Claude Finance Agents: What the Ten Templates Do and Cost

Anthropic now sells ten ready-to-run finance agents that build pitchbooks, reconcile general ledgers, screen KYC files, and close the books at month-end. The templates are open source and free; running them is not, and the best model available still gets roughly one in three analyst-style questions wrong on the standard benchmark. This guide covers what each agent actually does, what a real deployment costs, and how to pilot one without creating a compliance problem.

What Anthropic actually shipped

In May 2026, Anthropic released ten agent templates for financial services work, packaged as plugins for Claude Cowork and Claude Code, and as cookbooks for Claude Managed Agents. The whole set lives in a public GitHub repository, anthropics/financial-services, under an Apache 2.0 license.

Each template bundles three things: skills (the instructions and domain knowledge for the task), connectors (governed access to the data the task runs on), and subagents (additional Claude calls for sub-tasks like comparables selection or methodology checks). You are expected to adapt them to your own modeling conventions, risk policies, and approval flows.

The ten agents, grouped by who uses them:

Research and client coverage

• Pitch builder creates target lists, runs comparables, and drafts pitchbooks.
• Meeting preparer assembles client and counterparty briefs before calls.
• Earnings reviewer reads transcripts and filings, updates models, and flags thesis-relevant changes.
• Model builder creates and maintains financial models (DCF, LBO, three-statement, comps) in Excel.
• Market researcher tracks sector and issuer developments across news, filings, and broker research.

Finance and operations

• Valuation reviewer checks valuations against comparables, methodology, and your review standards; it ingests GP packages and stages LP reporting.
• General ledger reconciler reconciles GL accounts and runs net asset value calculations against the books of record.
• Month-end closer runs the close checklist, prepares journal entries, and produces close reports.
• Statement auditor reviews financial statements for consistency, completeness, and audit-readiness.
• KYC screener assembles entity files, reviews source documents, and packages escalations for compliance review.

The agents draw on a connector ecosystem that now includes FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG, Daloopa, Chronograph, Dun and Bradstreet, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, and Verisk, plus a Moody's MCP app covering credit ratings and data on more than 600 million companies. Claude add-ins for Excel, PowerPoint, and Word are generally available, so a model built in Excel can flow into a deck without re-explaining the work. The launch extends the job-specific agent strategy we covered in Anthropic's enterprise agents push.

Plugin or Managed Agent: two ways to run the same template

The same template runs in two modes, and choosing between them is the first real decision.

As a plugin in Cowork or Claude Code, the agent works alongside an analyst on their desktop. You hand the Pitch builder a target list and get back a comps model in Excel, a pitchbook draft in PowerPoint, and a cover note. The human drives; the agent drafts. This mode is included on all paid Claude plans.

As a Claude Managed Agent (in public beta on the Claude Platform), the same template runs autonomously on a schedule: a nightly reconciliation across a whole book of funds, or a KYC sweep across every new onboarding file. Managed Agents add the infrastructure a firm would otherwise build itself: long-running sessions that survive a multi-hour deal close, per-tool permissions, managed credential vaults, and a full audit log in the Claude Console where compliance can inspect every tool call and decision.

The practical split: start in plugin mode to validate output quality on your own data, then graduate the workflows that prove themselves to Managed Agents for scheduled, unattended runs. In both modes, outputs are staged for human review before anything goes to a client, gets filed, or is acted on.

What a real deployment costs

The templates are free. Everything else has a meter. Here is an honest cost build-up for a pilot at a fund operations team of eight people, using verified prices.

Seats. The plugins run on any paid plan. A Team Standard seat costs $25 per user per month billed monthly, or $20 billed annually, according to Anthropic's Team plan documentation. Eight seats on annual billing: $160 per month. Enterprise adds SSO, SCIM, audit logs, and custom data retention at custom pricing, per the Claude pricing page. Regulated firms will likely end up on Enterprise for the audit controls, not the seats.

Model usage. Claude Opus 4.7, the model Anthropic pairs with these agents, costs $5 per million input tokens and $25 per million output tokens at API rates, per Anthropic's platform pricing. One budgeting wrinkle from the same docs: Claude 4.7 and later use a new tokenizer that produces about 30% more tokens for the same text, so cost estimates based on older models will come in low.

Managed Agent runtime. Sessions carry a per-session-hour runtime charge on top of tokens. For a nightly reconciliation that runs 90 minutes across 22 business days, the runtime line comes to a few dollars a month. Runtime is noise; tokens are the bill.

An illustrative monthly total. Say the nightly GL reconciler processes 2 million input tokens and produces 400,000 output tokens per run (a moderate workload with prompt caching on reference documents). That is 44 million input and 8.8 million output tokens a month: roughly $220 plus $220, or $440. Add seats and runtime and the whole pilot lands under $700 per month. If you are weighing these templates against a custom build, our AI agent development cost guide covers the build-it-yourself economics. For comparison, Ledge's 2025 close benchmark found finance teams spend 20 to 50 hours a month on cash reconciliation alone; even at a conservative $50 per hour fully loaded, that is $1,000 to $2,500 of staff time on one task the reconciler targets.

The hidden line item: data. The connectors assume you already pay for FactSet, PitchBook, Moody's, or S&P Capital IQ. Those subscriptions dwarf the AI costs. If your firm does not hold market-data licenses, the research agents lose most of their value, and the operations agents (reconciler, closer, KYC screener) become the better entry point because they run on your own documents.

One more honest caveat from the Vals AI benchmark: hard analyst queries can cost more than $5 each on frontier models because the agent makes many tool calls per answer. Costs scale with difficulty, not just volume.

The accuracy number you need to plan around

Claude Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37% accuracy. That benchmark, built with Stanford researchers and a global systemically important bank, tests 537 questions at the level of an entry-level analyst: retrieval, market research, projections.

Read that number both ways. It is the best score any model has posted, and when the benchmark launched in 2025 the leader scored 46.8%, so capability is climbing fast. It also means the best available model gets about one in three analyst-style questions wrong. Models perform best on simple quantitative and qualitative retrieval, exactly the work that eats analyst hours, and worst on multi-step reasoning and projections.

Two design consequences follow. First, every agent output needs a named human reviewer with the time and skill to catch errors; the benchmark tells you errors will be frequent. Anthropic's own repository states the agents do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding, and that every output is staged for professional sign-off. Treat that as architecture, not boilerplate. Second, put the agents on retrieval-heavy, checkable work first (reconciliation breaks, document assembly, comps pulls) where verification is cheap, and keep judgment-heavy work (valuation calls, credit decisions) human-led with agent support.

Which agents to pilot first

Not all ten are equal starting points. A sensible order for a fund or asset manager:

1. General ledger reconciler. Breaks are factual, outputs are checkable against the books of record, and the pain is measured: half of finance teams take more than five business days to close, per the Ledge survey cited above. For the accounting-firm angle on the same workflow, see our guide to AI agents for accounting automation.
2. Meeting preparer. Low risk, immediately visible time savings, easy to judge quality.
3. KYC screener. High volume, document-heavy, and it packages escalations rather than deciding them, which keeps the compliance officer in charge.
4. Earnings reviewer or market researcher if you hold the data subscriptions to feed them.
5. Month-end closer and statement auditor last, once the reconciler has proven itself, because these touch the numbers your auditor will ask about.

The pitch builder and model builder are strong for banks and advisory teams but are the hardest to evaluate objectively, which makes them poor first pilots.

When these agents are the wrong choice

Be suspicious of any vendor who skips this section. These agents are the wrong choice when:

• You do not hold the data licenses. Without FactSet, PitchBook, or similar subscriptions, the research agents have little to connect to.
• Your books are simple. A two-fund manager with clean books and a good administrator will not recover the setup effort. A spreadsheet and a competent accountant win.
• You need the model to decide, not draft. Anything that must be deterministic and auto-approved (payment release, trade execution, onboarding approval) is outside what these agents are designed or licensed to do.
• Your data cannot leave your environment. Claude runs in Anthropic's cloud (or AWS and Azure marketplaces). If your regulator or security policy requires fully private inference, the path is a privately deployed open-weight model instead. That is a different build: we fine-tune and deploy open-weight models for banks and insurers when Claude's hosted model is off the table.
• You want fine-tuned Claude. Anthropic does not offer customer fine-tuning of Claude models. You adapt behavior through skills, prompts, and connectors, not weights.

Common mistakes to avoid

• Reading the benchmark as a production guarantee. 64.37% is an average over 537 curated questions, not your data. Measure accuracy on your own documents before scaling.
• Piloting all ten agents at once. You will not be able to tell which one works. One workflow, one baseline, one review loop.
• Skipping the baseline. Record hours spent and error rates on the manual process first, or you will never prove the pilot worked.
• Giving the agent write access to systems of record. Stage outputs for review. The repository is explicit that agents do not post to ledgers.
• Budgeting tokens on old assumptions. The 4.7-series tokenizer inflates token counts roughly 30% versus earlier models; US-only inference adds another 1.1x multiplier.
• Ignoring the audit log until audit season. Configure the Claude Console audit logging and retention on day one, because your compliance team will ask for it eventually.

A 30-day pilot plan

• Week 1: pick and baseline. Choose one workflow (GL reconciliation is our default recommendation). Log current hours, cycle time, and error rates for two weeks of history.
• Week 2: plugin mode. Install the agent plugin from the financial services marketplace, connect read-only data, and run it alongside the manual process. Compare outputs line by line.
• Week 3: tune. Adjust the skills to your conventions: account mappings, materiality thresholds, escalation format. Expect several iterations; this is normal, not a failure.
• Week 4: measure and decide. Quantify hours saved and exception rate. If accuracy on your data clears your bar, scope a Managed Agent deployment for scheduled runs with the audit log and per-tool permissions configured. If it does not, you have spent under $1,000 to learn that, which is a cheap answer.

This is the implementation work we do under our Claude for Financial Services practice: connector setup, skills adaptation to your policies, review workflow design, and the governance documentation your risk team needs. If your team also wants Claude tuned into the rest of the desktop workflow, our Claude implementation service covers Cowork, Claude Code, MCP servers, and custom skills.

Frequently asked questions

Are the Claude finance agents free? The templates are open source under Apache 2.0 and free to adapt. You pay for Claude seats, model usage, Managed Agent runtime, and any underlying data subscriptions. A focused pilot typically runs under $1,000 per month in platform costs.

Do I need the Enterprise plan? No. The plugins work on all paid plans, including Pro and Team. Enterprise becomes relevant when you need SSO, SCIM, audit logs, and custom data retention, which most regulated firms will.

How accurate are the agents? The best model, Claude Opus 4.7, scores 64.37% on the Vals AI Finance Agent benchmark, strongest on retrieval tasks. Plan for human review of every output and measure accuracy on your own documents during a pilot.

Can the agents post to our general ledger or approve onboarding? No. Anthropic's repository explicitly scopes the agents to drafting work product for professional review. They do not post, execute, bind, or approve.

Do the agents work with our fund administration or accounting system? The shipped connectors target market data and research providers. Connecting an internal GL, fund admin platform, or data warehouse is custom work, usually via MCP servers, which is a standard part of an implementation engagement. For the accounting angle, see our guide to AI agents for accounting automation.

How does this relate to Anthropic's broader enterprise push? The finance agents are one vertical of a wider plugin strategy we covered in Anthropic's Claude enterprise plugins and the workplace AI contest with Microsoft.

Where to go from here

The finance agents are real, cheap to try, and genuinely limited: they draft and reconcile well, they decide nothing, and a third of hard analyst questions still come back wrong. A disciplined one-workflow pilot tells you within a month whether the economics work on your data. If you want that pilot designed, instrumented, and documented for your risk team, talk to us about our Claude for Financial Services work.

Related reading

• AI Reconciliation for Fund Administration: Costs and Limits