Claude for Finance: Anthropic Ships 10 Ready-Made Agents
On May 5, 2026, Anthropic released a bundle of ten Claude agents verticalised for financial services, covering everything from M&A pitchbooks to KYC screening, general ledger reconciliation, and month-end close. All of it runs on Claude Opus 4.7, which scores 64.37% on the Vals AI Finance Agent benchmark — one of the strongest public results on finance-specific tasks today.
The most telling signal isn't the agent list itself. It's the list of eight financial institutions already in production: Citadel, BNY, Carlyle, FIS, Walleye Capital, Mizuho, Travelers, Hg. When a 400-person hedge fund (Walleye) reports 100% Claude Code adoption and an AML player like FIS talks about "compressing a days-long investigation into minutes," we're clearly past the prototype phase. For MLOps teams wondering what a serious production AI agent actually looks like in 2026, this is the case study to dissect.
The context: from horizontal LLM to vertical agent
For three years, the implicit bet has been that one generalist LLM would suffice with good prompt engineering. The reality has been different: for regulated, document-heavy sectors (finance, healthcare, legal), the boundary between "impressive demo" and "approved production system" hinges on dozens of hours of integration on business specifics — modeling conventions, close-cycle quirks, regulatory wording, deliverable formats.
Those dozens of hours, multiplied by each team, multiplied by each use case, eventually justify a "vertical skill" layer between the model and the user. That's exactly what Anthropic ships here: an agent = a skill (instructions + domain knowledge) + connectors (data access) + sub-agents (methodology, checks). All exposed as a plugin in Claude Cowork / Claude Code, or as a cookbook for Managed Agents in public beta on Claude Platform.
The ten agents in two blocks
Research and client coverage
- Pitch builder — target lists, comparables, pitchbook generation
- Meeting preparer — briefs on client or counterparty ahead of meetings
- Earnings reviewer — transcript and filing analysis, model updates
- Model builder — financial model creation from filings and data feeds
- Market researcher — sector tracking, news synthesis, sell-side research
Finance and operations
- Valuation reviewer — methodology checks on comparables
- General ledger reconciler — account reconciliation, NAV calculations
- Month-end closer — close checklists, journal entries, reporting
- Statement auditor — consistency review and audit preparation
- KYC screener — entity file assembly, document review, compliance escalations
These ten templates cover roughly everything a back-office finance team or middle-office investment shop does on a daily basis. It's the inventory, not the revolution — but bundled in one drop, ready to plug in, this is the first time a model vendor offers this at this level of granularity.
The underlying architecture
Three things are worth stopping on.
Native connectors to 18 data partners. FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, Daloopa, Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, Verisk, and Moody's via an MCP app exposing 600M+ public and private companies. It's the data BYOK pattern: plug in existing subscriptions, the agent uses them, you pay the data sub once instead of per call.
Native Microsoft 365. Excel, PowerPoint, Word add-ins generally available. Outlook coming. Context from an Excel file propagates automatically to a PowerPoint slide with no copy-paste. That's the friction that killed 80% of finance-agent prototypes — gone in one move.
Auditable governance. Per-tool permissions, managed credentials vault, full audit logs in Claude Console for compliance or engineering inspection. Combined with the "user firmly in the loop" posture — every output destined for a client or filing runs through human review — it gives a clean reading of how these agents fit into the AI Act compliance stack we covered last week.
The eight customers already in production
The logo list isn't trivial. Anthropic publishes references you can verify on LinkedIn, with named individuals speaking for themselves:
- Citadel — investment research and coverage models. Atte Lahtiranta (Head of Core Engineering) talks about a "step-change in efficiency" with Claude for Excel.
- FIS — AML investigation compression. Stephanie Ferris (CEO) announces a move from days to minutes on the investigation cycle.
- BNY — the Eliza agent processing full case files end to end. Leigh-Ann Russell (CIO): "digital employees who work the case end to end."
- Carlyle — adoption at the core of the AI stack, investing and operations.
- Mizuho — meeting prep transformed. "Prep time has been transformed into idea time."
- Travelers — engineering productivity and operational excellence.
- Walleye Capital — 100% Claude Code adoption across the 400-person hedge fund. Probably the most telling stat in the pack.
- Hg — due diligence and financial modeling on live deals.
When eight institutions of this size accept their logo on a product page, it's a commercial signal — but also a practical one: the friction of putting these systems into production has dropped enough that financial-institution CTOs go in on the record.
What this actually changes for AI teams
Three concrete implications:
1. Vertical agents become the reference standard. Maintaining a homegrown library of prompts to replicate an "earnings reviewer" becomes a waste of time when an official template exists, is maintained by the model vendor, and benefits from version upgrades. The MLOps competence that gains value isn't "prompting an LLM well in finance" anymore — it's "integrating + customising + observing an existing vertical agent into the business system".
2. Audit + permissions is no longer optional for regulated sectors. Anthropic ships natively what homegrown teams have been painfully building since 2024: per-tool-call audit logs, credentials vault, fine-grained permissions scoping. Exactly what articles 12 and 14 of the AI Act require. Competitors (OpenAI, Mistral) will have to align — or lose the sectors where compliance traceability wins.
3. For MLOps freelancers, the market shifts from build to tuning and integration. The work expected from an MLOps freelance in 2026 looks less and less like "write an agent" and more like "integrate this existing vertical agent into your SI, customise to your modeling conventions, validate compliance coverage, wire observability, write incident playbooks". It's a shift of value — not a reduction, but a recalibration to anticipate commercially.
For a finance team that hasn't started yet: audit Anthropic's financial-agent marketplace on GitHub, plug in a single agent (pitch builder or KYC screener depending on urgency), evaluate over 30 days, generalise. Entry ticket has dropped from months to a handful of person-days.
TL;DR
- 10 verticalised Claude finance agents released by Anthropic on May 5, 2026, running on Claude Opus 4.7 (64.37% on Vals AI Finance Agent benchmark)
- Coverage: pitch building, meeting prep, earnings review, model building, market research, valuation review, GL reconciliation, month-end close, statement audit, KYC screening
- Native connectors: 18 data partners (FactSet, S&P Capital IQ, MSCI, PitchBook, Moody's MCP with 600M+ companies, etc.)
- Microsoft 365: Excel/PowerPoint/Word add-ins GA, Outlook coming
- 8 institutions in production: Citadel, BNY, Carlyle, FIS, Walleye (100% Claude Code adoption), Mizuho, Travelers, Hg
- Governance: audit logs, per-tool permissions, credentials vault — aligned with AI Act articles 12 and 14
Industrialising AI agents? SeedVision offers 3-5 day AI audits and 15-30 day production rollouts. See the packages or book a 30-min call.