ALTITUDE
The Agents Came to You. The Operating Model Didn't.
Anthropic has launched Claude for Small Business: a plugin pack of ready-made skills for the smallest teams. What it actually is, what it isn't, and where the boundary sits.
IN 30 SECONDS
In May, Anthropic launched Claude for Small Business. The name suggests a business system; the package is smaller: thirty-one ready-made skill files inside Claude Cowork, plus connectors to a selection of mainstream tools, included with paid Claude plans. Useful, if those tools match yours, and it asks before anything touches money or customers. But it is the same for every business that installs it, and there is no operating model in the box. That part, the part fitted to one business, is the part that compounds.
In May, Anthropic, the AI lab behind Claude, launched Claude for Small Business. The ambition was put plainly by its president: AI is the first technology that can close the gap between small firms and large ones, taking on "the work that piles up after hours". The name and the framing promise a great deal. What actually shipped is smaller and simpler, and worth seeing clearly.
It is a plugin pack: thirty-one ready-made skill files across finance, operations, sales and marketing (payroll planning, monthly close, lead triage, invoice chasing, contract review, a weekly business pulse), plus eleven connectors to a selection of mainstream software, QuickBooks, PayPal, Stripe, Square, HubSpot, Canva, Docusign, Slack, Google Workspace and Microsoft 365. It runs inside Claude Cowork, the desktop agent that works in plain language, and installs at no extra charge on paid Claude plans. Alongside it, a free AI-fluency course built with PayPal.
Taken for what it is, it has real uses. The skills target familiar bottlenecks, and the approval pattern is the right one; Anthropic's own description puts it well: "you approve every step that touches money or customers." If your tools happen to match the connector list, a handful of these will save real time.
But the name is carrying more than the package holds. "For Small Business" suggests something shaped to a business; what arrives is generic by design. The thirty-one skill files that land in a Surrey land management practice are the thirty-one that land in a Chicago bakery. The connector list is a particular selection of mainstream, largely American software; if your accounting, your bookings or your point of sale live elsewhere, whole strands of the pack do nothing on day one. Nothing in it knows the business it has landed in: the clients, the seasons, the pricing, the functions that make up how the work actually runs, the things the owner has decided never to do. Nothing it does accumulates in a data foundation the business owns. There is no picture of what is running and what it can reach. And there is no depth in any particular domain, which is where most of the judgement in real work lives.
One default is worth knowing about before anything runs. On Team and Enterprise plans, business data is excluded from model training. On the individual Pro and Max plans most small operators actually hold, training on chats is switched on by default, and turns off manually in settings. A business running client work through an agent wants that switch set deliberately, not inherited.
None of this reads as a reason to hold back. It reads as a map of where the boundary sits. The generic agent is becoming a commodity, and this launch just accelerated that: whatever a canned workflow can do is now included with the subscription. What no platform ships is the layer around the agent, because it cannot be packaged: context written down where the AI reads it, data in accounts the business owns, clear rules about what runs unattended, and the domain knowledge that carries the judgement. The agents came to you. The operating model didn't, because it can't. It has to be built to fit.
If you hold a paid Claude plan, the pack is already there in Cowork, and the thirty-one skills take an afternoon to sift; if the connectors match your tools, a handful may fit a real bottleneck. The training toggle is in settings. And the skills that almost fit but not quite are worth noting as they turn up, because that gap, between a generic skill file and one particular business, is the most accurate map there is of where an operating model needs building next.