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 (systems that can act, not just generate content) is just beginning to appear in U.S. financial firms’ public disclosures and remains rare through 2025, but adoption shows an accelerating pattern: agentic mentions are absent in 2021–2023, appear in 2024 (0.4% of firm‑years) and rise in 2025 (1.6% of firm‑years). When firms do disclose autonomy-positive language, it disproportionately co-occurs with dense governance/control language — supporting a “governance‑bottleneck” view that credible agentic deployment in regulated finance tends to require mature control frameworks.

Key Points

• Agentic ≠ generative: the paper defines agentic AI as goal-directed systems integrated into workflows with delegated action/execution rights (planning, tool use, triggering downstream tasks), not merely text generation.
• Conservative, auditable measurement: the authors use a dictionary-and-context (anchor/autonomy/controls) approach designed to minimize false positives and distinguish symbolic AI mentions from plausible operational agentic claims.
• Diffusion pattern: no agentic disclosures 2021–23; emergence in 2024 and modest acceleration in 2025 — consistent with a late but rising diffusion phase rather than widespread current deployment.
• Governance link: autonomy-positive cases are concentrated in passages with higher “controls density” (stronger governance/safety language), consistent with the idea that organizations only grant execution rights where control infrastructure exists.
• Limitations acknowledged: disclosure language is strategic and imperfect; the number of autonomy-positive cases is intentionally small; results are descriptive and not causal.

Data & Methods

• Data: Balanced panel of 500 publicly listed U.S. financial institutions observed each year 2021–2025 (2,500 firm‑year Form 10‑K filings). Firms had to file annually for the entire window; SIC codes used to define finance firms.
• Text preprocessing: Full narrative text retained (not restricted to one item), HTML/tables/boilerplate removed; parsing to plain text to avoid spurious matches.
• Measurement design:
  • Three domainadapted keyword dictionaries:
    • Anchor dictionary to locate candidate passages discussing agentic/system topics (e.g., “agentic”, “LLM”, “workflow automation”, “copilot”).
    • Autonomy dictionary to flag execution/orchestration/action language (e.g., “execute”, “trigger”, “orchestrate”, “function-calling”, “straight-through processing”).
    • Controls dictionary to capture governance/safeguard language (e.g., “approvals”, “audit logs”, “monitoring”, “access controls”, “model risk governance”).
  • Conservative matching rules (case-insensitive, word/phrase boundaries, plural/tense variants).
  • Anchored extraction: agentic/autonomy signals must appear in context windows identified by anchors; controls density computed as intensity of controls language within the same local disclosure window.
• Analysis:
  • Time-series counts and rates of agentic mentions by year.
  • Within-anchor analyses evaluating co-occurrence between autonomy-positive language and controls density.
  • Robustness checks for rare events and sensitivity to dictionary choices.
• Limitations of method: public disclosure is a noisy and strategic signal; small number of positive cases limits inference; the association between autonomy and controls is correlational.

Implications for AI Economics

• Measurement advance: provides a transparent, replicable framework and baseline statistics for tracking agentic adoption in regulated sectors where internal deployment data are private. This helps separate hype from operational adoption in empirical work.
• Diffusion theory refinement: adoption of agentic AI in finance appears constrained by institutional governance capacity rather than by purely technical availability — implying that models of technology diffusion should explicitly incorporate governance/organizational costs and auditability as adoption frictions.
• Policy and regulation: regulators and supervisors should prioritize standards, tooling, and incentives for auditable controls (approvals, monitoring, logging, escalation mechanisms) because governance maturity seems to be a prerequisite for credible agentic deployment.
• Research opportunities:
  • Extend measurement to other filings, higher-frequency disclosures, or cross‑country samples to trace diffusion beyond 2025.
  • Link disclosure signals to firm outcomes (productivity, incidents, risk events, costs, employment changes) and to supervisory actions to estimate economic impacts.
  • Combine public-disclosure measures with proprietary operational data (where available) to validate and refine measurement of action‑taking deployments.
  • Model adoption with explicit governance costs: endogenous trade‑off between efficiency gains from automation and investments in control frameworks; regulatory responses and systemic risk externalities.
• Caution for economists: current disclosure evidence indicates emergence not ubiquity — empirical claims about widespread agentic automation in finance should account for governance constraints and the slow, institutionally mediated nature of deployment.