Last week, Anthropic put Jamie Dimon on a stage in Manhattan and announced a suite of AI agent templates for financial services. Two days later, they launched a separate package for small businesses. The headlines were big. “The operating layer for Wall Street.” AI walking through the doors of global finance.
I’ve been watching these releases closely. And I want to give you an honest read on what’s actually here — because I think the hype and the reality are sitting pretty far apart right now.
These are useful tools. Genuinely interesting. But they are small pieces of a very large puzzle. They are not rewriting the rules of accounting and finance.
What Anthropic Actually Launched
On May 5th, Anthropic launched 10 pre-built agent templates for financial services: a GL reconciler, a month-end closer, a statement auditor, a KYC screener, a pitch builder, and more. Then on May 13th, they dropped 15 agentic workflows for small businesses built around tools like QuickBooks, PayPal, and HubSpot — covering payroll planning, cash flow forecasting, and the monthly close.
Each template is what Anthropic calls a “reference architecture” — a starting point, not a finished product. It bundles three things: skills (markdown files with instructions and domain knowledge), connectors (governed access to external data like Moody’s, QuickBooks, Dun & Bradstreet), and sub-agents (focused Claude models called in for specific sub-tasks).
Importantly, all of these are publicly available on GitHub. You can go read the GL reconciler skill right now. It’s a markdown file. Plain English instructions telling the agent what steps to follow. That transparency is actually useful — if you want to understand what good finance-domain prompts look like, these are great examples.
What They Do — And Where They Stop
This is the part that gets glossed over in the coverage. Let me be specific.
The GL reconciler? It matches settlements against your general ledger, flags mismatches, and stages a close packet for you to review. It does not post to your ledger. Nothing moves downstream without a human sign-off.
The month-end closer? It assembles data, writes a plain-English P&L summary, and exports a packet. It does not close your books. You close your books.
The statement auditor? It reviews financial statements for consistency and flags potential issues. It does not issue an audit opinion.
The small business close workflow? Anthropic’s own description says it clearly: “Claude does the work; you approve before anything sends, posts, or pays.”
Why is everything staged for human review? Partly liability. But mostly because it has to be. The best model Anthropic has — Claude Opus 4.7, specifically positioned for financial work — scores 64% on the Vals AI Finance Agent benchmark. That’s a 36% error rate. A human who scored 64% on a professional competency exam wouldn’t pass. The review requirement isn’t a formality. It’s structurally necessary.
Where the Real Value Is
I don’t want to dismiss these tools, because there is genuine value here. I think about it as the jump-off point problem.
Getting from zero to a structured reconciliation file, a first-pass close packet, or a formatted credit memo is often the most time-consuming part of the work. Not intellectually hard. Just slow and manual. These agents handle that scaffolding. You’re not staring at a blank spreadsheet — you’re reviewing a structured draft.
For a small business, that might do a lot of the heavy lifting on its own. For a larger accounting or finance department, it gives you a foundation you can customize into something more niche and purpose-built. Either way, the value is real. It’s just not the same thing as replacing the accountant.
What This Means for You
Three things I’d take away.
First, assembly is automatable. Interpretation isn’t. Pulling data, formatting it, flagging obvious mismatches, writing a first-pass narrative — agents are good at that. Why did this GL account spike in March? Is this reconciliation exception a problem or a timing difference? Those calls require context, judgment, and accountability. That’s your job.
Second, your small business clients will start using this. When someone with QuickBooks can run an AI-assembled close and hand it to their accountant, the nature of what you receive changes. You might get cleaner data. You might also get AI-assembled packets with errors baked in that the client couldn’t catch. Think of it like spell check — genuinely useful, but confident and fast in ways that can hide mistakes from someone who doesn’t know what they’re looking at.
Third, nothing about accounting changed. The professional responsibility, the judgment, the client relationship, the sign-off — none of that moved. The scaffolding shifted. That’s meaningful. But it’s not the revolution the headlines are selling.
Key Takeaways
- Anthropic launched 10 financial services agent templates and 15 small business workflows this week — real products, already in use at firms like JPMorgan, Goldman, and Citi.
- These agents handle assembly and scaffolding. They do not post to ledgers, issue opinions, or make professional calls. Every output requires human sign-off by design.
- The best model scores 64% on the finance agent benchmark. Human review isn’t optional — it’s structurally required.
- The value is the jump-off point: getting from zero to a structured draft fast. The judgment on what that draft means stays with the accountant.
- Your small business clients will start using these tools. Knowing how they fail is part of serving those clients well.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply