The Day Wall Street Met Its New Operating System: Inside Anthropic’s Briefing for Financial Services
--
Dario Amodei and Jamie Dimon on stage. Ten new AI agents for banking. A capability curve that is not slowing down. Here is everything that happened.
By Luis Lopez | [email protected]
The following is a summary based on virtual attendance of Anthropic’s “The Briefing: Financial Services,” a livestreamed event held on May 5, 2026. The event was hosted live in New York City and broadcast globally to financial services executives. This piece draws on the full event transcript, presentation slides, and publicly available coverage.
One year ago, Anthropic made a promise to financial services executives: AI would fundamentally change how their people work, how their businesses operate, and what products their organizations could build. On the morning of May 5, 2026, live from New York City, that promise was cashed. In a rare gathering of senior leaders from banking, insurance, and capital markets, Anthropic brought together its own leadership, the Chairman and CEO of JPMorganChase, the CIO of Goldman Sachs, the CIO of JPMorganChase, and the CEO of AIG for a single, focused conversation. It was not a product pitch. It was a briefing on where the AI transformation in finance actually stands, and where it is headed.
“Is This Hype or Is It Real?” Dario Amodei and Jamie Dimon Answer the Question Every Board Is Asking
The event opened with a moment that would have been unthinkable eighteen months ago: Anthropic CEO Dario Amodei and JPMorganChase Chairman and CEO Jamie Dimon sharing a stage, moderated by CNBC anchor and New York Times DealBook editor Andrew Ross Sorkin. The message from both was unambiguous. The question that every board has been asking is now settled. This is real.
Dimon offered a window into how deeply embedded AI already is at JPMorgan. Just the weekend before the event, he used Claude Code personally, prompting it as a bank CEO to produce a deep analysis of asset swaps, treasury basis spreads, liquidity conditions, and investment-grade risk. “In 20 minutes it created this huge dashboard with all the backup and all the research, and it was very accurate about what I wanted,” he said. He noted that the bank first started using AI in 2012, initially for pattern recognition in mortgage data. “I call it more advanced math,” he said. “It was like machine learning where we notice patterns and run through lots of data that a human being couldn’t do.” Since then, JPMorgan has built hundreds of use cases spanning risk, fraud, marketing, document review, and customer personalization. His view on what comes next: “There’ll be second-order effects we haven’t thought of yet.”
Amodei offered a nuanced forecast. While the exponential curve of model capability has held accurately, he acknowledged the inherent unpredictability of how technology manifests in the real world. What he is confident about is clear: “Those error rates are going to go down. We’re going to see also more agency, more autonomy, ability to do things end to end.” Tasks that once required a team working for a week or two will increasingly be completed without human intervention.
The PE Joint Venture, Cybersecurity, and the Workforce Question
Sorkin pressed both leaders on three uncomfortable topics: the private equity joint venture announced the day before, cybersecurity risks, and the workforce implications of automation at scale.
On the joint venture, a collaboration with Blackstone and other PE firms to embed Claude directly into portfolio companies, Amodei was transparent about the motivation. Anthropic, with roughly 3,500 employees and a go-to-market team of several hundred, cannot match the 50,000-person sales forces of legacy enterprise software vendors. The PE channel offers distribution at a scale that organic growth cannot replicate. The bottleneck to AI’s broader impact, he argued, is not model capability. It is the speed of diffusion through large organizations. The joint venture is a structural answer to that problem.
On cybersecurity, Dimon was blunt. “Cyber is our biggest risk. It’s been our biggest risk for years. And it’ll be made worse by AI,” he said, referencing a section of his annual chairman’s letter. Amodei detailed the escalation. A year ago, Claude was used to review code for security vulnerabilities with modest results. Six months ago, Anthropic identified the first attempts by Chinese state-linked actors to conduct cyberattacks using Claude. Three months before the event, Anthropic used the model prior to Mythos to find what Amodei described as “something like 20 vulnerabilities” in the Firefox codebase, disclosing them publicly in coordination with Mozilla. With Claude Mythos, that number grew to almost 300. Amodei noted that Anthropic has identified tens of thousands of vulnerabilities in closed-door testing, the majority not yet announced because they have not yet been patched. He confirmed meeting directly with Treasury Secretary Scott Bessent on the issue. Both leaders agreed the window to address this is roughly six to twelve months, the estimated lead time before Chinese models close the capability gap.
On the workforce question, both leaders rejected the binary framing of replacement versus preservation. Amodei described the company preference plainly: “We try and push people towards, hey, why not do more with the same amount of resources?” Dimon echoed the sentiment, noting JPMorgan’s long-standing practice of redeployment over redundancy, while acknowledging that competitive pressure from AI-proficient organizations will not wait for gradual adoption.
Financial Services Is Now Anthropic’s Second-Largest Industry Vertical
When Chief Commercial Officer Paul Smith took the stage following the fireside chat, he grounded the conversation in numbers. A chart showing Anthropic’s enterprise revenue growth from September 2024 through April 2026 depicted a curve that went from flat to near-vertical in roughly twelve months. Financial services now ranks as Anthropic’s second-largest industry by enterprise revenue, a position it has held every month for twenty consecutive months. Forty percent of Anthropic’s top fifty customers are financial institutions. The first dedicated financial services vertical solution launched in July 2025.
What struck Smith was not managed growth but organic adoption. A snapshot taken the prior week showed the full map of workflows already running on Claude across the financial services industry: loan origination, covenant extraction, and financial spreading in lending and credit; KYC screening, SAR narratives, trade surveillance, and sanctions screening in risk and compliance; pitchbook assembly, earnings synthesis, IC memos, and portfolio commentary in investing and research; claims FNOL, underwriting, and loss run analysis in insurance; client onboarding, NAV reconciliation, and advisor support in client operations. “This is a Cambrian explosion of creativity and productivity across the organization,” Smith said.
Three customer deployments anchored the scale of what is already in production. Allianz has a target rollout of Claude Cowork to more than 50,000 employees. BNY has 220 AI solutions currently running in production. Moody’s has reduced credit memo preparation from forty hours to two minutes using Claude-powered agents.
“Coding Has Changed Forever. Finance Is Next.”
Lisa, a research product management leader at Anthropic who works at the intersection of customer deployment and model improvement, presented the section that carries the most weight for any senior leader evaluating AI strategy. Drawing a direct parallel between what AI did to software engineering and what it is now doing to finance, she made the case with two charts.
The first, titled “Scaling laws are real,” mapped the evolution of Claude models against what the model can actually do in practice, not just benchmark scores. Claude 2, launched in July 2023, answered questions. Claude Sonnet 3.5 did first-pass work. Claude Opus 4 owned defined tasks. Claude Opus 4.5 began contributing judgment. Claude Opus 4.7, released in April 2026, is operating at senior analyst level in finance. Above it on the curve sits Claude Mythos Preview, already capable of driving end-to-end outcomes.
The second chart, titled “Coding has changed forever. Finance is next.,” plotted Claude’s finance-specific capability against a projected roadmap. The trajectory from today’s inflection point moves toward VP-level finance work, then toward autonomous teams of engineers, and ultimately toward what the slide labels “Autonomous CxO.” Less than a year ago, she noted, “Claude could barely format a table without errors. And today it’s doing senior analyst level work, and we expect it to improve rapidly from here.”
The autonomy claim was backed by a concrete example. An AI finance analyst agent, given no playbook, was set the task of building weekly price forecasters for energy markets from scratch, reading news, pulling data, and testing multiple modeling approaches autonomously. Scored weekly against actual outcomes, the agent’s error rate fell from 0.27 in week one to 0.09 by week ten, beating both the LEER state-of-the-art benchmark and the DNN industry average. The agent built and improved its own forecasters without instruction.
Two safety metrics addressed the skepticism that any compliance-minded reader would reasonably bring to these claims. Claude Opus 4.7 is five times more resistant to prompt injection than the next-best frontier model, meaning a counterparty document cannot quietly rewrite the agent’s instructions. Across one model generation, the hallucination rate on impossible tasks has dropped by a factor of three. As Lisa put it: “When it can’t do the analysis, it says so. It doesn’t make up a number.”
The flywheel that sustains this trajectory was explained with unusual transparency: customer pain points become training targets. Research ships improvements as new Claude capabilities. Products get deployed in real teams. Real deployments generate real failure signals. Those signals feed back into research. Anthropic is not training on customer data, but it is working with financial institutions to identify which problems are worth solving next.
The Claude Thinking Engine: A Full Platform, Not Just a Model
Nick Lin, who leads financial services product at Anthropic and describes himself as “a recovering investment banker and private equity investor,” presented the architecture that translates the capability curve into deployable products.
The framing is a single thinking engine with three layers. At the base sit the models: Opus for industry-leading intelligence, Sonnet for general-purpose work, and Haiku for cost-effective tasks. On top of these models sits a platform comprising APIs, tools, skills, connectors, managed agent infrastructure, and a full operating system for deployment, governance, and monitoring at scale. On top of the platform, Anthropic and its customers build the outcomes: smarter employees through Chat, Code, and Cowork; faster processes through agents running core workflows end to end; and transformative products built directly into customer-facing experiences.
Every agent is built from three components. Skills encode the institutional knowledge of how a task is done at a specific firm. Connectors are the data sources the agent can query, including Factset, LSEG, Moody’s, MSCI, S&P Global, Morningstar, PitchBook, and Verisk, plus eight newly announced partners: Dun and Bradstreet, Fiscal.ai, GLG, Guidepoint, IBISWorld, SS&C, and Third Bridge. Subagents handle discrete sub-tasks such as data fetching and variance flagging. These three building blocks combine to produce a deployable agent template.
The ten agent templates announced at the event cover the workflows that consume the most analyst hours in the industry. Five are designed to make employees more effective: a pitch builder that assembles the book, a meeting prep agent that briefs bankers before client meetings, an earnings reviewer that reads the print, a model builder that spreads and projects financials, and a market researcher that conducts sector analysis. Five are designed to run processes autonomously: a valuation reviewer, a general ledger reconciler, a statement auditor, a KYC screener, and a month-end closer.
The demo that brought this to life followed Sarah, an Executive Director in natural resources coverage at a fictional investment bank. She receives an urgent take-private inquiry from a client during an industry dinner. In the old workflow, her analysts lose the weekend. With Claude, she forwards the email via voice mode. Claude activates the Pitch Builder plugin, pre-wired to S&P, Factset, and LSEG, pulls the firm’s comps analysis skill, builds the LBO model and comparable companies analysis in Excel, assembles a first draft of the pitch deck with interactive assumption sliders, drafts the client email in Outlook, and delivers a completed package before she gets home. Every banker in the room recognized the scenario.
On the Microsoft 365 integration, Lin confirmed that Claude is now generally available across Excel, PowerPoint, and Word, with Claude for Outlook launching simultaneously in beta. The same agent carries full context across all four applications in a single coherent workflow. Citadel’s Head of Core Engineering, Atte Lahtiranta, offered this assessment: “Our investment professionals live in data and analytical models. Claude for Excel meets them there and gives them a step-change in how fast they can build, test, and trust their work.”
For institutions that want agents running inside their systems rather than alongside their people, Claude Managed Agents provide the necessary infrastructure: long-running and autonomous, connected to existing technology stacks, fully configurable to institutional processes, and production-ready with Anthropic managing the operational layer.
The C-Suite Panel: Goldman Sachs, JPMorgan, AIG on What Actually Matters
The closing panel brought together three of the most senior technology and operations leaders in global finance: Marco Argenti, Chief Information Officer of Goldman Sachs; Lori Beer, Chief Information Officer of JPMorganChase; and Peter Zaffino, Chief Executive Officer of AIG.
The discussion on ROI surfaced a point that none of the three treated as genuinely contested. The foundational efficiency use cases — productivity, code, operations — are no longer a debate worth having. The harder question is how to measure value as AI moves from internal metrics to client-visible outcomes. “A year from now, we will see success or AI will be seen from our client side,” said Argenti. “Our clients will clearly see that we’re leading on AI. That will transform the way our clients actually even interact with us.”
Beer was direct about the lessons she would apply differently. JPMorgan had traditional AI deeply embedded before the current generation of models arrived. The gap was not in adoption but in speed of organizational rewiring. “Rewiring faster,” she said. “Governance, the way we work, I would do that differently now sitting from this chair, right now, in this moment.”
Zaffino offered a warning against institutional hesitation. “Getting stifled by return on investment when you’re going to have to compete in a world that’s going to be very proficient with large language models,” he said, was the risk he was most focused on avoiding. His goal for the year ahead: enterprise-wide rollout, and being able to say twelve months from now that AIG is materially better run because of it.
The Briefing Has Ended. The Work Has Not.
What Anthropic staged on May 5, 2026 was not a product launch. It was a marker. The institutions in that room and the many thousands watching the livestream heard from credible voices in both AI and finance that the transition from pilot to infrastructure is complete. The agents that close books at month-end, draft credit memos, run KYC, and analyze risk are not prototypes. They are in production. The question is no longer whether this technology meets the industry’s demands for precision, compliance, and oversight. It does. The question is now organizational: how fast can institutions rewire their governance, redeploy their people, and build the internal capability to keep pace with a capability curve that shows no sign of slowing. The briefing is over. The work has just begun.
Contact: [email protected]
Sources
Anthropic, “The Briefing: Financial Services” — Event page and livestream, May 5, 2026. https://www.anthropic.com/events/the-briefing-financial-services-virtual-event
Event transcript and presentation materials, Anthropic, May 5, 2026.
This summary was produced with the assistance of Claude Sonnet 4.6 (Anthropic).
Tags: Artificial Intelligence, Financial Services, Banking, Fintech, AI Strategy, JPMorgan, Goldman Sachs, Anthropic, Claude, Enterprise AI
