This is outside our normal publishing schedule and a little different from the usual structure. On Tuesday, May 5th, Anthropic held its first Financial Services Briefing in New York. You can watch the full event here. Let’s talk about it.

Jamie Dimon, CEO of JPMorgan Chase, opened the morning with a story about his weekend on Claude Code. He had built himself a research dashboard on asset swaps, treasury bid-ask spreads, market liquidity, and investment grade in twenty minutes. Love a story about a prototyping CEO.

We’ll walk through four other pieces from the morning:

Anthropic next put three of its financial-services customers on stage: Allianz, Bank of New York Mellon, and Moody's.

Anthropic said Allianz is putting fifty thousand Cowork seats into the hands of underwriters, claims assessors, and claims administrators across the company.

BNY built what Paul Smith, Anthropic's chief commercial officer, called digital coworkers. These are agents that show up in the corporate address book and take assignments the way employees do.

Moody's exposed credit-rating data through Model Context Protocol (MCP, a standard for connecting services with AI agents), making the data queryable from inside other companies' agentic workflows.

The three Anthropic stories share a pattern: business leadership owned the transformation and engineering served it. McKinsey says, "delegating the agentic transformation to your technology leader, as you would with a software deployment, will not suffice." Wharton's Ethan Mollick puts the same point more bluntly. "It's not actually a technology problem. It's a process problem."

Allianz's engineering team gained productivity first. That gave the company evidence to commit to a wider, business-led transformation.

Credit unions face the same questions about which workflow to change, who on the business side owns the redesign, and what internal capacity comes together with the vendor.

Nick Lin, who leads financial-services product at Anthropic and described himself on stage as "a recovering investment banker," walked the room through ten new agent templates and eight new connectors.

To show what these templates can do, Anthropic gave Claude a single instruction. Forecast next week's energy prices. Claude figured out it should read the news, pull data, try several approaches, and check itself against actuals. Anthropic said the model beat published state-of-the-art benchmarks from week one. Three weeks of quant work compressed to two hours.

Each template runs two ways. As a plugin inside Claude Cowork or Claude Code, where a person prompts it and watches it work. Or as a Claude Managed Agent, a new product Anthropic announced for long, asynchronous workflows that run in Anthropic's cloud with audit logs and human approval points.

Multiple templates can chain together. Anthropic showed a fund accountant launching a valuation review across twelve portfolio companies, with sub-agents running valuations in parallel and handing off to a statement auditor agent for the final tie-out before client delivery. End-to-end workflows now run as a sequence of agents with human approval points between them.

Audit logs and human approval points are what a CISO and a compliance officer need before any of this work lives near member data or regulatory filings.

Anthropic also announced Claude is now generally available across Microsoft Office for the first time. Excel, PowerPoint, and Word are live. Outlook is coming soon.

The day before the event, Anthropic and FIS announced a partnership to build a Financial Crimes AI Agent that compresses anti-money-laundering investigations from hours to minutes, with BMO and Amalgamated Bank as first deployments.

Anthropic's head of economics Peter McCrory presented data from a survey of eighty-one thousand Claude users (a Claude-user survey, not a representative labor-market sample). The most productive AI users worry the most about losing their jobs. Junior workers worry twice as much as senior workers. Managers reporting the same productivity gains worry much less about their own jobs.

McCrory called this a leadership gap. The workers see what's coming. The senior staff who would normally develop the next generation of leaders don't. He closed his section with the question.

Every credit union develops its next generation of leaders through some combination of formal training, lateral hiring, and learning by doing. AI is reshaping the entry-level work that has historically anchored the learning-by-doing piece. Credit union CEOs and HR leads have a window of a few years to build mechanisms so the people running the institution in 2032 develop tacit knowledge through paths AI has not dissolved.

The closing panel had Lori Beer (CIO of JPMorgan Chase), Marco Argenti (CIO of Goldman), and Peter Zaffino (CEO of AIG) on what worked, what didn't, and what they'd do differently.

Argenti described the pace as faster than previous waves of enterprise technology. Ten years of AWS adoption are happening in one year, with the cadence of the business shifted from quarterly to daily. The core, payments, fraud, and lending platforms credit unions and community banks rely on are moving at the same pace.

The internal appetite has flipped, too. Argenti said until two or three months ago he felt like he was pushing a car up a hill. Now the car has gone over the top and is rolling down. He's trying to hold it. The same shift is starting to show up at credit unions and community banks, where staff demand for AI tools is starting to move faster than leadership can govern.

Argenti also framed the deployment in three waves. Wave one is coding and engineering. Wave two is operational processes. Wave three is decision-making, where Goldman expects the biggest growth.

1. Move before full consensus is reached.

Zaffino's leadership team wanted to move slower than he did. He moved anyway. Zaffino argued that moving before full consensus put AIG ahead of where it otherwise would have been. He added a second lesson on rollout speed. The technology outpaced linear, sequential pilots. Parallel rollouts with clear approval gates worked better than running one at a time.

2. Pair top-down mandate with peer champions.

Argenti described two ways AI rollouts get stuck. A portfolio of pilots that never produces a usable outcome. A CEO mandate that nobody on the floor takes seriously. Goldman mitigates the latter by pairing mandate from the top with what Argenti called “bar raisers,” peer AI champions inside the organization who become the place colleagues go for help.

3. Fund the foundational platform without measuring its return.

Beer split JPMorgan's AI spending into two categories. The first is foundational platform investment, including model risk governance, agent orchestration, security controls, and training. JPMorgan treats those as must-do. The second is downstream revenue-generating applications, where JPMorgan tracks ROI closely. JPMorgan funded the foundational layer without trying to prove its return inside eighteen months.

At the end of his economic breakdown, Peter McCrory reminded the audience that we have an opportunity "to shape the future so that the benefits are broad-based and the costs of the transition are not unequally borne," he said. "We have agency over our future."

Anthropic plans to convene the same group again in twelve months.