Anthropic just turned Claude into back-office infrastructure for Wall Street. On Tuesday, May 5, 2026, at an invite-only event in New York — with JPMorganChase CEO Jamie Dimon onstage with Dario Amodei for the first time — Anthropic released Claude Opus 4.7, a library of ten pre-built financial services agent templates, full add-ins for Excel, PowerPoint, and Word, and a native Moody's app inside Claude covering 600 million public and private companies. Plus a Financial Crimes AI Agent built jointly with FIS that BMO and Amalgamated Bank are already piloting against the industry's roughly $40 billion annual AML problem.

This is the day-after-the-JV moment. Twenty-four hours after the $1.5 billion enterprise services joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, Anthropic shipped what the JV will actually sell: not "Claude API access," but a vertical agent stack with the data feeds, the application surface, and the regulator-ready scaffolding banks have spent two years asking for. Citadel, FIS, BNY, Carlyle, Mizuho, Travelers, Walleye Capital, and Hg are named in the launch materials. JPMorgan, Goldman, Citi, AIG, and Visa are already in production.

For enterprise AI buyers outside finance, the read-through is bigger than one industry. Anthropic just demonstrated the playbook every model vendor is going to copy: own the model, own the data layer, own the application surface, and let the agent templates carry the workflow. The question on every CIO's desk is which of those four layers their incumbent vendors still own — and which ones just got commoditized.

What Anthropic Actually Shipped

Strip away the Dimon-on-stage choreography and four discrete artifacts hit the market on Tuesday.

1. Claude Opus 4.7. The new flagship is state-of-the-art on Vals AI's Finance Agent benchmark at 64.37%, edging out Claude Sonnet 4.6 (63.33%), Muse Spark (60.59%), DeepSeek V4 (60.39%), and Claude Opus 4.6 in thinking mode (60.05%). Finance Agent is not a static QA benchmark — it measures multi-step financial analysis: planning, tool use, model-building from filings, and producing audit-quality outputs. The 1.04-point delta between Opus 4.7 and Sonnet 4.6 looks small in isolation, but Anthropic's internal positioning is that Opus 4.7 is built for multi-turn agentic workflows where each step compounds errors — exactly the regime where pitchbook generation and credit-memo drafting have historically broken down. Context window stays at 1M tokens with a January 2026 knowledge cutoff.

2. Ten financial services agent templates. Splitting the agent library into the two halves of the bank tells you how seriously Anthropic took the workflow research:

Research and client coverage: Pitch builder (target lists, comparables, drafted pitchbooks), Meeting preparer (client and counterparty briefs), Earnings reviewer (transcripts, filings, model updates, change-flagging), Model builder (financial models from filings and data feeds), Market researcher (sector synthesis).

Finance and operations: Valuation reviewer (valuations against comparables and standards), General ledger reconciler (account reconciliation, NAV calculations), Month-end closer (close checklists, journal entries), Statement auditor (consistency and audit-readiness review), KYC screener (entity files, document review, escalation packaging).

These are not prompts. They are templates with embedded tool-use scaffolding, data-source bindings, output formats, and review checkpoints — the kind of thing a managing director spent the last 18 months trying and failing to get a generic ChatGPT Enterprise license to do reliably. Plugins are available today on paid Claude plans; Managed Agents (the hosted, observable runtime that actually executes them) are in public beta.

3. Full Microsoft 365 integration. Add-ins for Excel, PowerPoint, and Word are generally available today; Outlook is forthcoming. The technically interesting line in Anthropic's launch post is this: "Context carries automatically between applications, so work that starts in a model can end in a deck without re-explaining." That is a single agent session preserving state across three application contexts — Excel cells, PowerPoint slides, Word paragraphs — without the analyst manually re-prompting at each handoff. This is the production-grade version of what Microsoft Copilot has been promising for two years and what the unbundled Claude desktop app could not deliver.

4. Moody's as a native app — plus 15 other data partners. Moody's Corporation is embedding its full platform into Claude as a native MCP application, exposing proprietary credit ratings and data on more than 600 million public and private companies without leaving the Claude interface. The wider data partner roster: FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG, Daloopa, Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C IntraLinks, Third Bridge, Verisk — sixteen of the most expensive data subscriptions on a Wall Street desk, all reachable through the same agent runtime.

A separate FIS partnership shipped the Financial Crimes AI Agent — FIS provides financial-data infrastructure, regulatory connectors, and bank-system integration; Anthropic provides Claude reasoning, agent orchestration, and the embedded engineering team. BMO and Amalgamated Bank are the announced first deployments, with broader rollout planned for 2H 2026. Stated target: collapse AML investigations from hours-or-days to minutes against a roughly $40 billion annual industry cost in fines, investigations, and false-positive review work.

Why This Hits Different

Anthropic is not the first vendor to ship "AI for banks." It is the first vendor to ship the whole sandwich in a single product event, and the layering is what enterprise buyers should pay attention to.

The model layer (Opus 4.7) wins a vertical benchmark — not because Anthropic invented finance reasoning, but because the model was specifically post-trained on the failure modes that show up in multi-step agentic workflows. The data layer (Moody's, FactSet, S&P, MSCI, et al.) eliminates the integration tax that has been the actual bottleneck on most internal Claude deployments — analysts kept hitting walls every time the model needed a number that lived inside a paid feed. The application layer (Excel, PowerPoint, Word) puts the agent in the surface where the work already happens — which is, empirically, the only place enterprise AI sticks. And the agent template layer (the ten pre-builts) takes the workflow expertise that used to require a forward-deployed engineering team and packages it as something a head of operations can deploy without a custom engagement.

That last layer is the one that gets attention but earns its own paragraph: Anthropic has effectively published the workflow design for 10 of the highest-value back-office processes in banking and insurance. Every consulting firm that was selling a six-week engagement to "stand up a Claude pitchbook agent" just got their deliverable shipped as a checkbox in the marketplace. Anthropic's own Claude Partner Network — Accenture, Deloitte, PwC, plus the JV — gets to focus on the harder, more profitable problem of customizing those templates against a specific bank's general ledger and risk-policy environment.

The competitive read against OpenAI is also telling. OpenAI's PwC partnership announced May 4 is structurally similar — finance-focused agents for forecasting, planning, reporting, procurement, payments, treasury — but OpenAI did not pair its announcement with a model release optimized for finance, did not ship a native Microsoft 365 surface (it has its own M365 Copilot competition problem), and did not lock down a Moody's-grade native data app. The announcements look symmetric on the press-release page; the products are not.

The Customer List Tells the Story

Public customer disclosures at this stage of an enterprise AI rollout are usually marketing theater. This list is something else.

In production: JPMorganChase, Goldman Sachs, Citi, AIG, Visa.

Named launch partners: Citadel, FIS, BNY, Carlyle, Mizuho, Travelers, Walleye Capital, Hg.

Joint venture sponsors: Blackstone, Hellman & Friedman, Goldman Sachs (each contributing $300M of the $1.5B vehicle).

Financial Crimes Agent first deployments: BMO, Amalgamated Bank.

Three of the top four U.S. money-center banks. The largest custodian (BNY). Two of the largest alternative asset managers (Blackstone, Carlyle). One of the largest hedge funds (Citadel). The dominant core-banking technology vendor (FIS). A multi-line insurer (AIG) and a P&C carrier (Travelers). Plus Visa, which is not technically a bank but is in nearly every payments-fraud workflow on earth.

The strategic point is that Anthropic is no longer prospecting Wall Street; Wall Street is co-designing the product. When Goldman Sachs is putting $300M into the JV, Citi is in production, JPMorgan's CEO is on stage, and Citadel is named as a launch partner, the marketing department isn't selecting these logos — the engineering roadmap is being shaped by them. That is the moat that takes competitors twelve to eighteen months to overcome, regardless of whose model wins the next benchmark cycle.

What Gets Disrupted

For enterprise buyers, the question is not "should we use Claude in finance?" — most banks reading this already are. The question is which line items in next year's budget just became negotiable.

Mid-market financial-data subscriptions. When Moody's, FactSet, S&P, MSCI, PitchBook, Morningstar, LSEG, and Verisk are all reachable through one agent runtime, the analyst desk's case for maintaining all of them shifts. The data is still licensed; the seats and workflow tools attached to those licenses are not.

Pitchbook and presentation tooling. The Pitch builder, Meeting preparer, and Statement auditor agents working through the PowerPoint and Word add-ins compress a junior-banker workstream that has historically supported its own ecosystem of templates, plug-ins, and outsourced production. Vendors selling presentation-automation tooling against a per-seat banker model have to argue why their layer is better than Claude's native one.

Legacy GRC and AML tooling. The FIS-Anthropic Financial Crimes Agent is the most concrete example. False-positive AML alerts are the single largest line item in financial-crime operations, and the agent's stated value prop is to compress investigation duration by an order of magnitude. Vendors selling rules-engine alert triage and case-management tooling have a 2H 2026 problem.

The "ChatGPT Enterprise but for finance" startup category. Roughly two dozen Series A/B startups have been raising on a "Claude/GPT for [pitchbooks / credit memos / KYC / month-end close]" thesis. Those theses just got harder to defend. Anthropic shipping the templates as marketplace artifacts means the wedge has to be either deeper (true workflow customization for a specific institution) or narrower (a defensible compliance moat the model vendor will not cross). The middle of that distribution is now under direct competitive pressure.

Generic Claude Enterprise deployments inside finance. This is the awkward one. If you are a bank running a horizontal Claude rollout with internally-built agents, you now have to decide whether to keep maintaining your custom pitchbook agent or switch to the marketplace template. The marketplace template is going to get more reps, more updates, and more vendor support. The internal one belongs to your team. Most CIOs are going to want to switch — and most heads of internal AI engineering are going to push back.

The Enterprise Buyer's Read

Three concrete actions for any enterprise AI leader, not just in financial services.

One — pull the contract on your most-expensive vertical data feeds. If Moody's just agreed to expose its full platform as a Claude-native MCP app, your Bloomberg, Refinitiv, or sector-specific data licenses are in the renegotiation window. Vendor pricing power is a function of distribution; Anthropic just compressed distribution. Even if you're not in finance, the pattern repeats — Reuters, LexisNexis, Westlaw, IQVIA, Cotality, IHS Markit are all sitting on the same kind of contracts and the same kind of MCP-exposure decision.

Two — re-evaluate "build" versus "marketplace template" for back-office agents. Most internal AI teams that have been building custom pitchbook / KYC / month-end / month-end agents over the last 18 months should hold their next sprint review against Anthropic's templates. If 70% of your custom build is replicated by the template, the question is whether the remaining 30% is worth the maintenance burden. Often it isn't. The right answer is fork the template, customize the last mile, and reallocate the team.

Three — watch the Microsoft 365 surface as a leading indicator. The Excel-PowerPoint-Word context-carry feature is the single most important thing in this announcement that nobody outside enterprise AI infrastructure will write about. Whoever owns the application surface where work actually happens — and not just the chat interface — owns the next phase of agent adoption. Microsoft, Google Workspace, Notion, Atlassian, and Salesforce are now in a race that just got more concrete. Track which agent platforms get add-in parity and which ones don't; that will tell you who's still in the running for production deployment in 2027.

The deeper pattern is the one Anthropic just made explicit: enterprise AI is no longer a model bake-off — it's a stack bake-off. Model + data + application surface + workflow templates, all four layers shipped together. Every vendor going to market with one of those four layers had a worse Tuesday than they admit. Every enterprise buyer with a fragmented AI vendor portfolio just got a clearer picture of where the consolidation is heading.

Sources:

• Agents for financial services and insurance — Anthropic
• Anthropic deepens push into Wall Street with new AI agents — Fortune
• Anthropic Unveils AI Agents to Field Financial Services Tasks — Bloomberg
• The day after the $1.5bn JV, Anthropic shipped what the JV will sell — TNW
• FIS Brings Agentic AI to Banking with Anthropic — FIS
• Vals AI Finance Agent v1.1 benchmark