Anthropic Shipped Workflows, Not Just a Model

Anthropic added ready-to-use artifacts to Claude for Legal yesterday: interactive trackers and dashboards built for named tasks, which a team adapts with its own documents and then shares with colleagues. No new model, no benchmark score. The company says the most common requests since Claude for Legal launched in May were simpler setup and easier sharing, and this is the answer to both.

What is in the gallery

None of those require a capability that did not exist last month. A competent person with a general assistant could assemble any of them given a few days, the right prompts and the discipline to maintain the result. Almost nobody does, because the few days are the problem.

Two design choices worth copying

Every output links to its source. Anthropic states that each answer carries a link back to where it came from. In work where being wrong is expensive, that is the difference between a tool you can use and a tool you have to re-verify from scratch, which costs more than doing it yourself.

Each artifact opens with a worked example. You see the finished shape before supplying your own data. That sounds like a nicety and it is the adoption mechanism: it converts a blank prompt box, which is a design task, into an editing task, which people will actually start.

Both of those transfer to anything you build internally. An internal tool that cites its sources and ships with one filled-in example gets used; one that presents an empty field and an instruction to describe what you need gets opened once.

The cite check is pointed

A tool whose stated job is comparing a document against its sources, flagging quotes that do not match and sources that do not support the claim, is a direct response to a well-documented failure mode in professional writing.

Shipping that as a named artifact rather than a reassurance is the honest version of addressing it. The problem does not get solved by asking people to trust the output more; it gets solved by building the check into the workflow, which is the argument in quality assurance when AI writes the first draft.

The division of labour, stated out loud

Anthropic describes the arrangement plainly: the professional brings their documents and sources, Claude completes the first pass, and the lawyer applies judgment to the results.

That is unusually restrained positioning for a product launch, and it is the sentence worth borrowing. Any workflow you deploy should be able to state the same split as clearly, including who owns the judgment step and what they are expected to check.

Where that split is vague, the judgment step erodes quietly until nobody is performing it and everyone assumes somebody is. Writing it into the procedure is what prevents that, as in checklists that hold up.

What this signals for other professions

Legal is a reasonable first vertical because the tasks are well defined, the documents are structured, the stakes are high enough to justify the spend, and the profession already thinks in precedents and playbooks.

The same packaging pattern applies anywhere work is repetitive, document-heavy and judgment-gated: accounting, insurance, construction administration, clinical documentation, property management. Expect named workflows for those rather than new models, and expect the vendors who get the examples right to win regardless of which model sits underneath.

For a small business that is mostly good news, since a packaged workflow is far cheaper to adopt than a build. The questions to carry into it are the usual ones about what the tool can reach and what the terms say, covered in what to check in AI terms of service, and whether the thing actually works on your documents rather than the demo ones, which is a proof of concept that decides.

For professional services firms specifically, the existing coverage of what these tools catch and miss is in AI contract review and document review for law firms.

Frequently Asked Questions