A comparison of how 12 leading banks deploy, organize and govern AI, built from public disclosures. It ends with the strategic questions I think matter most for a firmwide technology agenda.
Employee access to AI is close to universal at the leading banks; the open questions are now governance and value. Banks have converged on a governed central platform with business-owned use cases. Agentic AI is moving faster than supervisory guidance, which leaves each bank to set its own agent perimeter and its own way of proving value.
Reported employee access: BNY 99%, Morgan Stanley ~98%, Citi 87%+, Barclays ~90% on Copilot. Evident finds only 12% of use cases report impact against operational KPIs and about 1% disclose concrete financial returns.
LLM Suite, Eliza, Lumina, Claves and Capital One's enterprise platform all centralize models, data access and controls. Use cases sit with the businesses. A named AI executive is now common: UBS (Jan 2026), RBC (Feb), HSBC (Apr), Citi (Apr), Wells Fargo (Oct).
SR 26-2 (Apr 17, 2026) replaced SR 11-7 and puts generative and agentic AI outside its scope until a future RFI. Banks are filling the gap themselves. Goldman has a CIO/CRO co-chaired AI Risk Committee, and BofA reviews each project on 16 risk dimensions.
JPMorganChase and RBC block third-party AI agents from customer accounts. Morgan Stanley lets clients' agents connect over MCP. BofA argues for human authorization of agent payments. No industry standard exists yet.
14,000 Citi engineers on Devin. Hundreds of agents alongside Goldman's 12,000 engineers. About 40% of BNY code AI-authored. Claude Code targeted at half of Barclays developers by year-end. The likely next effect is on the shape of technology teams.
RBC targets C$700M–1B by 2027. BofA guides $300M of savings in 2026 rising to $700M in 2027. BNY tracks ~$450M. Capital One will "not quantify." JPMorganChase calls its earlier $2B benefit figure "stale."
Ordered by 2026 Evident AI Index rank. Open a row for leadership, platform, agentic footprint, governance and value. The operating-model labels are my synthesis of each bank's disclosures, not the banks' own terms.
Evident AI Index 2026 pillar ranks out of 50 banks (lower is better). The index weights Talent at 45%, Innovation 30%, Leadership 15% and Transparency 10%. Transparency is the pillar closest to responsible AI. Several leaders rank high on talent and innovation but low on leadership and transparency.
The dates that shape AI operating-model decisions across US, EU, UK and Canadian banking.
Developed with the FSSCC, US Treasury and more than 100 institutions, with NIST input. It is the closest thing to a shared industry baseline.
Narrower model definition, oversight scaled to risk and materiality, aimed mainly at banks above $30B in assets, nonbinding. Generative and agentic AI are explicitly out of scope; the agencies plan an RFI.
Barclays and UBS are participating in the UK regulator's live testing programme.
Canada's prudential regulator addressed gen and agentic AI ahead of Guideline E-23.
Regulation (EU) 2026/1744 delays high-risk obligations. Art. 50 transparency (chatbot disclosure, labelling) still applies from Aug 2, 2026. AI-literacy duties remain, in softer form.
Enterprise model-risk expectations for Canadian banks, including AI models.
Covers creditworthiness and credit scoring, and life and health insurance pricing: risk management, data quality, human oversight, documentation, conformity assessment.
The disclosures above point to five decisions where leading banks still differ. Each needs a clear, defensible position and a plan to carry it out.
Separate internal agents, agents acting for clients and third-party agents acting on clients' behalf. Each needs its own identity, authorization and liability model. Blocking outside agents works today, but it may not hold once clients expect their own agents to transact.
SR 26-2 leaves the gap open. Leaders are writing their own controls, such as agent identity and access, post-deployment monitoring and human sign-off for high-risk actions. Getting ahead of the RFI lets a bank help shape it rather than react to it.
Options include model-agnostic routing (JPMorganChase, UBS, BNY), proprietary or open-weight models (Capital One, RBC's ATOM) and a single strategic partner (Morgan Stanley with OpenAI). Each choice affects cost, concentration risk and how much of the stack the bank controls.
Moving from counting use cases and users to outcomes attributable to the P&L means agreeing on baselines with Finance. Evident adds an outcomes pillar in 2027, so the market will soon compare banks on it.
If engineering is the first workload to scale with agents, the mix of junior and senior engineers, vendor spend and team structure will change first, and the strategy needs to plan for that.
Researched and built in one session using agentic AI tools: parallel research agents read and cross-checked the sources below, and the page was generated and checked in code. Figures are as disclosed by the banks or reported by the sources, are self-reported and unaudited, and carry the dates noted. Several bank profiles draw on BankingNewsAI's per-bank summaries of earnings calls, filings and press. Where counts conflicted, I used the most recent dated figure. This brief is not affiliated with or endorsed by any bank named.