Anthropic's Finance Agents: Not 10 Templates, but a Deliverable Workflow Operating System

Anthropic's Finance Agents: Not 10 Templates, but a Deliverable Workflow Operating System

中文 EN

Anthropic's May 5 release, Agents for financial services, looks on the surface like 10 finance agent templates: pitch builder, meeting preparer, earnings reviewer, model builder, market researcher, valuation reviewer, general ledger reconciler, month-end closer, statement auditor, and KYC screener.

But the more important move is this: Anthropic is packaging one of the hardest types of AI work to sell, deploy, and audit into something that looks like a deliverable operating system.

Not "Claude can help you make an investment banking deck."

More like: take part of the work of a junior analyst, an associate, a controller, or a compliance reviewer; break it into a describable process; connect it to data sources; assign tool permissions; delegate pieces to subagents; and produce the artifacts finance teams actually pass around: Excel, PowerPoint, Word, exception reports. Every artifact then has to be reviewable, traceable, and auditable.

That is the step from agent demo to enterprise system.

This is not a prompt library. It is a reference architecture.

Anthropic describes each agent template as a bundle of three things.

First, skills: task instructions, domain knowledge, and workflow methods. Second, connectors: governed access to data. Third, subagents: additional Claude models that the main agent can call for specialized sub-tasks such as comparables selection or methodology checks.

That three-part structure matters. Finance does not lack smart models. It lacks a reliable skeleton that can turn "I need a pitchbook" into a controlled chain of actions.

I pulled down Anthropic's official financial-services marketplace repo. The repo is more revealing than the announcement. It contains agent plugins, vertical plugins, Managed Agent cookbooks, and MCP connector definitions in one place.

The README groups the 10 agents into finance's real work map:

  • Coverage and advisory: Pitch Agent, Meeting Prep Agent
  • Research and modeling: Market Researcher, Earnings Reviewer, Model Builder
  • Fund admin and finance ops: Valuation Reviewer, GL Reconciler, Month-End Closer, Statement Auditor
  • Operations and onboarding: KYC Screener

That is the important product move. Anthropic does not begin with "what can the model do?" It begins with "which expensive, repetitive, format-bound workflows do financial firms run every day?"

Pitch Agent: model, deck, QC

The clearest example is the Pitch Agent system prompt.

It defines the agent as a "senior investment banking associate." The input is a target company and a strategic situation. The output is not a chat answer. It is two artifacts.

The first is an Excel valuation workbook with trading comps, precedent transactions, DCF, and a football-field summary. Every output cell must be a live formula traceable to an input.

The second is a pitch deck on the bank's PowerPoint template: situation overview, company snapshot, valuation summary, comps detail, precedents detail, and illustrative process. Every chart must be tied back to the Excel model.

The point is not that the agent can "write slides." The pain in an investment banking pitch is keeping the deck, model, source data, footnotes, dates, and valuation methods internally consistent. Anthropic writes that work into the agent contract: scope the ask, identify comps and precedents, pull CapIQ data and filings, build comps, LBO, DCF, and three-statement models, generate the football field, populate the deck, and run deck QC.

If you have ever produced analysis that someone else will act on, you know the first draft is rarely the expensive part. The expensive part is "where did this number come from?", "why does page 18 not match the model?", and "is that footnote date wrong?" Anthropic puts that unglamorous work into the workflow.

The guardrails are finance-native: no external communications, cite every number, mark unavailable sources as [UNSOURCED] instead of inventing them, stop for banker review after the Excel model, and stop again after the deck.

This is not an autonomy fantasy. It is human approval written into the process.

KYC Screener: finance agents are about isolation, not freedom

The KYC Screener shows the security philosophy even more clearly.

The agent reads an onboarding packet, extracts legal names, beneficial owners, addresses, identifiers, and document inventory; runs KYC/AML rules; screens for sanctions, PEP, and adverse media hits; and produces an escalation packet.

The guardrails are the important part:

  • onboarding documents are untrusted;
  • the doc-reader worker has only Read/Grep and no MCP access;
  • the orchestrator never writes;
  • only the escalator subagent holds Write;
  • the agent recommends a risk rating, but the compliance officer decides.

That is the core design principle for finance agents: do not maximize freedom. Minimize unnecessary authority. Treat external documents as untrusted. Concentrate write access at the end of the workflow. Keep decisions with accountable humans.

This is almost the opposite of the generic agent demo where the agent browses, fills forms, sends messages, and does whatever it wants. A production finance agent needs bounded autonomy.

It has to do work. But every step needs to answer: who gave it the data, which tool did it call, could it write, where did it write, and who approved the result?

Claude Managed Agents: from Cowork to schedulable back-office processes

Anthropic offers two deployment paths.

The first is plugin mode, running inside Claude Cowork or Claude Code alongside an analyst's workspace. Anthropic's examples include a Pitch Agent that returns an Excel comps model, a PowerPoint pitchbook, and an Outlook cover note.

The second is Claude Managed Agents, running on Claude Platform for background workflows such as an entire book of deals, nightly schedules, or long-running close processes. Anthropic says the Managed Agent cookbooks support long-running sessions, per-tool permissions, managed credential vaults, and a full audit log in Claude Console so compliance and engineering teams can inspect tool calls and decisions.

The Pitch Agent Managed Agent cookbook makes this concrete. It uses claude-opus-4-7, connects CapIQ and Daloopa MCP servers, and defines callable agents: researcher, modeler, and deck-writer. The README states that only the deck-writer leaf agent holds Write; the researcher and modeler primarily read data, build intermediate artifacts, and hand work forward.

This may be Anthropic's most interesting enterprise agent architecture right now. The main agent does not need to do everything. It acts more like an orchestrator. Specialized leaf agents receive limited tools, limited data, limited write permissions, and a specific job.

In enterprise software language, this is close to an agentic workflow engine: process, roles, tool ACLs, data connectors, artifact output, and audit.

Microsoft 365 integration: agents enter the native artifact environment

Anthropic also announced that Claude can work through Microsoft 365 add-ins inside Excel, PowerPoint, and Word, with Outlook coming soon. It says context can flow across applications, so a model created in Excel can carry into a PowerPoint deck without being re-explained.

That matters more than it sounds.

Finance work does not end in a chat window. Investment banking, private equity, research, finance, risk, and compliance teams produce Excel workbooks, PowerPoint decks, Word memos, Outlook threads, PDF statements, and data-room files.

The competition is not "who can answer like an analyst?" It is "who can stand on the artifact production line?"

Excel is the production environment for models. PowerPoint is the production environment for client communication. Word is the production environment for memos and policy. Outlook is the production environment for task flow. If Claude can maintain context across these surfaces, the agent stops being an assistant next to the work and starts becoming part of the work itself.

MCP connectors: data rights are the moat

Anthropic's data layer is just as important. The announcement lists FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, Daloopa, and others. It also adds connectors for Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, Verisk, and more. Moody's also launched an MCP app that brings proprietary credit ratings and data on more than 600 million public and private companies into Claude.

The repo's financial-analysis MCP config turns that into implementation detail: Daloopa, Morningstar, S&P Global, FactSet, Moody's, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, and Egnyte are all MCP servers.

The business implication is clear. Finance agent performance will not be decided by the foundation model alone.

The model must be strong. But the real questions are:

  • Can it legally access the data?
  • Can it preserve data-provider entitlement?
  • Can each data query leave an audit trail?
  • Can internal CRM, research repositories, and data warehouses be connected?
  • Can the same agent behave differently under each firm's policy?

That is why financial-data and software stocks sold off after the announcement, according to Bloomberg coverage cited in market reporting. Investors were asking a simple question: if model companies connect the data layer, workflow agent, and Office artifact production line, how much front-end value shifts away from terminals and traditional workflow software?

I do not think data providers disappear. They still control core data, licensing, trust, and lineage. But the user entry point may move. Instead of opening a terminal, looking up data, and moving it into Excel by hand, an analyst may ask an agent to query MCP sources behind the scenes and return an auditable workbook and deck.

When the entry point moves, bargaining power moves with it.

The benchmark: finance agents are not "done," but they are valuable enough to deploy

Anthropic cites the Vals AI Finance Agent benchmark, where Claude Opus 4.7 leads with 64.37% accuracy.

That number should not be misread as "Claude can replace a financial analyst." The Vals benchmark has 537 questions designed around entry-level financial analyst research tasks: SEC filings research, simple retrieval, market research, projections, adjustments, beat/miss analysis, financial modeling, and market analysis. Agents receive EDGAR search, Google search, an HTML parser, and a retrieval tool, then are graded on final answer accuracy.

Three points matter.

First, 64.37% is not a graduation score. High-risk financial work cannot move to no-review mode because a benchmark leader cleared the field. Anthropic's own repo repeats the right disclaimer: these agents draft analyst work product. They do not give investment, legal, tax, or accounting advice; execute transactions; approve onboarding; or remove human sign-off.

Second, Vals notes that top performers tend to use more tool calls. Finance agent capability is not just "the model knows more." It is whether the agent can query EDGAR, parse filings, retrieve data, and return to the source before answering. That matches real analyst work. A lot of the job is not inspiration. It is process discipline and data handling.

Third, cost becomes product design. Vals notes that some hard queries can cost more than $5 on models such as o3 or Claude Opus 4.1 Thinking. Finance has budget, but not infinite token budget. Production agentic workflows need to decide where to use frontier models, where to use cheaper models, what to cache, and which intermediate results should persist.

This is another reason templates matter. Templates are not just demo packaging. They are a container for cost control.

Market context: Anthropic wants Wall Street's AI operating layer

This release did not happen in isolation.

Business Insider reported that financial services is Anthropic's second-largest enterprise-revenue vertical after technology. It also reported that 40% of Anthropic's top 50 customers are in finance. Axios quoted Anthropic financial-services product lead Nicholas Lin saying the goal is to reduce deployment from months to days.

One day earlier, Anthropic announced a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs. Anthropic Applied AI staff and the new company's engineers will work with customers to transform core operations.

Put the two announcements together and the strategy is clear.

One side templates the most common high-value financial workflows so large institutions can install plugins, run Managed Agents, and tune policy themselves. The other side uses a services company to help firms without enough engineering resources put Claude into operations.

This looks like a hybrid of Palantir, Accenture, Bloomberg, Microsoft Office, and a model company. It is not pure SaaS. It is not pure API. It is a stack of model capability, data connectors, desktop software, engineering deployment, and compliance controls.

Competition will not only come from OpenAI. Finance-native startups such as Rogo and Hebbia are already working on pitch decks, research, meeting prep, and model building. Their edge is not the model alone. It is domain expertise, data integration, customer workflow, and model-agnostic routing.

So Anthropic's hard problem is not whether Claude is strong enough. It is whether these workflows can land inside each bank's systems, data permissions, approval flows, deck templates, and modeling conventions instead of remaining impressive demos.

The core judgment: the unit of a finance agent is not chat. It is the artifact.

The biggest lesson from this release is that Anthropic moves the agent output unit from "answer" to "reviewable work product."

Pitch Agent delivers a workbook and a deck. KYC Screener delivers an entity file, rules-engine result, screening result, and escalation packet. GL Reconciler delivers a break list, root-cause trace, and exception report. Model Builder delivers a linked Excel workbook. Statement Auditor delivers an audit-ready review.

These outputs share the same properties:

  • fixed format;
  • data sources;
  • intermediate calculations;
  • approval points;
  • error cost;
  • compliance requirements;
  • downstream users.

That is more practical than a generic agent that can "do anything on the internet." Enterprises do not buy autonomy for its own sake. They buy throughput, consistency, and traceability.

Looking through the repo's agent prompts, I had one strong feeling: this is less like an AI writing tool and more like workflow-as-code. Finance know-how used to live in people, templates, Slack threads, Excel workbooks, and approval habits. Now some of it is being written into markdown skills, agent YAML, MCP config, and tool permissions.

If this direction works, every industry will get its own agent marketplace. But the valuable unit will not be the label "legal agent," "healthcare agent," or "finance agent." The valuable unit will be whether the workflow is auditable, data-connected, and capable of producing a usable artifact.

The risk: errors become more dangerous when they look official

The risks are also direct.

First: source hallucination and numeric error. In finance, one wrong multiple, share count, guidance midpoint, or non-GAAP adjustment can distort an entire deck. Anthropic's templates require every number to be citable and unavailable sources to be marked [UNSOURCED]. Necessary, but not sufficient.

Second: data permissions. MCP connectors make cross-source querying convenient, but every query touches licensing, entitlement, client confidentiality, and internal data separation. If agents can read across systems more easily than humans, they need stricter policy enforcement than humans.

Third: responsibility laundering. Once an agent produces a deck, memo, or KYC escalation packet, who owns it? Anthropic's repo keeps saying human sign-off. That needs to stay. Financial firms cannot let "AI suggested it" become a grey zone where nobody is accountable.

Fourth: organizational friction. Connecting an agent to Excel, PowerPoint, Outlook, and data sources does not mean analysts trust it. Finance quality control contains a lot of tacit knowledge: which MD hates which chart, which sponsor case is too aggressive, which company habitually sandbags guidance, which footnote legal always flags. Firm-specific skills will need to absorb that knowledge over time.

Builder takeaway

The lesson is not "go build a finance agent."

The real lesson is that if you want to sell agents into high-value industries, the product should look more like this:

  1. Start from a frequent, expensive, format-bound workflow, not from chat.
  2. Write domain knowledge as skills, not as hidden prompt soup.
  3. Connect data through governed connectors, not copy-paste.
  4. Split sub-tasks into subagents, but give each only the necessary permissions.
  5. Produce the artifact the industry actually uses: workbook, deck, memo, report.
  6. Put human approval into the workflow instead of treating it as a product failure.
  7. Treat audit logs, source trace, and cost control as core features, not enterprise add-ons.

Anthropic is not selling 10 templates. It is selling a new enterprise AI delivery format: model + skill + connector + subagent + artifact + audit.

Finance is the first battlefield because the data is expensive, the processes are expensive, the errors are expensive, and the labor is expensive. If an agent can reliably shorten even one part of the workflow, the ROI is high enough for firms to tolerate the deployment cost.

Next, I am watching two things.

First, how deeply these templates get modified inside real financial institutions. If every firm has to rewrite the prompts, connectors, and approval flows, Anthropic is mainly selling a framework. If most firms only tune the templates, it becomes closer to a marketplace.

Second, how Rogo, Hebbia, Bloomberg, FactSet, S&P, Moody's, and other workflow/data players respond. Model companies are moving up into financial workflow. Data companies and vertical SaaS companies will move down into the agent layer. The winner may not be the company with the strongest model. It may be the company that controls the artifact production line and the data lineage.

References

Related

When AI Agents Meet OS Design: A Deep Dive into Anthropic Managed Agents Architecture
When 90% of the code is written by AI: Mike Krieger (Anthropic CPO) on what changes next
AI Closes Its Doors: The April 2026 Privatization Inflection