Here’s a pattern worth noticing. When a company decides to “invest in AI,” the seats almost always go to engineering first. It feels obviously right — the tools came out of developer culture, the early demos were all code, and the engineering org is used to adopting new software.
Then Anthropic published usage data from around 1.2 million Claude Cowork sessions, and the largest single category of use turned out to be business processes and operations, at 33.4%. Content creation was second at 16.4%. Software development came in under 9%.
The value isn’t concentrated where the budget is pointed. This guide is about correcting that.
Step 1 — Pick the cohort by task shape, not job title
Don’t start by asking who’s senior or who’s keen. Ask a narrower question: who in this building spends the most time moving information between files?
The reliable signals:
- They own something that repeats on a calendar — a monthly report, a weekly digest, a quarterly pack.
- Their work involves opening several files to produce one file.
- They’d describe their most time-consuming task as “not hard, just tedious.”
- They already have a personal system of templates and folders, because they had to invent one.
That last one is the strongest tell. People who’ve built their own workarounds are the ones who’ll immediately see what to hand over.
In most organisations that lands on operations, finance, HR, and marketing — and a founder or chief of staff doing six jobs at once. Pick three or four people across different functions, not four from one team; you want a spread of use cases from the first fortnight, not four versions of the same one.
Step 2 — Start in a sandbox folder, deliberately
Cowork works on a folder-permission model: you grant read, write, and create access to specific directories, and it genuinely cannot see anything outside them. That’s not a policy promise, it’s the mechanism — which makes the safe rollout pattern very simple.
For the first two weeks:
- Each pilot person makes a new folder — call it
cowork-pilot. - They put copies of the files for one real recurring task in it.
- They point Cowork there and nowhere else.
Copies, not originals, for the first fortnight. Not because Cowork is reckless — it asks before anything consequential and shows you results for approval — but because the point of a pilot is to let people be fearless. Someone worrying about the live ledger won’t try the ambitious prompt, and the ambitious prompt is exactly what you’re trying to discover.
Widen the scope once people have watched it work a few times. They’ll ask to.
Step 3 — Give each person one job to hand over, not a tool to explore
“Here’s Claude, have a play” produces enthusiasm and no evidence. Instead, have each pilot person name one specific recurring task before they start, and write down two numbers: roughly how long it takes today, and how often it happens.
Good first candidates:
- Ops — consolidating several regional spreadsheets into one report and flagging what doesn’t reconcile
- Finance — turning a month’s raw exports into the standing management pack
- HR — reducing a pile of applications into a structured comparison table against the actual job spec
- Marketing — turning a quarter of campaign data and notes into the performance write-up
- Anyone — the folder of meeting notes that should be a summary with owners and dates
The brief for each person is one sentence: make Claude do this one, and write down where it fell short. The failures are as valuable as the wins — they tell you where your data or your templates are the actual problem.
Step 4 — Connect it to where the work really lives
A folder is a good sandbox and a poor permanent home. Once the pilot has proved itself, the step change comes from connecting Cowork to the systems your work already lives in — Google Drive, Gmail, document signing, and industry data systems.
This is the point where IT joins the conversation, and it’s the right point: connectors touch real company data and should be granted, reviewed, and revocable like any other integration. Do this after the pilot, not before — you’ll be asking for access with actual evidence in hand rather than a hypothesis, which is a materially easier conversation.
Our security and data guide covers the questions your IT and legal teams will ask, including the training-data distinction between consumer and commercial plans that most people get wrong.
Step 5 — Turn one person’s know-how into everyone’s
This is the step most rollouts never reach, and it’s where the compounding actually happens.
In every team there’s someone who knows the real rules — how the board pack has to be formatted, which checks compliance always asks for, the three things that make a brief acceptable. That knowledge normally lives in one head and leaves when they do.
Organisations can deploy plugins that encode this: institutional knowledge and workflow, packaged so that everyone’s Claude already knows how your company does financial analysis, or hiring, or review. The first time a new joiner produces a correctly formatted board pack on day three, you’ll understand why this step matters more than any individual time saving.
Practically: during the pilot, have people capture the prompt that finally worked. Those saved prompts are the raw material — the difference between “we bought AI seats” and “we encoded how we work.”
Step 6 — Use the fact that it no longer needs a laptop
Cowork sessions now run remotely on Anthropic’s servers, which means scheduled work runs with no device switched on at all. “Every Monday at 7am, pull last week’s numbers into the standing report” simply happens, and the result is waiting when someone opens the app.
Two things follow for a team rollout. First, the standing weekly jobs are the highest-value thing to automate second — right after the one-off tasks that prove the concept. Second, review discipline becomes the thing that matters. Work produced while nobody was watching still needs a named human who checks it before it goes anywhere. Make that explicit early; it’s much harder to introduce later, once people have got comfortable.
Step 7 — Close the loop with evidence
Measure the two things separately.
Did it save time? Go back to the numbers each pilot person wrote down in step three and re-measure the same tasks. Three named jobs with before-and-after hours is a far better case for expanding seats than a room full of positive anecdotes.
Did the capability stick? This is the one rollouts consistently fail to evidence. Enthusiasm in week two is not capability in month six. Evidence turns “we rolled out Cowork” into “these named people run these named jobs this way, and here is the before and after” — which is something a manager can point at, and something that survives the person who championed it moving on. Keep the artefacts as you go: the measured hours, the prompts that worked, and the standing jobs with their named reviewers.
A 30-day arc
- Days 1–3 — Pick three or four people across functions. Each names one recurring task and its current cost in hours.
- Days 4–14 — Sandbox folders, copies only, one task each. Capture what worked and what didn’t, verbatim.
- Days 15–21 — Widen to real folders. Bring IT in for the first connector. Start a shared file of the prompts that worked.
- Days 22–30 — Re-measure the three tasks. Schedule the first standing weekly job and name its reviewer. Decide the second cohort based on task shape, using the same test as step one.
Then expand. The second cohort is easier than the first every time, because by then you’re not selling a tool — you’re pointing at a colleague whose Monday morning got two hours shorter.
New to the product itself? Start with what Cowork is, or how it differs from Claude Code if you’re deciding which your team needs.