Banks went from barely mentioning AI to almost all of them mentioning it in five years. Their efficiency ratios didn't follow the same curve.

Direct answer

McKinsey describes generative AI adoption as the fastest technology rollout on record, with a projection that early-adopter banks "could see" a 4-point improvement in Return on Tangible Equity. [C01] Testing the disclosure side of that claim directly: the share of 58 reviewed public bank holding companies whose 10-K mentions AI-related terms rose from 5.2% in 2020 to 86.2% in 2025, with the steepest jump between 2023 and 2024. [C02] But a direct read of the actual filing text behind that curve found most of it is risk-factor language treating AI as one of several "emerging technology" risks, not a claim of realized deployment. [C03] And the regulatory efficiency ratio — the actual financial-statement metric AI adoption is supposed to move — shows no cohort-wide improvement timed to that disclosure surge: the clearest broad-based efficiency gain across large-bank cohorts happened in 2022, before the surge, and 2023 saw efficiency broadly worsen even as survey-reported adoption was already rising. [C04, C05]

Key findings

The disclosure curve is real and dramatic

Searching each institution's 10-K filings for a disclosed five-term AI vocabulary ("artificial intelligence," "machine learning," "generative AI," "large language model," "GenAI"), the share of institutions with at least one mention climbed steadily and then sharply: 5.2% in 2020, 8.6% in 2021, 15.5% in 2022, 17.2% in 2023 — then a jump to 56.9% in 2024 and 86.2% in 2025. [C02] That timing lines up with the broader post-ChatGPT generative-AI wave: 10-Ks filed in 2024 (for fiscal 2023) and 2025 (for fiscal 2024) are the first annual reports where most banks would have had a full year of generative-AI-specific developments to discuss.

What the mentions actually say

A rising mention count is not the same as a rising count of real deployments. Reading the actual text behind three example hits:

• Capital One's 2020 10-K groups AI and machine learning with "distributed ledger technologies" inside a regulatory-compliance risk factor: adopting new technologies "can present unforeseen challenges in applying and relying on existing compliance systems." This is a risk disclosure about regulatory uncertainty, not a claim that Capital One had deployed AI operationally. [C03]
• Fulton Financial's 2025 10-K discloses "our implementation of certain new technologies, such as those related to artificial intelligence, machine learning and automated decision making, in our business processes" — genuinely acknowledging real implementation, but framed entirely as an operational-risk caveat about "unintended consequences," not an efficiency claim. [C03]
• WSFS Financial's 2025 10-K is the exception: a strategic disclosure listing "the deployment of artificial intelligence, and predictive modeling to create operational efficiencies and redesign business models" as part of its own stated initiatives — the one example that actually claims what the survey narrative describes. [C03]

One bank's filing said what the AI-efficiency narrative claims. Two others mentioned AI only as a source of risk to manage. A rising mention count blends both — and cannot tell them apart without reading the text.

The efficiency ratio doesn't show a matching curve

If the survey-reported AI-adoption acceleration were already producing broad, measurable efficiency gains, the population-level efficiency ratio — noninterest expense over revenue, where lower is better — should show its own inflection roughly where the disclosure curve inflects, in 2024-2025. Reusing OptimaYield's already-computed cohort medians:

The clearest, most broad-based improvement across all three cohorts happened in 2022 — two years before the AI-disclosure surge began. [C04] 2023, a year every survey source already described as one of rising AI adoption, instead shows the median efficiency ratio getting worse at all three cohorts. [C04] 2024 and 2025 — the years disclosure actually surged — show real improvement at the two largest cohorts, but the $10-50B cohort still worsened in 2024 before improving in 2025. [C05] There is no clean, population-wide inflection lined up with the disclosure timeline in either direction.

Methodology

AI-disclosure side: reused the reviewed-CIK population and EDGAR full-text-search client (src/data-platform/edgar-fulltext-search.ts) built for the vendor-concentration report — 58 of 159 $10B+ institutions with a reviewed public-holding-company SEC_CIK link. Queried each institution's CIK for a combined, disclosed five-term AI vocabulary, restricted to Form 10-K, for filings dated within each calendar year 2020-2025. Efficiency-ratio side: reused OptimaYield's already-computed efficiency_ratio cohort percentiles in semantic.peer_statistics (2016-2026, no new derivation) for the three largest asset cohorts. Example AI-term hits were spot-checked by fetching the actual filing text directly. Live query: scripts/analysis-ai-adoption-efficiency.ts.

Limitations and counter-evidence

• This is a time-series comparison, not a matched correlation. [C06] The AI-disclosure population (58 institutions) and the efficiency-ratio population (up to 159 institutions) are different sets; no claim is made that specific banks in one dataset drove the pattern in the other.
• "Filed in year Y" is a disclosed proxy for fiscal year Y or Y-1, not exact fiscal-year attribution. [C07]
• The AI vocabulary and Form 10-K restriction are disclosed, non-exhaustive choices. [C08] A broader vocabulary or additional filing types could show a different curve.
• This cannot test individual banks' actual AI-driven outcomes, nor rule out that AI is improving a different metric (revenue growth, headcount, customer experience) not measured here. [C09]
• A negative population-level result is not proof AI has no effect. It is evidence that, so far, no broad, population-wide efficiency-ratio inflection is visible at the timing the disclosure surge and survey narrative would predict — a meaningfully different, and more testable, claim than "AI doesn't work."

Related on OptimaYield

• Fiserv shows up in bank SEC filings more than every other named core-processing vendor combined — the same EDGAR full-text-search infrastructure and the same discipline of spot-checking what a filing "mention" actually means before trusting the raw count.

Official sources

• McKinsey & Company, "Global Banking Annual Review 2026," 2026.