The release, in one paragraph
On May 5, 2026, Anthropic shipped a set of ten reference agent templates aimed at banking, insurance, asset management, and fintech: Pitch builder, Meeting preparer, Earnings reviewer, Model builder, Market researcher, Valuation reviewer, General ledger reconciler, Month-end closer, Statement auditor, KYC screener. The same release rolled out full Microsoft 365 integration — Claude carrying context across Excel, PowerPoint, Word, and Outlook simultaneously — plus a Moody's data partnership for one of the templates and an updated MCP server pack for vendors like FIS. Vals AI's Finance Agent benchmark has Opus 4.7 leading at 64.4%, and Anthropic's deck names JPMorganChase, Goldman Sachs, Citi, AIG, and Visa as production customers since the original Claude for Financial Services launch last July.
The headline is "ten agents shipped." The substance is what each template is: a packaged bundle of skills (instructions and domain knowledge), connectors (governed access to source systems), and subagents (additional Claude calls fanned out under a coordinator). That bundle — not the model, not the IDE, not the platform — is now Anthropic's unit of vertical procurement.
Why "the template" is the new procurement object
For two years the buying-side question was "which model do we standardize on" or "which platform do we build on." Both are still real questions, but they are now table-stakes questions. The new differentiating question is "which template do we adopt — and how do we adapt it to the bank we actually run?"
A template is the right unit because it answers three questions at once that buyers were previously answering separately:
1. What does the agent know? The skills layer encodes the regulatory framework, the firm's internal style, the rubric an analyst would apply. Without it, every agent is a blank slate. With it, the agent inherits a baseline of domain literacy that didn't exist in the model weights.
2. What is the agent allowed to touch? Connectors are the IAM-aware data plane. A KYC screener that can read sanctions lists, customer onboarding docs, and transaction history — but cannot mutate them — is a different artifact than a coding agent with raw shell access. Templates push the buyer to think about least-privilege at the agent level instead of treating IAM as a deployment afterthought.
3. How does the agent decompose work? The subagent layer says: the coordinator handles intake, fans out specialist subagents for each step, and consolidates. That decomposition is encoded in the template, not in a prompt the user has to invent. For workflows with five-to-fifteen serial steps — which is most of finance back-office — having the decomposition pre-baked is the difference between an agent that works in a demo and one that survives an audit.
This is the same packaging shift that turned cloud infrastructure from "rent a VM and configure it" into "rent a managed service tuned for the workload." The model is the VM; the template is the managed service.
What this changes for buyers in regulated industries
The "build vs buy" line shifts. A year ago the choice was build a custom agent on a frontier model or buy a SaaS that does the workflow. Templates introduce a third option that didn't cleanly exist before: fork a vendor reference template and adapt it. That is operationally closer to forking an open-source project than to either building or buying. It also implies an integration practice most enterprises haven't staffed for — the work to take the Pitch builder template and adapt it to this firm's compliance posture, this firm's style guide, this firm's CRM, and this firm's approval workflow is real engineering, not a configuration screen.
Microsoft 365 as a single context surface is genuinely new. A Claude session that reads a deal model in Excel, drafts the deck in PowerPoint, pastes the right exhibits into a Word memo, and queues the cover email in Outlook — all carrying context across the four — is what most banking teams have been improvising with copy-paste for two years. The integration is more interesting than it sounds because it changes the cost model of the next ten workflows: once Claude is wired into the Office surface, every additional template the bank deploys piggybacks on the same plumbing.
Vendor data partnerships become procurement objects. A Moody's-backed template implies the buyer is licensing two things at once — Anthropic's agent runtime and Moody's data — through a single template. That bundling is convenient for buyers who already license Moody's; it is a lock-in question for everyone else. Read the small print on which data goes with which template and what swap-out costs look like.
Where we'd push back on the narrative
Two gaps worth being honest about.
A reference template is a starting point, not a product. The Pitch builder Anthropic ships is calibrated against a generic investment-banking pitch. The Pitch builder your firm needs has to follow your style guide, pull from your deal database, respect your compliance review process, and route to the right MD for sign-off. Adapting the template to that reality is the real engineering work, and it is invisible in the launch demo. A bank that installs the template and expects production-ready output on day one is going to have a rough quarter.
Ten templates is not the same as ten complete workflows. The templates Anthropic shipped cover discrete, well-bounded tasks — review a financial statement, build a pitch, prepare a meeting. Real finance work is not ten discrete tasks; it is hundreds of micro-decisions glued together by tribal knowledge. Templates close the gap on the discrete pieces; the glue is still missing, and the firms that adapt fastest will be the ones that build their own glue layer between the templates rather than waiting for Anthropic to ship it.
What we'd build differently this quarter
If our team were standing up a financial-services AI program off the back of this release, the operational defaults we'd reach for:
- Treat the templates as forks, not installs. Stand up an internal repository of forked templates per workflow. Version them. Code-review the prompts and connector configs. The discipline is exactly what an SRE team applies to Helm charts; the artifact is different but the operational posture is the same.
- Stand up the connector layer once, not per template. Excel, PowerPoint, your CRM, your data warehouse, your KYC vendor — the connector configurations should live in a shared layer that every template inherits. Otherwise every new template re-litigates IAM.
- Author firm-specific skills as a first-class artifact. The compliance reviewer at your firm has rules in their head that aren't in any public document. Capture those rules, version them, and bind them to the templates that need them. This is the closest thing to a durable moat a bank can build on top of a frontier-model agent.
- Build a workload-specific eval suite per template. Vals AI's Finance Agent benchmark is a useful sanity check; it does not tell you whether the Earnings reviewer template, adapted to your firm's coverage universe, with your analyst rubric, makes decisions you'd ship. The eval that matters is "the last 30 earnings reviews my analysts wrote, replayed through the template, graded by a senior analyst against the original."
Sonnet Code's take
Vertical templates are the most consequential shift in enterprise AI procurement this quarter, and the firms that win the next two years aren't the ones that bought Anthropic's templates first — they're the ones that adapted them fastest to the firm's actual conventions, connectors, and compliance posture. That adaptation work is the same shape we already run for AI development clients in regulated industries: forking the template, wiring its connectors to internal systems, encoding firm-specific skills, and standing up the eval suite that tells the auditor whether the agent is still trustworthy six months from deploy. We pair that with AI training engagements where senior domain reviewers — analysts, underwriters, compliance officers, depending on the workflow — author the rubrics, golden examples, and red-team prompts that calibrate the templates against the firm's actual standards. If your firm spent yesterday's all-hands looking at Anthropic's ten finance templates and is now wondering which one to pilot first, the next conversation isn't about picking the template. It's about who owns the fork.
