Claude’s New Finance Agent Templates: What Non-Technical Professionals Should Actually Care About

What actually launched on May 5

On May 5, 2026, Anthropic released ten ready-to-run agent templates for the most time-consuming work in financial services: building pitchbooks, screening KYC files, reviewing earnings, and closing the books at month-end. [1]

If you’ve been hearing the phrase “AI agents” thrown around for the last 18 months and quietly wondering when it would mean something concrete, this is one of the more concrete examples to date. These aren’t demos. They’re working templates that ship as plugins inside Claude Cowork, Claude Code, and as cookbooks for Claude Managed Agents. [1]

Each one is what Anthropic calls a “reference architecture.” That’s a fancy phrase for “here’s a working starting point you can copy and adjust to fit your team.” It saves the months of trial and error a finance team would otherwise spend trying to get an AI agent to behave reliably on real work.

The 10 agents, in plain English

The templates split into two groups of five. [2] Here’s what each one actually does, stripped of jargon.

Research and client coverage agents

1. Pitch builder. Takes a client name, an industry, and a meeting agenda, then drafts the deck. Pulls in market data, builds the slides, and flags the sections that need a human eye before the meeting.
2. Meeting preparer. Reads everything available about a client (filings, news, prior meeting notes) and produces a one-page brief: what they care about, what changed since last time, what to ask.
3. Earnings reviewer. Listens to or reads earnings calls, summarises the analyst Q&A, and extracts the surprises that move price targets.
4. Model builder. Builds a financial model from a set of assumptions and historical data. Drops it into Excel via the Microsoft 365 add-in.
5. Market researcher. Goes wide on a topic (a sector, a geography, a regulatory shift), pulls together sources, and produces a structured brief with cited links.

Finance and operations agents

6. Valuation reviewer. Checks a draft valuation against comparables, methodology choices, and recent transactions. Flags inconsistencies for a human to confirm.
7. General ledger reconciler. Matches transactions across accounts, surfaces discrepancies, and proposes journal entries.
8. Month-end closer. Walks through the close checklist: accruals, intercompany matches, variance commentary. Produces the binder of documentation auditors expect.
9. Statement auditor. Cross-checks financial statements against source data, prior periods, and disclosure requirements.
10. KYC screener. Reviews onboarding documents against watchlists, sanctions lists, and risk policies. Flags edge cases for a compliance officer to clear.

If you don’t work in finance, three or four of these will sound vaguely familiar from corporate function-speak. The point isn’t to learn the templates. It’s to notice the pattern Anthropic chose: every single one of these agents replaces work that’s structured, repetitive, and partially deterministic. Not creative work. Not strategy. The grindy stuff that fills calendars and burns out senior staff.

Why Anthropic started with finance

Bloomberg called it a strategic move into one of Anthropic’s most lucrative customer segments. [3] That’s true and not very interesting. The more useful read: finance is where the pain is most measurable.

A junior banker building a pitchbook spends 40 to 60 hours on slides that are 80% template. A KYC analyst spends a third of her week clicking through documents that almost always pass. A controller’s team works through three days of weekends every month-end. There’s a lot of slack in the system, and finance teams measure their own time in basis points. When you tell a CFO an agent will save 15 hours a week per analyst, the math is automatic.

For the rest of us, the lesson is simple: AI agents land first where the work is repetitive, the cost of mistakes is bounded, and the volume is high. Look at your own function and ask which tasks check those three boxes. That’s where your version of these agents will land first.

The three-part recipe behind every agent

Here’s the part you should actually screenshot and keep. Anthropic published the architecture every one of these agents uses. [2] It’s three pieces:

1. Skills

The instructions and domain knowledge for the task. For the pitch builder, skills include things like “how a pitchbook is structured,” “what a comparable company set looks like,” and “what the firm’s house style for slides is.” Skills are reusable text files that tell the agent how the work should be done.

You can write skills for almost any function. A marketer’s skill might be “how we write a campaign brief.” An HR director’s skill might be “how we screen senior leadership candidates.” If you can document the way good work gets done in your team, you can write a skill.

2. Connectors

Governed access to the data the task runs on. The KYC screener has a connector to the sanctions database. The earnings reviewer has a connector to your firm’s research repository. Connectors are how an agent gets to the real systems it needs without you giving it your password and praying.

For non-finance teams, this means the agent isn’t useful until it can see the systems you live in. Salesforce, your CMS, Notion, your file share. Anthropic, OpenAI, and Microsoft all ship connector libraries now. Pick the one that talks to most of your stack. (We covered the Microsoft Copilot vs. Gemini comparison here if you’re weighing the options.)

3. Subagents

Additional Claude models the main agent calls for sub-tasks. The pitch builder might call a subagent specifically for picking comparables. Another for doing methodology checks. The main agent stays focused on the overall task; subagents handle the parts that need their own context window or specialised behaviour.

This is the part most professionals miss. A good agent isn’t one giant prompt. It’s a small team of specialists with one orchestrator. Architecturally that mirrors how human teams already work: a senior leads the project and pulls in specialists when needed.

How non-finance teams can borrow the same pattern

This is where it gets interesting. The 10 templates aren’t going to be much use to a marketing manager or an operations lead directly. But the pattern absolutely is. Here are four examples of agents you could build using the exact same recipe:

For marketing: a campaign post-mortem agent

Skills: your campaign brief template, your performance benchmarks, the questions your CMO always asks at debriefs. Connectors: Google Analytics, Meta Ads Manager, your CRM. Subagents: one for creative analysis, one for media performance, one for competitive context.

You’d put a campaign URL in, get back a structured post-mortem with the things you’d usually spend a Friday afternoon assembling.

For HR: a candidate brief agent

Skills: your scorecard template, the seniority level definitions, your firm’s interview rubric. Connectors: your ATS, LinkedIn (where allowed by your contract), the candidate’s portfolio. Subagents: one for a public footprint summary, one for skills mapping against the job requirements.

For consulting and project work: a client status agent

Skills: your project plan template, your status report format, the typical risks for the engagement type. Connectors: your time tracking, your project management tool, the client’s shared drive. Subagents: one for risk identification, one for budget burn analysis.

For operations: an SOP compliance agent

Skills: the standard operating procedure document, the exception handling rules. Connectors: your ticketing system, the relevant data warehouses. Subagents: one for anomaly detection, one for root cause analysis.

None of these require a developer. With Claude Cowork or ChatGPT’s custom GPTs, a non-technical person who’s done the work to write good skills documents can stand up a v1 in a week. (We covered building custom GPTs without code here.)

What to do Monday morning

Three concrete moves, in order:

1. Pick one task you do every week that follows a template. Not the most strategic thing on your plate. The most repetitive thing. Status reports, candidate screens, client briefs, weekly performance recaps. Anything where you keep ending up at the same shape of output.

2. Write the skill document. One or two pages, no more. Headed sections: what the input looks like, what the output looks like, the rules a good version follows, the rules it never breaks. This is the highest-leverage two hours of work you can do this week. It’s also useful even if you never build the agent. You’ve now documented institutional knowledge.

3. Run it manually with Claude or ChatGPT first. Paste your skill document, paste a real input, see what the output looks like. Iterate on the skill until the output is consistently usable. Once you have that, you have the prompt half of an agent. Connectors and subagents come later.

Most professionals stall on step 2 because it forces you to admit that the way you do work isn’t actually written down anywhere. That’s the work. The agent is just the multiplier on top.

The bigger picture: when Anthropic publishes a recipe, it tends to become the way the industry builds for the next two years. Last year it was prompt patterns. This year it’s the skills + connectors + subagents pattern. Worth learning even if you never touch a finance template.