Six reasons financial firms can't just install Anthropic's new agents

On May 5, Anthropic announced ten production-grade Claude agents for financial services: a Pitch Builder, a Meeting Preparer, an Earnings Reviewer, a Model Builder, a Market Researcher, a Valuation Reviewer, a GL Reconciler, a Month-End Closer, a Statement Auditor, and a KYC Screener. All ten shipped as Apache-licensed templates in the anthropics/financial-services repo. Microsoft 365 integration ships at the same time, putting Claude inside Excel, PowerPoint, Word, and (soon) Outlook with context preserved across all four. Moody’s, Dun & Bradstreet, Verisk, Guidepoint, IBISWorld, Third Bridge, Fiscal AI, and Financial Modeling Prep all launched MCP data connectors at the same time. Citadel, BNY, Carlyle, FIS, Mizuho, Travelers, Walleye Capital, and Hg are quoted as customers in the announcement.

This is the most consequential financial-services announcement in AI to date. The breadth alone reshapes what a credible “AI roadmap” at a bank, insurer, or asset manager should look like. If you run operations at a $500M-$5B firm and you haven’t read the Anthropic announcement yet, stop reading this and read that one first.

But here’s the part the press cycle has skipped past: the templates are starting points, not shipping products. Anthropic says it explicitly in the README: “These are reference templates — they get better when you tune them to how your firm works.” Every output is “staged for human sign-off.” Nothing in the repo “constitutes investment, legal, tax, or accounting advice.” Anthropic is not delivering you ten production tools. They’re delivering ten scaffolds that need to be productionized inside a specific firm, against specific data, for specific workflows, for specific compliance regimes.

Financial firms outside the top of the market have a real problem on their hands. The templates exist. The CEO is asking when they ship. Internal engineering can’t pause everything to do this. The big SIs cost too much. The new Anthropic-backed services joint venture is built for the largest names first. So how does a $1B-revenue regional bank or a $5B-AUM asset manager actually get one of these agents from “marketplace download” to “the operations team uses it Monday morning”?

Here are the six gaps to know about before signing a vendor.

1. The data connectors are paywalled

The new agents are powerful because of the data they can access. Moody’s MCP server gives you “proprietary credit ratings and data on more than 600 million public and private companies.” PitchBook gives you private equity data. FactSet, S&P Capital IQ, LSEG, Morningstar, Daloopa, Aiera — all integrated. All gated behind enterprise subscriptions.

If your firm doesn’t already have these subscriptions, the agents work on whatever data you already have. Sometimes that’s enough. Often it isn’t. A KYC Screener with no D&B subscription is doing string matching against public records. A Market Researcher with no Capital IQ subscription is summarizing what’s on Yahoo Finance. The cost analysis on the front end of a pilot needs to include connector-licensing math, not just engineering time.

What’s the right move: in the pilot’s first week, map every step of the target workflow against the data sources required. Identify which connectors you already have, which you can buy, and which you can substitute with internal data. A KYC Screener pointed at your own customer database with limited external lookup may be more useful than a generic one pointed at premium data your team will never see again after the pilot.

2. The templates assume someone else’s workflow

The GL Reconciler in the Anthropic repo runs a “close checklist,” prepares “journal entries,” and produces “variance commentary.” That’s the abstract pattern. Your firm’s checklist is different. Your chart of accounts is different. The fiscal calendar treatment of accruals at a Bermuda-domiciled reinsurer is different from a Connecticut hedge fund administrator, which is different from a regional bank.

The template gets you 60% of the way to a tool that works at your firm. The remaining 40% is your firm’s actual workflow encoded into the agent’s skill files, prompts, and output formats. You can either do that customization in-house — assuming you have engineers who understand both Claude Skills syntax and your firm’s operations deeply — or you bring in a partner who has shipped this kind of customization before.

What’s the right move: scope your pilot as “the GL Reconciler running OUR close checklist on OUR ledger” instead of “deploy Anthropic’s GL Reconciler.” Same architecture, very different deliverable.

3. Adoption beats deployment, every time

Anthropic’s announcement doesn’t say much about what happens after the agent is installed. That’s not a criticism — it’s not their job. Their job is to ship the agent. The firm’s job is to get the agent used.

Most internal AI rollouts die here. The IT team installs the GL Reconciler. The operations VP demos it in an all-hands. Three people try it. Two go back to their Excel macros within a week. The fourth quietly uses it but never tells anyone. Six months later the tool is technically deployed and effectively dead.

The fix is unglamorous: someone has to sit next to the operators during real production work and watch what breaks. Reconciliation entries that the agent gets wrong because the chart-of-accounts mapping is fuzzy. Variance commentary that uses your firm’s internal terminology incorrectly. Edge cases nobody thought to specify. After two weeks of this, the tool either works for real or you find out exactly why it doesn’t.

What’s the right move: the pilot should include two weeks of paired adoption work after the agent ships. The deliverable is not the code. The deliverable is your operations team using the tool by Monday of week three.

4. Your eval set doesn’t exist yet

The 10 templates work well on Anthropic’s demo data. The question is how well they work on yours.

Every production AI tool needs an eval set: a structured collection of representative inputs paired with known-correct outputs that proves the tool is doing what you think it’s doing, and lets you catch regressions when you update prompts or models. A Statement Auditor that catches 95% of issues in Anthropic’s test corpus might catch 70% on your firm’s PE limited-partnership financial statements because the formatting conventions are different.

Building an eval set sounds easy and isn’t. The work is: identify 30-50 representative cases from real production data, label them with what the agent should produce, run the agent against them, measure the gap, iterate. This is the part most pilots skip. It’s also the part that determines whether the tool is trustworthy six months in.

What’s the right move: budget for eval-set construction inside the pilot. Plan to spend a week of work with your operations team to assemble it. Without an eval set you cannot make safe changes to the agent’s prompts or upgrade the underlying model when Anthropic ships Claude Opus 4.8.

5. Compliance is firm-specific, not industry-specific

Anthropic ships the KYC Screener as “assembles entity files, reviews documents, packages escalations for compliance teams.” That sentence is correct for the general case. The specifics at your firm — which document types are required for politically-exposed-person screening, what the escalation criteria are, who in the compliance team gets which case, how the audit log needs to be structured for your regulator — none of that lives in the template. It lives in your firm’s policies and procedures manual, and it varies by regulator (OCC vs. state banking commissioner vs. SEC vs. FCA) and by product line.

Same is true for the Statement Auditor and the Valuation Reviewer. The template runs the math. The escalation logic, the audit-trail format, the retention requirements — those are your compliance officer’s job, not Anthropic’s.

What’s the right move: bring your compliance lead into the scoping conversation in week one. Their input determines what the agent escalates and what it auto-resolves. Skip this and you get a tool your compliance officer vetos in week six.

6. System integration is the unsexy 70% of the work

The agent has to read from your CRM, write to your portfolio accounting system, push files to your document management platform, and post events to your audit log. Each of those integrations is straightforward in isolation and brutal in aggregate.

The 30 lines of code that call Claude with the right prompt are easy. The 700 lines of code that authenticate to your CRM via OAuth, parse the input format that the CRM exposes, handle the rate limits when you batch 200 entities through the KYC Screener overnight, retry on transient failures, log every call with the right audit fields, and surface errors to operators in a way they can act on — that’s the work. It’s also where most internal AI pilots get stuck for months.

What’s the right move: pick a workflow where the integration surface is shallow. A pilot that connects to one system you already control beats a pilot that has to negotiate with eight teams to get a service account. If your firm’s first agent has to integrate with seven systems, the pilot is the wrong shape — scope to one or two.

---

What this means for the next 60 days

The Anthropic announcement reframes the conversation at most firms. It used to be: “should we build an AI roadmap?” Now it’s: “Anthropic just shipped ten things our competitors will deploy in the next six months. What do we do about it?”

Three things, in order:

One. Pick the agent that addresses the workflow that’s costing your operations team the most hours. Probably KYC Screener, Month-End Closer, or GL Reconciler. The first deployment doesn’t need to be the most sophisticated one. It needs to be the one that ships.

Two. Scope the pilot around YOUR workflow, not the template. Six to eight weeks. Fixed scope. One specific workflow your team does every week.

Three. Build the eval set first, the adoption playbook second, the integration third. In that order. Most pilots build the integration first and discover the eval set issue at week seven. Don’t be most pilots.

If you saw the Anthropic announcement last week and thought “cool, but who actually does this for us at our scale” — that’s the conversation worth having. The templates are real. Productionizing them takes a partner who has done the customization, the eval-set work, and the adoption pairing before. That’s what we do at Polystat. We deploy production AI agents at financial firms in six weeks, fixed price, one workflow at a time.

If this resonates, visit our financial services practice page or book a 30-minute conversation. We’ll spend it on your specific workflow, not on a sales pitch.