Anthropic just launched thirteen enterprise plugins for Claude. The industries it chose first — financial analysis, investment banking, equity research, private equity, wealth management — aren't random. They're a displacement thesis published as a product roadmap.
On February 24, Anthropic launched thirteen new enterprise connectors for Claude Cowork: DocuSign, FactSet, MSCI, LegalZoom, Google Workspace, Apollo, Clay, Outreach, Similarweb, WordPress, and Harvey. Alongside the connectors, they released industry-specific agent templates for financial analysis, investment banking, equity research, private equity, wealth management, HR, design, and engineering operations. And they opened a private plugin marketplace where enterprises can build, curate, and distribute their own agent tools internally.
Three days later, they acquired Vercept — a Seattle startup that builds cloud-based macOS agents capable of operating any desktop application the way a human would. Vercept's investors included Eric Schmidt and Jeff Dean. UiPath shares dropped 3.6% on the news.
Taken separately, these are product announcements. Taken together, they are something else: a published theory of which human work is most ready to be absorbed by AI. The plugin list is a catalog of the displacement.
What the Choices Reveal
When a platform company launches an app store, the categories it opens first are diagnostic. Apple launched the App Store in 2008 with games, productivity, and social networking — the domains where mobile-native interaction patterns were already understood. Salesforce's AppExchange started with sales automation and CRM extensions — the workflows closest to its own platform capabilities. The initial catalog isn't a random sample of possibility. It's a bet about where the new platform creates the most value fastest.
Anthropic's thirteen connectors sort into three clusters. The first and largest is financial services: FactSet for market data, MSCI for risk analytics, and industry templates spanning financial analysis, investment banking, equity research, private equity, and wealth management. Five templates out of eight target finance. The second cluster is legal: LegalZoom for document preparation and Harvey for AI-native legal work. The third is sales and marketing operations: Apollo, Clay, Outreach, and Similarweb.
The financial services concentration is not proportional to the economy. It is proportional to revenue per worker. A junior analyst at an investment bank generates between $300,000 and $500,000 in revenue annually. An equity researcher generates similar figures. A wealth manager handling $100 million in assets under management generates roughly $1 million in fees. These are the most expensive information workers in the economy, and their work is almost entirely digital — reading documents, building models in spreadsheets, writing reports, pulling data from terminals, and communicating through structured channels.
The work is expensive because it is judgment-intensive. But a specific kind of judgment-intensive: the kind where the judgment operates on structured data through known analytical frameworks. Comparable company analysis follows a template. Discounted cash flow models have standard assumptions. Equity research reports have conventional structures. The judgment is real, but it runs on rails.
That is exactly the kind of work an agent with access to FactSet data, MSCI risk models, and Excel can begin to perform.
The Two Layers
The Vercept acquisition closes a gap that the plugin architecture alone cannot. Plugins connect Claude to applications through APIs — structured, documented interfaces where the data flows in predictable formats. But most enterprise software was not designed for agent consumption. It was designed for humans sitting at screens, clicking through menus, dragging files between windows, reading visual layouts that have no API equivalent.
Vercept's technology lets Claude operate the screen itself. Their product, Vy, ran a macOS instance in the cloud and navigated it the way a person would — clicking, typing, scrolling, reading what appeared on screen. On the OSWorld benchmark, Claude's accuracy at desktop tasks improved from under fifteen percent in late 2024 to 72.5 percent in early 2026. Not perfect. But approaching the threshold where the remaining errors are edge cases rather than fundamental limitations.
Combined, the two layers cover the full surface area of enterprise computing. Plugins handle applications with APIs: financial terminals, CRM systems, document platforms, communication tools. Computer use handles everything else: legacy software, proprietary internal tools, applications where the only interface is a window on a screen. The result is that there is no category of digital work that is architecturally unreachable.
The constraint on what Claude can automate has shifted from access — which applications it can interact with — to judgment — which decisions it can make reliably enough to trust. And the judgment threshold is moving.
The Marketplace as Platform Strategy
The private enterprise marketplace is the most structurally significant announcement, though it received the least attention. Enterprises can now build custom plugins, curate them into internal catalogs, and distribute them to specific teams with per-user provisioning and automatic installation.
This is the app store pattern applied to agent capabilities. And the history of app stores has a consistent lesson: the platform that hosts the store captures more long-term value than any individual app within it. Apple didn't need to build the best weather app. It needed to be the place where weather apps lived.
For SaaS companies, the marketplace creates a specific strategic dilemma. A FactSet plugin that lets Claude pull financial data and build models directly competes with FactSet's own interface — the terminal where humans sit for eight hours a day, for which FactSet charges $12,000 to $24,000 per user per year. If the plugin becomes the primary way analysts interact with FactSet data, the value has migrated from FactSet's interface to Anthropic's platform. FactSet becomes a data provider. Anthropic becomes the workspace.
The same logic applies across every connector. DocuSign becomes a signing API. LegalZoom becomes a document template service. MSCI becomes a risk data feed. The companies retain their data and their regulatory positions. They lose the interface — the surface area where users spend time, form habits, and experience the product.
Jefferies downgraded both Workday and DocuSign in the same week, citing AI disruption. FinancialContent headlined the Cowork launch as 'The SaaSpocalypse Arrives.' Between mid-January and mid-February 2026, roughly a trillion dollars was wiped from enterprise software valuations. The market is pricing the pattern before it fully manifests.
The Layoff List and the Plugin List
In the same month that Anthropic published its plugin catalog, the layoff announcements accelerated. Block eliminated nearly half its workforce and the stock surged twenty-four percent. Dell cut its sales organization. IBM reduced entry-level positions. Google announced restructuring across multiple divisions. Each company offered the same explanation: AI makes the work more efficient, fewer people can produce the same output.
The correlation between the plugin list and the layoff list is not causal in any direct sense — Anthropic did not cause Block to cut workers. But the two lists share a common underlying variable: both identify which professional work is most compressible by AI. The plugin list identifies it from the supply side — here is where we think an agent can do the work. The layoff list identifies it from the demand side — here is where companies believe they need fewer humans.
Financial analysis. Sales operations. Legal document preparation. HR onboarding. Engineering operations. The overlap between the thirteen plugin categories and the recent restructuring announcements is almost complete.
There is a temporal pattern worth noting. The restructuring announcements tend to follow the capability announcements by roughly two to four quarters. The SaaS companies that faced the most pressure after the August 2025 AI stock rotation were the ones whose core workflows most closely matched the agent capabilities announced in Q1 2025. The February 2026 plugin list may be a leading indicator for which specific professional roles face restructuring pressure in Q3 and Q4 of 2026.
What the Security Model Reveals
The Cowork plugin security architecture has three components: admin-controlled governance policies, optional human approval prompts for sensitive actions, and audit logs tracking agent activity. Enterprises can configure which plugins each user accesses and require confirmation before specified categories of action.
What the architecture does not have is identity verification. When Claude executes a DocuSign transaction, pulls confidential FactSet data, or submits a LegalZoom filing, the 'approval' is a prompt dialog. There is no cryptographic binding between the approval and a specific human principal. No biometric attestation that the person who clicked 'confirm' is the person authorized to make that decision. No tamper-evident record distinguishing 'someone approved this' from 'the authorized decision-maker verified this exact action.'
The security researchers at PromptArmor demonstrated a data exfiltration attack against Cowork that bypassed the approval mechanism entirely. Anthropic acknowledged that agent safety 'is still an active area of development.'
This is the pattern that repeats across every major agent platform launched this week. OpenAI partnered with McKinsey, BCG, Accenture, and Capgemini to deploy enterprise agents at scale — through consulting engagements, not through verification infrastructure. Microsoft previewed Copilot Tasks, which runs autonomously in a cloud browser and requests 'consent' before significant actions — the same lightweight approval model. Perplexity Computer coordinates nineteen models across four hundred apps with human checkpoints that amount to confirmation dialogs.
Five major announcements in four days. Anthropic, OpenAI, Microsoft, Perplexity, and Meta (which acquired the agent orchestration startup Manus for two billion dollars in December). All solving orchestration. All shipping agent capability into production. None shipping identity verification for the agents they're deploying.
The capability layer is being built at extraordinary speed. The verification layer — cryptographic proof that a specific human approved a specific agent action — does not yet exist at the product level in any of these offerings. The gap between what agents can do and what they can prove is growing with every connector, every template, every marketplace listing.
The Catalog as Forecast
Product catalogs have always been forecasts disguised as inventory lists. The Sears catalog of 1896 didn't just sell goods — it predicted which rural markets were ready for direct commerce. The iTunes Store of 2003 didn't just sell music — it predicted that digital singles would replace physical albums. The predictions were embedded in the choices: what to include, what to exclude, and how to organize what remained.
Anthropic's plugin catalog is a forecast of which professional work is ready for agent-mediated execution. The five financial services templates say: this work — the analysis, the modeling, the research, the portfolio management — can be performed by an agent with access to the right data and the right tools. Not perfectly. Not autonomously. But well enough to change the economics of who does it and how many people are needed.
The industries that don't appear in the catalog are equally informative. No healthcare templates. No education templates. No construction or manufacturing templates. The absence reflects either regulatory complexity that makes the work harder to automate, physical-world requirements that digital agents cannot meet, or judgment requirements that remain beyond current capability.
The catalog will grow. Each new plugin extends the surface area of what Claude can reach. Each new template encodes another workflow that an agent can perform. The private marketplace means enterprises will build their own templates faster than Anthropic alone could — turning their institutional knowledge into agent instructions, their internal processes into executable plugins.
When a company publishes which jobs it intends to augment, the professional class should read the list carefully. Not because the current tools are good enough to replace anyone today. But because the catalog is the roadmap, the marketplace is the distribution mechanism, and the financial services industry — with its high revenue per worker, its digital-native workflows, and its structured analytical frameworks — is where the transformation begins.
The Sears catalog didn't kill general stores overnight. But the general stores that didn't notice the catalog was also a forecast learned the lesson late.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)