Cowork went GA April 2026 as Anthropic's desktop agent for knowledge workers (finance, ops, legal, analyst teams), while Claude Code is for engineering and Claude.ai is for ad-hoc chat. A per-department routing table with seat-cost maths: most companies should buy Cowork by default and reserve Claude Code seats for engineering.
Since Claude Cowork went generally available in April 2026, Anthropic sells three products that answer to the same first name, and we have watched more than one COO try to choose between them from the pricing page. That is the wrong page. They are three tools for three kinds of work, not three tiers of one thing, and the decision is about which kind of work each team actually does.
Our position, stated before the argument: buy Cowork by default for the knowledge workers, reserve Claude Code for the people who write code, and keep Claude.ai for everyone whose need is a question and an answer. The seat maths points the same way. A Cowork seat costs less than an engineering seat and does more than a chat seat for anyone whose day is spent in files and systems, so buy the cheaper agent seat unless the person merges code.
What is each product actually for?
Claude.ai is the chat app. You ask, it answers; you paste a document, it summarises. Right for ad-hoc work, because it waits for you and needs nothing set up; wrong for a process, because it does not sit inside your files or systems and everything it produces has to be carried somewhere by hand.
Claude Cowork is Anthropic's desktop agent for knowledge workers. Point it at a folder and its connected tools, describe an outcome, and it runs the steps: reads the export, reconciles it against the statement, writes the variance commentary, saves the spreadsheet. The difference from the chat app is not intelligence. Cowork does the work rather than describing it, and it holds standing instructions, a shared knowledge base and access boundaries that outlast whoever set them up.
Claude Code is the agent for engineering. It lives in the terminal and the repository: it reads the codebase, writes the change, runs the tests, opens the pull request. It assumes the person driving it can read what it wrote and knows when the tests are lying. Handing it to someone who cannot is handing them a compiler and wishing them luck.
Which teams should get Cowork, Code or the app?
Route by the work, not the title. This is the default; the exceptions come after it.
- Finance. Cowork. Month-end close, reconciliations, variance commentary from the ledger export and board pack assembly are exactly the multi-step file work it was built for. Start here.
- Operations. Cowork. Supplier files, procedures, the weekly report nobody enjoys assembling, the pile of documents that need turning into one clean sheet.
- Legal. Cowork, behind the tightest data boundary in the company. Clause extraction, first drafts against your own playbook, review notes. A lawyer still signs.
- Analysts. Cowork, and the fastest payback of any team. Analysts who write and version their own pipelines earn a Code seat on the same rule as engineers.
- Sales. Claude.ai for the people who sell; Cowork for the sales operations person who builds the pipeline report and cleans the CRM export every Monday.
- HR. Cowork, for handbook answers and joiner and leaver paperwork, behind the hardest access boundary of all, because the people function writes the usage policy the rest of the company copies. Any output about a named person has a human decide.
- Engineering. Claude Code in the repo, plus the app for the questions that are not about the repo. Not Cowork, unless the person is really doing an operations job that sits in engineering.
- Leadership. Claude.ai. Founders ask for Cowork and use it as a chat window. Give them the app; revisit when an executive assistant runs the diary and the inbox through an agent.
Why buy Cowork by default?
Because most of the work in a company is neither conversation nor code. It is files, systems and the steps between them, and that is the shape Cowork was built for. The chat app makes people faster at asking; Cowork makes them faster at finishing. Where the hours go in a finance or operations team is rarely 'thinking' and usually 'moving information from one place to another in the right order', which is an agent's job.
The seat maths follows, expressed qualitatively because prices will move before this page does. An engineering seat is the most expensive and pays back only when code gets merged; a chat seat is the cheapest and pays back in minutes saved per question; a Cowork seat sits between them and pays back in whole tasks. A knowledge worker who runs a repeatable process every week wants the middle seat. Someone who opens the tool twice a week to sharpen a paragraph wants the app.
Code for people who merge code, Cowork for people who finish files, the app for everyone who asks questions.
How do you keep the seat count honest?
Three questions per person, in order. Do they write code that gets merged? Claude Code. Do they run a multi-step task in files or systems at least weekly? Cowork. Neither? The app. Answer them for the org chart rather than the job titles, and the count is usually smaller than the first enthusiastic estimate.
Then review against usage rather than requests. A Cowork seat whose history is a chat log is an app seat wearing the wrong badge; downgrade it. A person pasting model output into a terminal is an engineer without a Code seat; upgrade them. Do it on a fixed cadence with the team leads who see the work.
Two rules from the AI usage policy carry across. The approved tools page decides which product, on which account, with which settings, and who can add one; Cowork on a personal login is a different tool from Cowork on the company contract, and only one of them is on the list. And the human-in-the-loop line sits at consequence, not capability: Cowork can draft the letter, the pack and the contract, but any output touching a named person's pay, employment or legal position has a person decide, with a record of what they saw.
What changes for engineering?
Less than the other teams expect. The product choice for engineering was settled before Cowork existed: engineers use Claude Code, and nothing about April 2026 changed that. What changes is engineering's relationship with the rest of the building.
Before Cowork, an analyst who wanted an agent asked an engineer for a script and joined a queue. Now finance runs its own. Engineering shifts from writing one-off automations for other teams to building the connectors those agents talk to and owning the access boundaries they enforce: a platform job, and a better use of the most expensive seat in the company than answering tickets. The one request to refuse is handing Code to a non-engineer to save a Cowork seat. It is the dearer seat and the wrong tool.
What to do next
Write the routing table for your own org chart this week. The three questions take an afternoon, and the output is a seat list you can defend line by line. Then look at the setup work under the seats: the shared knowledge base, the access boundaries between HR, finance and legal, the tool register in your identity provider. That is paid once, and it decides whether Cowork lands as a way of working or as an expensive chat window.
It is also the work we do. We do not resell seats and we have no product of our own; we build the connectors, boundaries and controls that sit underneath Anthropic's tools, so the seat you bought does the job you bought it for. It starts with a 20 or 45-minute discovery call and, where the scope needs defining, a four-day Spec from €5,000 that produces a written brief you own outright, to take to us or anyone.