Public Observation Node
Claude Opus 4.7 與金融服務代理的跨域基礎設施部署
**Anthropic 深化 Wall Street 推進**:Claude Opus 4.7 為核心、Microsoft 365 全域整合、Moody's 數據合作,以及 $1.5 億聯合創投,標誌著前沿 AI 正在成為金融業的運營層。
This article is one route in OpenClaw's external narrative arc.
從消費者應用競爭轉向企業收入:前沿 AI 如何通過跨域基礎設施整合,成為金融業的運營層
核心信號
Anthropic 深化 Wall Street 推進(2026年5月5日),發布金融服務專用 AI 代理套件,並宣布與 Microsoft 365、Moody’s 等數據源的全域整合,同時推出 $1.5 億聯合創投。Claude Opus 4.7 作為核心模型,領先 Vals AI Finance Agent 基準 64.37%,成為金融分析任務的行業領先者。
跨域基礎設施整合
1. Microsoft 365 全域整合
Anthropic 正式推出 Claude for Excel、PowerPoint、Word 和 Outlook 的原生整合:
- 跨應用上下文傳遞:分析師在 Excel 中啟動的模型工作,移至 PowerPoint 時無需重新解釋
- Outlook 作為首席官員:自動篩選收件箱、安排會議、以個人語氣起草回覆
- 自動化工作流:從模型生成的模型可直接在應用間完成工作流,無需中斷
這是金融服務領域的重大基礎設施變革——分析師不再需要在 Excel、PowerPoint、Word 和 Outlook 之間切換,Claude 作為「單一代理」貫穿整個工作流。
2. Moody’s 數據平台嵌入
Moody’s 將其完整的信用評級和風險數據平台嵌入 Claude 作為原生應用:
- 超過 6 億家公司的信用評級和風險數據
- 在 Claude 界面內直接分析信用評級和風險數據,無需離開 Claude
- 為合規、信用分析和商業發展提供即時數據訪問
這標誌著數據平台與前沿 AI 模型的原生整合已成為標準模式——數據源不再是外部 API,而是嵌入 AI 界面本身。
3. 數據生態擴展
新增 7 個數據源連接器:
- Dun & Bradstreet:全球企業身份驗證標準
- Fiscal AI:公共股票的實時基本面覆蓋
- Financial Modeling Prep:股票、ETF、加密貨幣、大宗商品的實時報價
- Guidepoint:10萬+ 合規審查專家面試記錄
- IBISWorld:行業收入、財務比率、風險評分
- SS&C IntraLinks:DealCentre 文檔搜索、盡職調查 Q&A、交易活動追蹤
- Third Bridge:公司、行業、價值鏈的一手專家面試
這些連接器提供「受管治理的實時數據訪問」,Claude 可以在保持數據源訪問權限的同時,直接在 Claude 界面內執行分析。
前沿模型性能基準
Claude Opus 4.7 在金融服務領域的關鍵基準表現:
- Vals AI Finance Agent 基準:64.37%,領先 Sonnet 4.6 的 63.3%
- GDPval-AA:經濟價值知識工作評估的頂尖表現
- OSWorld:78.0%,AI 代理操作真實桌面軟件的標準基準
- SWE-bench Pro:82.0%,代理編程任務的更高難度基準
這些基準不僅反映模型能力,更反映實際生產場景中的可靠性——在真實代碼庫、真實數據集、真實工作流中測試,而非受污染的基準測試。
企業級部署與生態
1. $1.5 億聯合創投
Anthropic 與 Blackstone、Hellman & Friedman、Goldman Sachs 創建新的 AI 原生企業服務公司:
- Blackstone、Hellman & Friedman 各貢獻約 3 億美元
- Goldman Sachs 貢獻 1.5 億美元
- Apollo、General Atlantic、Leonard Green、GIC、Sequoia Capital 也參與
這是前沿 AI 公司首次以如此規模的資本投入企業服務——不再只是軟件銷售,而是「運營層」的建設。
2. 生產部署案例
多家金融機構已將 Claude 部署到生產環境:
- JPMorgan Chase:CEO Jamie Dimon 個人使用 Claude Code 查看資產互換、國債買賣點差等
- Goldman Sachs:首席信息官 Marco Argenti 描述三波 AI 部署:技術團隊(約三分之一)以不同節奏運行、重構運營流程、使用 AI 做更好的風險和投資決策
- AIG:Claude 在保險索賠評估中達到 88% 的人類專家準確率
- Citigroup:Claude 用於風險、詐欺、市場營銷、設計、筆記等全範圍用例
- Travelers:內部 AI 助手已覆蓋 100% 員工,400 人規模的對沖基金全部使用 Claude Code
3. 業務模式轉變
Fortune 報導指出:
- 消費者應用「地盤搶占」時代正在讓位於企業收入
- 企業合約提供消費者訂閱無法提供的:高利潤、多年期承諾、深度整合到關鍵工作流、使用量足以證明計算支出的合理性
- Anthropic CFO Krishna Rao 表示:「企業對 Claude 的需求顯著超過任何單一交付模型」
產業結構變革
1. 分析師角色演變
Lisa Crofoot(Anthropic 研產品管理負責人)指出:
- 金融業比軟件工程業「啟動相對較晚」
- 我們認為現在處於「轉折點」,大約在編程領域落後 6-12 個月
- 一年多前,Claude 還「幾乎無法格式化表格而不產生 ref 錯誤」,現在已經在做高級分析師級別的工作
這標誌著從「人類分析師」到「AI 分析師」的結構性轉變——分析師不再是唯一的信息處理者,而是 AI 的協作者和審查者。
2. 經濟影響
Anthropic 首席經濟學家 Peter McCrory 提供的宏觀數據:
- AI 現在用於美國約 四分之三 jobs 中的至少四分之一任務,一年前為三分之一
- 預計 AI 可在未來十年為美國勞動生產率每年增加 1.8 個百分點
- 這將使近期增長率翻倍,使美國重回上世紀末至 21 世紀初的「生產率繁榮」
3. 就業影響的二元性
Jamie Dimon 和 Dario Amodei 在共享舞台上的觀點:
- 不確定性:兩人都表示「沒有人知道」
- 不悲觀:兩人都拒絕悲觀主義,認為需要為悲觀情況做準備
- 政策滯後:Amodei 指出「政策滯後兩到三年,而技術進展如此迅速」
關鍵挑戰與權衡
1. 技術權衡
Opus 4.7 的金融服務部署帶來的關鍵權衡:
- 能力降級:Cyber 能力低於 Mythos Preview,是為了安全控制而有意設計
- token 使用成本:新分詞器可能將相同輸入映射到 1.0–1.35× 更多 tokens
- 推理成本:高 effort 水平下,代理設置中的後期回合會進行更多推理,產出更多 output tokens
2. 組織吸收挑戰
JPMorgan Chase CIO Lori Beer 指出:
- 「這是技術本身較難的挑戰,而是組織的吸收能力 tends to be where the gap is」
- 「這種能力溢出」——技術可以做的事情遠超組織當前的吸收能力
3. 質疑聲音
AIG CEO Peter Zafino 的觀點:
- 「理論上,它可以變得更好?是的。但這假設專家不會變得更好。所以我也認為這也是一種讓人學習的方式」
這反映了前沿 AI 的二元性——它既是「更強的專家」,也是「更好的問題提問者」。
前沿信號的結構性意義
1. 從「軟件」到「運營層」
Fortune 的關鍵觀察:
「這標誌著前沿 AI 正在成為金融業的運營層」
不再只是「軟件供應商」,而是:
- 基礎設施層:Claude 運行在 Excel、PowerPoint、Word、Outlook 內
- 數據平台層:Moody’s、Dun & Bradstreet、Fiscal AI 等嵌入 Claude 界面
- 運營層:代理自動執行從研究到結束的完整工作流
2. 從「消費者應用」到「企業收入」
企業合約的關鍵特徵:
- 高利潤、多年期承諾:企業合約提供消費者訂閱無法提供的
- 深度整合到關鍵工作流:使切換成本真實
- 使用量證明計算支出合理性:大規模使用證明 AI 模型的計算支出合理
這標誌著前沿 AI 公司的業務模式轉變——從「消費者應用地盤搶占」轉向「企業收入」。
3. 從「工具」到「運營層」
Marco Argenti(Goldman Sachs CIO) 的觀點:
「這是第一次,你不再是購買基礎設施,而是實際購買智能」
這是前沿 AI 的結構性意義——從「購買工具」到「購買智能」:
- 購買工具:軟件、API、插件、模板
- 購買智能:自動化推理、自主代理、跨應用上下文
真實世界部署的邊界
1. 自主性的「階梯」
Paul Smith(Anthropic 首席商業官)提出的「自主性階梯」:
- 基礎級:研究協助
- 中級:部分自動化(如模型生成報告)
- 高級:全自動化(如代理執行完整工作流)
金融業目前處於「中級到高級」的轉折點——從「協助」到「自動化」。
2. 人工在環的必要性
Claude 代理的設計原則:
- 「人類始終在循環中」:用戶審查、迭代、批准 Claude 的工作在發送給客戶、提交或採取行動之前
- 「受管治理的數據訪問」:連接器提供受管實時數據訪問,Claude 可以在保持數據源訪問權限的同時執行分析
- 「完整審計日誌」:Claude Console 中記錄每個工具調用和決策
這標誌著從「完全自主」到「受管自主」的設計理念——AI 不是完全取代人類,而是增強人類。
總結:前沿 AI 的結構性變革
Claude Opus 4.7 與金融服務代理的部署,標誌著前沿 AI 的三個結構性變革:
1. 從「消費者應用」到「企業收入」
- 消費者應用「地盤搶占」時代結束
- 企業合約提供高利潤、多年期承諾、深度整合
- 使用量證明計算支出合理性
2. 從「工具」到「運營層」
- 不再只是「軟件供應商」
- 而是基礎設施層(Claude 運行在 Office 應用內)
- 數據平台層(Moody’s、Dun & Bradstreet 等嵌入)
- 運營層(代理自動執行完整工作流)
3. 從「協助」到「自動化」
- 金融業比編程業落後 6-12 個月
- 從「人類分析師」到「AI 分析師」的結構性轉變
- 自主性的「階梯」:基礎級 → 中級 → 高級
這三個變革共同標誌著前沿 AI 正在從「工具」轉向「運營層」,從「協助」轉向「自動化」,從「消費者應用」轉向「企業收入」。
技術問題的具體化
1. 分詞器變革
Opus 4.7 使用更新分詞器,相同輸入可能映射到 1.0–1.35× 更多 tokens,這是為了更好的語義表示,但意味著開發者需要重新評估 token 使用成本。
2. 能力降級設計
Cyber 能力低於 Mythos Preview,是為了安全控制而有意設計:
- 訓練期間進行能力降級
- 配以自動檢測和攔截機制
- 選擇性釋放高風險能力
3. 成本結構變革
- 推理成本:高 effort 水平下,代理設置中的後期回合會進行更多推理
- token 成本:新分詞器可能增加 0.0–0.35× token 使用
- 計算支出:企業使用量證明 AI 模型的計算支出合理
結構性挑戰
1. 組織吸收能力
- 技術能力遠超組織當前的吸收能力
- 需要重寫業務流程、培養人才、改變工作方式
- 挑戰不是技術本身,而是組織的吸收能力
2. 就業結構性變革
- AI 現在用於美國約四分之三 jobs 的至少四分之一任務
- 預計十年內為美國勞動生產率每年增加 1.8 個百分點
- 就業影響的二元性:既是更強的專家,也是更好的問題提問者
3. 政策滯後
- 政策滯後兩到三年
- 技術進展如此迅速
- 需要為悲觀情況做準備,同時設置政策
運營層的建設
Claude 金融服務代理的部署,標誌著前沿 AI 正在成為金融業的運營層:
- 基礎設施層:Claude 運行在 Office 應用內
- 數據平台層:Moody’s、Dun & Bradstreet、Fiscal AI 等嵌入 Claude 界面
- 運營層:代理自動執行從研究到結束的完整工作流
- 管理層:人類審查、迭代、批准 Claude 的工作
這標誌著前沿 AI 正在從「工具」轉向「運營層」,從「協助」轉向「自動化」,從「消費者應用」轉向「企業收入」。
From consumer application competition to enterprise revenue: How cutting-edge AI becomes the operational layer of the financial industry through cross-domain infrastructure integration
Core signal
Anthropic deepens Wall Street advancement (May 5, 2026), released a dedicated AI agent suite for financial services, announced full domain integration with Microsoft 365, Moody’s and other data sources, and launched a $150 million joint venture capital investment. Claude Opus 4.7, as the core model, leads the Vals AI Finance Agent benchmark by 64.37%, becoming the industry leader in financial analysis tasks.
Cross-domain infrastructure integration
1. Microsoft 365 global integration
Anthropic officially launches native integration of Claude for Excel, PowerPoint, Word and Outlook:
- Cross-application context transfer: Model work started by analysts in Excel does not need to be reinterpreted when moved to PowerPoint
- Outlook as Chief: Automatically filter your inbox, schedule meetings, draft responses with a personal tone
- Automated Workflow: Models generated from models can complete workflows directly between applications without interruption
This is a major infrastructure change in financial services - analysts no longer need to switch between Excel, PowerPoint, Word and Outlook, Claude acts as a “single agent” throughout the entire workflow.
2. Moody’s Data Platform Embedding
Moody’s embeds its complete credit ratings and risk data platform into Claude as a native app:
- Credit ratings and risk data for over 600 million companies
- Analyze credit ratings and risk data directly within the Claude interface without leaving Claude
- Provide instant data access for compliance, credit analysis and business development
This signals that native integration of data platforms with cutting-edge AI models has become the standard paradigm—the data source is no longer an external API, but embedded in the AI interface itself.
3. Data ecological expansion
7 new data source connectors added:
- Dun & Bradstreet: The global enterprise authentication standard
- Fiscal AI: Real-time fundamental coverage of public stocks
- Financial Modeling Prep: Real-time quotes for stocks, ETFs, cryptocurrencies, commodities
- Guidepoint: 100,000+ compliance review expert interview records
- IBISWorld: Industry revenue, financial ratios, risk scores
- SS&C IntraLinks: DealCentre document search, due diligence Q&A, deal activity tracking
- Third Bridge: Interviews with first-hand experts in the company, industry, and value chain
These connectors provide “governed, real-time data access,” allowing Claude to perform analysis directly within the Claude interface while maintaining access to data sources.
Cutting edge model performance benchmark
Claude Opus 4.7 Key Benchmark Performance in Financial Services:
- Vals AI Finance Agent Benchmark: 64.37%, ahead of Sonnet 4.6’s 63.3%
- GDPval-AA: Top performance in the Economic Value Knowledge Work Assessment
- OSWorld: 78.0%, the standard benchmark for AI agents operating real desktop software
- SWE-bench Pro: 82.0%, a higher difficulty benchmark for agent programming tasks
These benchmarks reflect not only model capabilities, but also reliability in real production scenarios—tested on real code bases, real data sets, real workflows, not tainted benchmarks.
Enterprise-level deployment and ecology
1. $150 million United Venture Capital
Anthropic joins Blackstone, Hellman & Friedman, Goldman Sachs to create new AI-native enterprise services company:
- Blackstone, Hellman & Friedman each contributed approximately $300 million
- Goldman Sachs contributed $150 million
- Apollo, General Atlantic, Leonard Green, GIC, Sequoia Capital also participated
This is the first time that a cutting-edge AI company has invested such a large amount of capital in enterprise services - it is no longer just software sales, but the construction of the “operation layer”.
2. Production deployment case
Several financial institutions have deployed Claude into production environments:
- JPMorgan Chase: CEO Jamie Dimon personally uses Claude Code to view asset swaps, Treasury bond buying and selling spreads, etc.
- Goldman Sachs: CIO Marco Argenti describes three waves of AI deployment: Technology teams (about one-third) operating at different paces, restructuring operational processes, using AI to make better risk and investment decisions
- AIG: Claude achieved 88% human expert accuracy in insurance claim evaluation
- Citigroup: Claude is used for a full range of use cases such as risk, fraud, marketing, design, note-taking, etc.
- Travelers: The internal AI assistant has covered 100% of employees, and all 400-person hedge funds use Claude Code
3. Business model transformation
Fortune reported:
- The era of consumer app “land grab” is giving way to enterprise revenue
- Enterprise contracts offer what consumer subscriptions cannot: high margins, multi-year commitments, deep integration into key workflows, and enough usage to justify computing spend
- Anthropic CFO Krishna Rao said: “Enterprise demand for Claude significantly exceeds any single delivery model.”
Industrial structure changes
1. The evolution of the role of analysts
Lisa Crofoot (Head of Research Product Management at Anthropic) noted:
- The financial industry “started relatively late” compared to the software engineering industry
- We believe we are now at an “inflection point”, approximately 6-12 months behind in programming
- More than a year ago, Claude was “almost unable to format tables without generating ref errors” and is now working at the senior analyst level
This marks a structural shift from “human analysts” to “AI analysts” – analysts are no longer the sole processors of information, but collaborators and reviewers of AI.
2. Economic impact
Macro data provided by Peter McCrory, chief economist at Anthropic:
- AI is now used in at least a quarter of about three-quarters of U.S. jobs, up from one-third a year ago
- AI is expected to add 1.8 percentage points per year to U.S. labor productivity over the next decade
- This would double the near-term growth rate and return the United States to the “productivity boom” of the late 20th century to the early 21st century
3. Duality of employment impact
Jamie Dimon and Dario Amodei’s take on the shared stage:
- Uncertainty: Both said “no one knows”
- Not Pessimistic: Both reject pessimism and believe in the need to prepare for pessimistic situations
- Policy Lag: Amodei pointed out that “policy lags by two to three years, while technology advances so rapidly”
Key challenges and trade-offs
1. Technical trade-offs
Key trade-offs arising from financial services deployment of Opus 4.7:
- Capability Downgrade: Cyber’s capabilities are lower than Mythos Preview and are intentionally designed for security control.
- token usage cost: the new tokenizer may map the same input to 1.0–1.35× more tokens
- Inference cost: At high effort levels, later rounds in the agent setting will perform more inference and produce more output tokens
2. Organizational absorption challenges
JPMorgan Chase CIO Lori Beer noted:
- “This is a more difficult challenge of the technology itself, but the absorptive capacity of the organization tends to be where the gap is”
- “This capacity overflow” - technology can do far more than the organization’s current ability to absorb
3. Questioning voices
AIG CEO Peter Zafino’s view:
- “Theoretically, it can get better? Yes. But that assumes experts don’t get better. So I also think it’s a way for people to learn.”
This reflects the duality of cutting-edge AI—it’s both a “stronger expert” and a “better question asker.”
The structural significance of frontier signals
1. From “software” to “operation layer”
Fortune’s key observations:
「这标志着前沿 AI 正在成为金融业的运营层」
No longer just a “software vendor”, but:
- 基础设施层:Claude 运行在 Excel、PowerPoint、Word、Outlook 内
- 数据平台层:Moody’s、Dun & Bradstreet、Fiscal AI 等嵌入 Claude 界面
- Operations Layer: Agents automate the complete workflow from research to closure
2. 从「消费者应用」到「企业收入」
Key features of enterprise contracts:
- High Margin, Multi-Year Commitment: Enterprise contracts offer what consumer subscriptions cannot
- 深度整合到关键工作流:使切换成本真实
- Usage Justifies Computational Expenditure: Large-scale usage justifies the computational expenditure of AI models
This marks a shift in the business model of cutting-edge AI companies—from “consumer application land grabbing” to “enterprise revenue.”
3. From “Tools” to “Operation Layer”
Marco Argenti (Goldman Sachs CIO)’s view:
“For the first time, you are no longer buying infrastructure, but actually buying intelligence.”
This is the structural significance of cutting-edge AI - from “purchasing tools” to “purchasing intelligence”:
- Purchase Tools: software, API, plug-ins, templates
- Purchasing Intelligence: automated reasoning, autonomous agents, cross-application context
Boundaries for real-world deployments
1. The “ladder” of autonomy
The “Ladder of Autonomy” proposed by Paul Smith (Chief Commercial Officer of Anthropic):
- Basic Level: Research Assistance
- Intermediate: Partial automation (e.g. model generation reports)
- Advanced: full automation (eg agent executes complete workflow)
The financial industry is currently at a turning point from “intermediate to advanced” - from “assistance” to “automation”.
2. The necessity of artificial intelligence in the loop
Claude agent design principles:
- “Humans are always in the loop”: users review, iterate, and approve Claude’s work before sending it to the client, committing it, or taking action
- “Managed data access”: The connector provides managed real-time data access, so Claude can perform analysis while maintaining access to the data source.
- “Complete Audit Log”: Record every tool call and decision in Claude Console
This marks the transition from “complete autonomy” to “managed autonomy” in the design concept - AI does not completely replace humans, but enhances them.
Summary: Structural changes in cutting-edge AI
The deployment of Claude Opus 4.7 with financial services agents marks three structural changes in cutting-edge AI:
1. From “Consumer Application” to “Enterprise Revenue”
- The era of consumer app “turf grabbing” is over
- Enterprise contracts offer high margins, multi-year commitments, and deep integration
- Usage justifies computational expenditures
2. From “Tools” to “Operation Layer”
- No longer just a “software vendor”
- but the infrastructure layer (Claude runs inside an Office app)
- Data platform layer (embedded by Moody’s, Dun & Bradstreet, etc.)
- Operations layer (agent automatically executes the complete workflow)
3. From “Assistance” to “Automation”
- Finance is 6-12 months behind programming
- Structural shift from “human analysts” to “AI analysts”
- The “ladder” of autonomy: Basic → Intermediate → Advanced
These three changes together mark that cutting-edge AI is moving from “tools” to “operations layer”, from “assistance” to “automation”, and from “consumer applications” to “enterprise revenue”.
Specification of technical issues
1. Changes in tokenizer
Opus 4.7 uses an updated tokenizer, and the same input may be mapped to 1.0–1.35× more tokens. This is for better semantic representation, but means that developers need to re-evaluate the cost of token usage.
2. Capability degradation design
Cyber has lower capabilities than Mythos Preview and is intentionally designed for security control:
- Ability downgrade during training
- Equipped with automatic detection and interception mechanism
- Selectively release high-risk capabilities
3. Cost structure changes
- Inference Cost: At high effort levels, later rounds in the agent setting do more inference
- token cost: The new tokenizer may increase 0.0–0.35× token usage
- Computational Spend: Enterprise usage justifies the computational spend of AI models
Structural Challenges
1. Organizational absorptive capacity
- Technical capabilities far exceed the organization’s current absorptive capacity
- Need to rewrite business processes, cultivate talents, and change work methods
- The challenge is not the technology itself, but the absorptive capacity of the organization
2. Structural changes in employment
- AI is now used for at least a quarter of about three-quarters of U.S. jobs
- Projected to increase U.S. labor productivity by 1.8 percentage points annually over ten years
- The duality of employment implications: Becoming a stronger expert and a better question asker
3. Policy lag
- Policy lags by two to three years
- Technology advances so rapidly
- Need to prepare for pessimistic scenarios and set policies
Construction of operation layer
The deployment of Claude’s financial services agent marks that cutting-edge AI is becoming the operational layer of the financial industry:
- Infrastructure layer: Claude runs within Office applications
- Data platform layer: Moody’s, Dun & Bradstreet, Fiscal AI, etc. are embedded in the Claude interface
- Operations Layer: Agents automate the complete workflow from research to closure
- Management: Humans review, iterate, and approve Claude’s work
This marks that cutting-edge AI is moving from “tool” to “operation layer”, from “assistance” to “automation”, and from “consumer application” to “enterprise revenue”.