appear in disclosure passages with higher “controls density,” consistent with governance maturity being a prerequisite for deploying action‑taking systems.

Key Points

• Scope: Balanced panel of 2,500 firm–year observations (500 publicly listed U.S. finance firms per year) for 2021–2025.
• Measurement approach: Text-as-data using an auditable dictionary-and-context method to flag agentic references and measure local governance/safety language.
• Temporal pattern: No agentic mentions in 2021–2023; 0.4% of firm‑years in 2024; 1.6% in 2025 — a late but accelerating diffusion signal.
• Autonomy vs. mention: Most agentic mentions do not contain clear autonomy evidence; when autonomy signals appear, they are concentrated in disclosure windows with higher controls density.
• Transparency: Emphasis on an auditable, reproducible measurement framework that can provide baseline statistics for ongoing monitoring.

Data & Methods

• Data: Annual filings (presumably 10‑K / equivalent) from 500 U.S. finance firms each year, 2021–2025, forming a balanced panel of 2,500 firm–year records.
• Identification: Dictionary-based lexicon of agentic terms/phrases applied to filings to flag potential mentions.
• Contextual validation: A context window around flagged terms is analyzed to reduce false positives and to detect indicators of autonomy (action-taking capability).
• Controls density metric: Quantifies presence of governance/safety language (e.g., risk controls, oversight, monitoring, testing, human-in-the-loop) within the same local disclosure window to measure governance maturity around agentic references.
• Auditability: Method framed to be transparent and reproducible (dictionary + contextual rules rather than opaque ML black boxes).

Implications for AI Economics

• Diffusion dynamics: Agentic AI adoption in finance appears late relative to generative AI discussion but is accelerating — useful baseline for modeling adoption curves and forecasting operational risk exposure.
• Governance as a barrier/precursor: Higher controls density around autonomy signals suggests governance investments are correlated with (or necessary for) deploying agentic systems; models of adoption should account for regulatory/compliance readiness costs.
• Risk measurement & supervision: Public filings provide a feasible monitoring signal for regulators and market participants to track the transition from content-generation to action-oriented AI; the auditable approach supports oversight and longitudinal studies.
• Strategic behavior & disclosure incentives: Low incidence of agentic mentions may reflect genuine limited deployment, disclosure lags, or strategic underreporting — relevant for interpretation of market risk and informational asymmetries.
• Research agenda: The framework enables time-series and cross-sectional analyses of who adopts agentic systems, how governance practices evolve, and how deployments affect firm performance, operational loss events, and systemic risk.