Anthropic's 10 AI Agents Just Rewired Wall Street, and the Jobs Math Is Getting Uncomfortable
In 48 hours, Anthropic unveiled 10 ready-to-run banking agents, a $1.5 billion joint venture with Blackstone and Goldman Sachs, and a new model that leads every finance benchmark. The pitch: Claude becomes the operating layer for global capital markets. The catch: junior bankers are first in line.
Jamie Dimon and Dario Amodei don't share a stage by accident. When the CEO of JPMorgan Chase and the co-founder of Anthropic appeared together at an invite-only financial services briefing in New York on May 5, 2026, both men knew the cameras were watching. Dimon built a live Treasury asset-swap analysis dashboard from a blank Excel sheet in under 20 minutes. The message was unmistakable: this isn't a pilot program. It's production.
The New York event capped what Fortune called a "48-hour blitz" that saw Anthropic drop 10 pre-built AI agent templates for financial services, debut a Microsoft 365 integration spanning Excel, Word, PowerPoint and Outlook, announce new data partnerships with Moody's, FactSet, Morningstar, S&P Global and Dun & Bradstreet, and reveal a $1.5 billion joint venture with Blackstone, Hellman & Friedman and Goldman Sachs to embed Claude directly into hundreds of enterprises. The underlying model powering it all, Claude Opus 4.7, now leads the Vals AI Finance Agent benchmark with a score of 64.37%.
Anthropic first entered financial services in July 2025. Thirteen months later, it has Claude in production at JPMorgan Chase, Goldman Sachs, Citi, AIG and Visa. The trajectory isn't incremental. It's a structural push to become what every software vendor dreams of being: infrastructure.
The 48-Hour Wall Street Blitz
The timing was deliberate. On May 4, the day before the flagship New York event, FIS announced a partnership with Anthropic to build a Financial Crimes AI Agent. FIS isn't a minor player. The company powers roughly 12% of the global economy's transactions, deposits, payments, and credit operations. Embedding Claude inside that infrastructure means reaching thousands of financial institutions without asking any of them to switch vendors.
The Financial Crimes AI Agent compresses anti-money-laundering investigations from days to minutes. It automatically assembles evidence across a bank's core systems, evaluates activity against known typologies, and surfaces the highest-risk cases for human investigator review. BMO and Amalgamated Bank are the first development partners, with general availability planned for the second half of 2026. Crucially, client data stays within FIS-controlled infrastructure at all times; Claude functions as the reasoning layer, one step removed from the source data.
"The future is about a trusted provider who manages the data, who governs the agents, and who stands between your customers and the AI making decisions about their money."
Stephanie Ferris, CEO, FIS — FIS Press Release, May 4, 2026
Then came the joint venture. Anthropic, Blackstone and Hellman & Friedman each contributed roughly $300 million, with Goldman Sachs at $150 million; Apollo Global Management, General Atlantic, Leonard Green, GIC and Sequoia Capital also participated. The Wall Street Journal reported the $1.5 billion total, which Anthropic hasn't formally confirmed. Whatever the exact figure, the structure is novel: a private equity-backed entity designed to forward-deploy Claude directly into the portfolios of some of the world's largest PE firms. No AI company has previously had distribution at that scale.
Enterprise demand context: Anthropic CFO Krishna Rao said at the joint venture announcement that "enterprise demand for Claude is significantly outpacing any single delivery model." Separately, CEO Dario Amodei disclosed at the May 5 event that the company had projected 10x growth internally, only to see 80x adoption instead.
10 Agents, One Mission
The ten agent templates released May 5 aren't tools. They're reference architectures. Each packages three components: skills (domain knowledge and instructions for the task), connectors (governed, real-time access to external data), and subagents (additional Claude models called in for specific sub-tasks like comparables selection or methodology checks). The architecture is designed so that a bank's compliance, risk, and engineering teams can customize the templates without touching the underlying model.
Pitch Agent
Hand it a target list. Get back a comps model in Excel, a pitchbook drafted in PowerPoint, and a cover note in Outlook. A full pitch package in hours.
KYC Screener
Screens know-your-customer files against verified identity data from Dun & Bradstreet's Commercial Graph and D-U-N-S system for auditable onboarding.
Month-End Close
Handles accounting reconciliation, cross-checks entries against linked workbooks, and produces close narratives against a firm's own templates in Word.
Market Researcher
Tracks sector and issuer developments, synthesizes news, filings, and broker research, and flags items for credit and risk review automatically.
DCF / Comps Agent
Builds discounted cash flow models from filings and data feeds, audits formulas across linked workbooks, and runs sensitivity analyses in Excel.
Financial Crimes Agent
Via the FIS partnership: compresses AML investigations from days to minutes, assembles evidence, evaluates typologies, and surfaces high-risk cases for review.
The Microsoft 365 integration is what makes the whole stack practically usable. Once the Claude add-ins are installed, context carries across applications. An analyst who starts a DCF model in Excel doesn't re-explain the deal when the work shifts to a PowerPoint pitchbook. The cross-application memory is a genuine workflow change, not a cosmetic one. Outlook integration is listed as "coming soon."
Data access scales to match. Claude now connects to FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, Daloopa, IBISWorld, SS&C Intralinks, Third Bridge, and Verisk, alongside internal data warehouses, research repositories, and CRMs. All connections operate under governed access controls. The question of which connectors a firm enables, and what permissions each carries, is where compliance teams will spend most of their implementation time.
Compliance note: Anthropic explicitly states in its GitHub repository documentation that all agent outputs are drafts requiring qualified human review. The agents don't execute transactions, don't approve onboarding independently, and don't write directly into books of record. Every output requires sign-off from a licensed professional.
What the Banks Are Actually Saying
The closing panel at the May 5 event was unusual for how specific it got. Goldman Sachs CIO Marco Argenti, JPMorgan Chase CIO Lori Beer, and AIG CEO Peter Zaffino each described real deployment data, not roadmaps.
Argenti outlined three sequential waves of AI adoption at Goldman. First: empowering the technology team, roughly a third of the firm, to work at what he called "a completely different pace." Second: reimagining operational processes end-to-end. Third, and most consequential long-term: using AI to improve risk and investment decisions themselves.
"This is the first time that instead of buying infrastructure, you can actually buy intelligence."
Marco Argenti, CIO, Goldman Sachs — Fortune, May 5, 2026
AIG's Zaffino shared what may be the most striking benchmark from the event: Claude, out of the box, scored 88% accuracy against expert-level claims assessors in AIG's underwriting workflow. He was careful to frame the implication correctly. "The theory is, can it get better? Yes. But that assumes that the claims expert doesn't get better," he said. The human professional isn't static either.
Beer's remarks from JPMorgan were more cautionary. The bank has built a cyberthreat model for security analysts and integrated security intelligence directly into its software development process. Her point was structural: governance and risk frameworks have to be built into the platform from day one, not retrofitted. "There are capabilities we need, platforms we need to build, agent orchestration to protect and secure... We don't measure ROI on those things. They are must-dos," she told the audience.
| Institution | Deployment Stage | Key Use Case | Notable Data Point |
|---|---|---|---|
| Goldman Sachs | Wave 2: operational redesign | Trade accounting, client onboarding, research acceleration | AI deployment scaling without proportional new hiring; COO John Waldron confirmed the bank is "scaling up without requiring much more hiring" |
| JPMorgan Chase | Production across front and back office | Cyberthreat modeling, wealth management coaching, dev tooling | Connect Coach AI product boosted advisor capacity to handle more clients per advisor |
| AIG | Production in underwriting and claims | Claims assessment, policy underwriting support | Claude scored 88% accuracy vs. expert claims assessors out of the box |
| BMO / Amalgamated Bank | Development (general availability H2 2026) | AML investigation via FIS Financial Crimes Agent | First institutions to deploy the FIS co-designed agent |
| Commonwealth Bank of Australia | Production | Fraud prevention, customer service | CTO Rodrigo Castillo called the partnership "foundational to our strategy to become a global leader in AI innovation in banking" |
| Bridgewater (AIA Labs) | Production since 2023 | Investment Analyst Assistant: Python code generation, data visualization, financial analysis | Claude Opus 4 passed 5 of 7 levels of the Financial Modeling World Cup; 83% accuracy on complex Excel tasks |
The Junior Banker Problem
Nicholas Lin spent two and a half years on Morgan Stanley's M&A team before moving to tech investing at Singapore's sovereign wealth fund. He's now Anthropic's product lead for financial services. Speaking to Bloomberg shortly before the May 5 launch, Lin said financial AI applications are "just a few months behind" coding applications, "which we've seen massive acceleration in." Thousands of coding jobs have already disappeared.
The ten agent templates released May 5 cover the exact tasks that define a junior analyst's day: pitchbooks, comps, earnings reviews, financial model construction, valuations. A video in the product announcement showed comparable company analysis completing in seconds. It's not a subtle implication.
"A lot of these problems we're hoping to solve are just so near and dear to my heart because I spent probably 75% of my time just doing this manual data analysis, PowerPoint creation, making sure that the text boxes really match."
Nicholas Lin, Product Lead for Financial Services, Anthropic — eFinancialCareers, May 2026
One junior banker, speaking anonymously to eFinancialCareers, confirmed the internal pressure is already real. His team has stopped hiring at analyst level. He's now expected to produce more output than before, with AI tools described by managing directors as a productivity multiplier rather than a workload reducer.
The macro picture is less clear-cut. A 2026 Cambridge Judge Business School report on AI in financial services found that 24% of industry respondents expect a net reduction in roles, up from 13% in the prior three years. But 25% expect significant reskilling without large net losses, and 10% expect a net increase in jobs. The distribution of outcomes is widening, not converging.
Deloitte's research puts potential upside at $3.5 million in additional front-office revenue per employee at the top 14 global investment banks, from productivity gains of 27% to 35% via generative AI. Those gains don't require firing people. They require the same headcount doing more, faster. Whether that translates to fewer new hires or fewer existing jobs depends on whether deal volumes grow to absorb the additional capacity. Right now, Goldman Sachs has actually cut its 2026 deal count outlook to roughly 100 IPOs, down from earlier projections, signaling the demand side isn't yet expanding to match AI supply.
What roles are actually at risk
- Junior analyst and associate roles focused on modeling, formatting and data processing face the most direct pressure from agent templates.
- KYC and AML analysts in compliance operations are the target of both the FIS Financial Crimes Agent and the standalone KYC Screener template.
- Financial reporting roles at mid-market firms: 87% of CFOs at that tier are already turning to AI for financial reporting work, per PYMNTS research.
- Senior and client-facing roles remain more insulated. They require contextual judgment, negotiation, and the kind of long-term relationship capital AI doesn't accumulate.
Regulators and the Mythos Shadow
Sitting behind all of Anthropic's financial services ambitions is a harder conversation that surfaced at the same May 5 event. Amodei warned publicly that Anthropic's restricted-access model, Claude Mythos Preview, has identified tens of thousands of high-severity software vulnerabilities, including nearly 300 in Firefox alone, and that there's a six-to-12-month window to patch them before adversarial AI systems catch up. Most of those vulnerabilities haven't been publicly disclosed because they remain unpatched.
An earlier Claude model found roughly 20 vulnerabilities in Firefox. Mythos found nearly 300. The scale of potential exploits has grown with each model generation. Anthropic has limited Mythos to a small number of partner companies precisely because of concerns about what criminals or adversarial nation-states could do with the same capability.
"The bad guys will exploit them if they are identified."
Dario Amodei, CEO, Anthropic — CNBC, May 5, 2026
UK financial regulators are taking the concern seriously. Officials from the Bank of England, Financial Conduct Authority and HM Treasury are holding urgent talks with banks and cybersecurity officials. The National Cyber Security Centre is involved. Major banks, insurers, and exchanges are to be warned about findings at a meeting scheduled within the fortnight following the Mythos disclosure. The matter is also set for discussion at the Cross Market Operational Resilience Group, co-chaired by the Bank of England's executive director for supervisory risk.
In the US, Treasury Secretary Scott Bessent convened Wall Street bank leaders separately to assess exposure. The dual reality is uncomfortable: the same company selling Wall Street its AI operating layer is simultaneously disclosing that AI has created a systemic vulnerability window the industry has roughly a year to close.
Regulatory signal: The Cambridge Judge 2026 report found that 78% of surveyed regulators view AI as significant or transformative for financial supervision by 2030, and that regulators are more concerned than industry about AI concentration risk, 43% versus 28%. Regulators are also more likely than industry to place primary accountability for AI outcomes on the regulated financial institution, not on the AI vendor.
OpenAI Is Watching
Anthropic isn't operating in a vacuum. OpenAI is pursuing a similar enterprise joint venture, reportedly raising $4 billion against a $10 billion valuation to create new channels for large-scale enterprise AI deals, per Bloomberg. The competitive dynamic is straightforward: both companies need enterprise contracts to justify the capital expenditures going into compute. Consumer subscriptions don't generate the kind of multi-year, high-margin commitments that make frontier model development economically viable at scale.
Anthropic's positioning in financial services rests on three differentiators: safety-first reputation, coding performance via Claude Code, and the depth of its vertical integration, specifically the combination of pre-built agents, Microsoft 365 plugins, and a growing ecosystem of data connectors. OpenAI's strength is brand recognition and developer mindshare. Which matters more to a Fortune 500 CIO depends on whether the bank's primary use case is customer-facing product development or back-office workflow automation.
| Dimension | Anthropic / Claude | OpenAI / GPT |
|---|---|---|
| Finance-specific benchmark | 64.37% on Vals AI Finance Agent (Claude Opus 4.7, category leader) | Not disclosed on Vals AI Finance Agent as of publication |
| Pre-built finance agents | 10 templates released May 2026; marketplace available now | Finance-focused tools via plugins; no equivalent marketplace |
| Financial data partnerships | FactSet, Moody's, Morningstar, PitchBook, S&P Capital IQ, MSCI, LSEG, Daloopa, Dun & Bradstreet, Verisk, Third Bridge | Partnerships in progress; fewer publicly confirmed financial-data connectors |
| Enterprise joint venture | $1.5B JV with Blackstone, H&F, Goldman Sachs (announced May 4, 2026) | Reportedly pursuing similar structure; $4B raise against $10B valuation |
| Systemic risk posture | Mythos vulnerability disclosure; proactive regulator engagement | No equivalent public disclosure as of publication |
The market data partnerships are where Anthropic's moat is deepest in the near term. A Claude agent that can natively reason over Moody's credit ratings, PitchBook private market data, and a firm's own internal research repository simultaneously is operationally different from a general-purpose chatbot with document upload. That kind of grounded, auditable reasoning is what enterprise risk officers need before signing off on production deployment.
Frequently Asked Questions
What are Anthropic's Claude finance agents?
Anthropic's Claude finance agents are 10 pre-built AI agent templates released in May 2026 for tasks including pitchbook creation, KYC screening, financial modeling, month-end close, and market research. Each runs as a plugin in Claude Cowork or Claude Code, and connects to major financial data providers like FactSet and Moody's. All outputs require human review before use.
Which banks are using Anthropic's Claude?
As of May 2026, Claude is in production at JPMorgan Chase, Goldman Sachs, Citi, AIG, Visa, Commonwealth Bank of Australia, and Bridgewater (via AIA Labs). BMO and Amalgamated Bank are in development with the FIS Financial Crimes AI Agent, with general availability planned for H2 2026.
Will AI replace investment banking jobs?
The short-term evidence shows hiring freezes at junior analyst levels rather than mass layoffs. A 2026 Cambridge report found 24% of financial industry respondents expect net job reductions, up from 13% in recent prior years. The jobs most at risk are modeling, formatting, KYC screening, and AML review. Senior client-facing roles are more insulated from near-term automation.
What is the Vals AI Finance Agent benchmark?
The Vals AI Finance Agent benchmark tests AI models on realistic financial analysis tasks, including equity research, financial modeling, and data synthesis. Claude Opus 4.7 currently leads the benchmark with a score of 64.37%. It is one of the primary publicly available evaluation frameworks for comparing AI performance on financial work.
What is the FIS Financial Crimes AI Agent?
The FIS Financial Crimes AI Agent is a tool co-designed by Anthropic and FIS that compresses anti-money-laundering alert investigations from days to minutes. It assembles evidence across a bank's core systems, evaluates transactions against known fraud typologies, and presents high-risk cases for human investigator review. Client data remains within FIS infrastructure throughout.
What is the Anthropic, Blackstone, and Goldman Sachs joint venture?
The joint venture, announced May 4, 2026, is a private equity-backed AI services company designed to embed Claude across hundreds of enterprises via forward-deployed engineering teams. Anthropic, Blackstone, and Hellman & Friedman each contributed roughly $300 million; Goldman Sachs contributed $150 million. The WSJ reported a $1.5 billion total, which Anthropic has not formally confirmed.
How does Claude Opus 4.7 differ from earlier Claude versions for financial work?
Claude Opus 4.7 is Anthropic's current flagship for financial tasks and leads industry benchmarks on the Vals AI Finance Agent evaluation. When deployed by FundamentalLabs on Excel work, Claude Opus 4 passed 5 of 7 levels in the Financial Modeling World Cup and scored 83% accuracy on complex Excel tasks. Earlier Claude models showed capable but lower performance across the same task categories.
What regulatory concerns exist around AI in financial services?
UK regulators from the Bank of England, FCA, and HM Treasury are currently assessing risks linked to Anthropic's Mythos model, which has identified tens of thousands of software vulnerabilities. The Cambridge Judge 2026 report found 78% of regulators view AI as transformative for financial supervision by 2030, with regulators more concerned than industry about AI concentration risk and accountability gaps.
The Operating Layer Play
Anthropic's financial services push makes sense as a business strategy only if you accept one underlying premise: that enterprise AI is a winner-take-most market, not a diversified one. If banks end up with one primary AI reasoning layer embedded across their Excel models, their compliance workflows, their AML systems, and their pitchbook assembly, then the company that occupies that layer collects the compound interest on every deal, every investigation, every quarter-end close. That's the prize Anthropic is building toward.
The joint venture with Blackstone and Goldman Sachs is the clearest signal of that ambition. Private equity firms have portfolio companies in manufacturing, healthcare, logistics, retail, and real estate. A forward-deployed Claude inside those companies, deployed through a PE-backed services firm, doesn't need Anthropic to sell each deal individually. The distribution builds itself.
What remains unresolved is the regulatory question. Anthropic has been unusually forthcoming about AI-generated vulnerabilities and systemic risks, a posture that distinguishes it from competitors but also raises the question of whether the product and the risk can be cleanly separated. You can't sell a bank on Claude as its compliance operating layer while simultaneously disclosing that AI has created a vulnerability window that bad actors may exploit before the industry patches it. Both things are true, and financial institutions are not historically comfortable operating in that kind of ambiguity.
The workforce question compounds the tension. Anthropic's own product lead describes the agent templates as solving problems that were "near and dear" to his heart as a junior banker. That's an honest framing. It's also a company describing its product as a direct replacement for early-career financial labor, in public, while pitching that same product to the firms that employ that labor. The banks aren't going to stop buying. But the junior analysts already on payroll are watching what their managing directors do next.
No comments:
Post a Comment