Claude Cowork use cases for business: the seven jobs I actually hand over

What makes a good Cowork job?

Chat answers. Cowork does.

It opens the folder you give it, reads the files and uses the tools you’ve connected.

Then it leaves something finished behind: a document, a spreadsheet, a draft, a page.

Before I hand anything over, I ask three questions.

1. Does it repeat? A job I’ll run every day or every week earns the setup. A one-off is just a chat.

2. Does it live somewhere Claude can reach? A folder, an inbox, a calendar, an export.

   If the facts only exist in my head, Claude will guess.

3. Does it end in a draft I approve? If the last step is sending, paying or deleting, the last step stays mine.

Three yeses and it’s a Cowork job. These seven passed.

1. A morning brief that reads the business before I do

Every morning one message is waiting on my phone. Wins first. Then one line of numbers, the one lever for today, and only the new things that need me.

It reads two dozen files that other jobs wrote overnight, which makes it the heaviest thing I run.

Its first run on the current prompt took 47 steps. It also has a budget: 140 words on weekdays. Anything longer stops being a brief.

What broke: the account it ran on ran out of credit, and the briefs simply stopped.

Silence looks exactly like a quiet morning. Now a tripwire fires when the balance drops below double the daily spend.

In Cowork this is a scheduled task, and daily briefings are the first example on Anthropic’s own list. Connect your email and calendar, then:

Every weekday at 7:00, read my email and calendar from the last 24 hours. Write one message under 140 words, in this order: what went well yesterday, one line of numbers, the one thing I should focus on today, then only new items that need me. If you can't read a source, write "no fresh data from [source] today". Never guess a number. Don't reply to or send anything.

2. Inbox triage that drafts and never sends

At 7:35 a job goes through every inbox I’ve connected, looks at the last 24 hours, writes draft replies and sends me one summary.

It has never sent a reply on its own. A second job sorts my personal messages into three piles: needs you, waiting on them, FYI.

Not sending is on purpose. Treat it like a new assistant in her first month.

She can write anything. She sends nothing until you’ve read enough of her drafts to trust her voice.

Messages to real people stay mine, and so do payments and anything that gets deleted.

Every weekday at 7:30, go through my inbox from the last 24 hours. Sort every thread into: needs me today, waiting on someone else, FYI. For anything in "needs me today", write a short draft reply in my voice under the summary. Do not send, archive or delete anything. Give me one list, most urgent first.

One caution. Connectors differ in what they’re allowed to do: some only read mail, some can save drafts.

Check yours before you build on it. Mine saves drafts through a separate integration.

3. A weekly business review from two snapshots

This is the one I’d start with, and it’s the job my free starter teaches.

You keep two dated snapshots of the business in a folder, last week’s and this week’s.

Claude compares them record by record and writes down what changed, what’s over budget and one thing to improve.

In the sample business, spending on the Oak project goes from €720 to €900.

The review finds the extra €180 and names the record behind it (e4).

It flags that Oak is now €100 over its limit and spots a possible €39 a month in overlapping software.

It also notices that nothing has been cancelled yet. A saving you haven’t made isn’t a saving.

Read last-week.json and this-week.json in this folder. Compare records by ID. Save weekly-review.md with: what changed, project spending against its limit, recurring tools that overlap, and one improvement tied to my goal. Show the evidence and the calculation for every number. Separate possible savings from savings already made. Put only decisions that need me into needs-you.md. Tell me which files you actually saved.

Run it by hand first.

When it comes back right twice, ask Cowork to schedule it for Friday at 9:00, then open the Scheduled page and check the task is really there.

(I’ll explain why below.) Because it reads a folder on your computer, the desktop app has to be open when it runs.

The long version is in how to make an AI weekly business review.

4. On-brand content from a working folder

When I ran the coaches workshop in April, the first thing we built wasn’t a prompt.

It was a folder called Claude Projects, with a subfolder for each part of the business: brand, voice, content, launches.

Cowork works inside it, so every output lands where you’ll find it next week. Your file system is your memory.

Then we filled it.

First a design system: Claude builds a page showing your colors, type and buttons.

It saves the same system as JSON at the bottom, so every later prompt can use it.

Then a client-voice file, built from real phrases people in the niche use about their problem.

A marketing plugin adds ready-made commands like /draft-content and /brand-review on top.

This is the rule that keeps the copy sounding like your clients:

Using client_voice.json, write one landing-page headline under 12 words, three bullet points that mirror pain_language, and one call to action from desire_language. Do NOT invent new phrases. Only recombine theirs. No coaching jargon unless it's literally in the JSON. If nothing fits, say so.

“If nothing fits, say so” is the most useful line in it. Polish, don’t pretend. If you coach, this is most of what AI for Coaches covers.

5. Checking that the work actually happened

In my first month testing the Social Machine on my own accounts, it published 906 posts and they got 811.4K views.

The lesson I paid most for: scheduled is not posted.

Posts failed for three days behind a dead connection before anyone noticed.

Another time a report quietly fell back to an older source and looked healthy.

I read it and concluded only two of my accounts existed. Both reports told me what I expected. Neither told me what happened.

So now a job checks every post’s real status every day, and a source that fails has to say it failed. A small Cowork version:

Every day at 18:00, read today's status report from my scheduling tool. Compare what was scheduled with what actually published. List anything that didn't go out, with the account and the time. If you can't read the status for an account, mark that account "unverified". Don't repost or reschedule anything.

If social media is the job you most want off your plate, that’s what The Social Machine is for.

6. Stress-testing a decision before I spend on it

I start things easily. I forget the main lever. That isn’t modesty.

It’s written in the notes my assistant works from, and it’s set up to stress-test my ideas, not applaud them.

When I wanted to start a newsletter, it ran the numbers and came back with a cheaper test: sell one workshop first.

It was faster to answer, and it had a clear pass or fail number. I parked the newsletter.

Cowork does this better than chat because it reads your real numbers instead of your summary of them.

Put your prices, costs and audience size in the folder, then:

Be my harshest critic. Read the files in this folder. I'm choosing between [A] and [B]. For each one, tell me what it costs, how long until I know whether it works, and what number would tell me it failed. Then recommend the cheapest test that answers the real question. Don't soften it.

7. Small recurring admin I’d otherwise forget

Until September, four of the sixteen scheduled jobs on my Mac were about my house, not my business.

Pay the pool guy on the 28th. My housekeeper’s salary on the 1st. The dog’s deworming. The water filter.

The water-filter one was the clever one.

It kept the date it last fired in a small text file and stayed quiet unless that was more than 40 days ago, so it never nagged twice.

A reminder that checks before it speaks.

On 22 September I moved all four into a plain to-do app. They were sending messages from a laptop, so they only fired when the laptop was open.

A reminder that depends on your lid being up isn’t a reminder. That’s the other lesson small jobs teach: some of them don’t need AI at all.

Still, small jobs are where you learn the scheduler with nothing at stake. If a reminder fails, the filter gets changed a week late.

If your weekly review fails, you make a decision on old numbers. A version you can paste:

On the 1st and 15th of every month at 9:00, read last-service.txt. If the date in it is more than 40 days ago, remind me to book the water-filter service. If not, do nothing.

What does Cowork cost, and will I hit the usage limits?

Cowork needs a paid plan: Pro, Max, Team or Enterprise. The account my Mac jobs run on is on Max.

The limits are real.

Under a Cowork tutorial with 1.37 million views, one of the most-liked comments is a single line: “The worst thing about Claude is the usage limits.”

Long, multi-step jobs use far more of your allowance than a chat, because Claude reads, checks and writes many times inside one task.

Remember the 47 steps.

Here’s where they bit me.

One morning in September, two jobs on my laptop that should have run before 8:00 fired at 9:23, when the laptop woke up, and produced nothing.

The likely cause was a usage window. That was the old rule: no awake computer, no job.

That rule is changing. From October 6, scheduled tasks on Pro and Max run in the cloud and keep going when your laptop is closed.

The catch: a task that reads files in a folder on your computer still needs the desktop app open.

If a job depends on a local folder, keep the app running or move its inputs into a tool you’ve connected.

What I don’t hand to Cowork

• Anything that speaks for me to a real person. Drafts, yes. Sending, no.
• Money. A reminder to pay, yes. Paying, no. The same goes for ad spend.
• Deleting real data. A wrong file sort is annoying. A wrong delete is gone.

• Code. Endii customers couldn’t add yesterday’s check-in.

  The fix was patched, tested, translated into French and Spanish and ready to ship the same day.

  That kind of job belongs to Claude Code, not Cowork.

• Jobs nobody reads. I had a job refreshing an ops dashboard every afternoon at 15:30.

  I wasn’t opening the dashboard. It’s on the cut list. A job you don’t read is a gym membership.

One more rule from my own setup: a cheap model can do the work, but a cheap model never judges what reaches me.

How do you know a scheduled job is actually running?

You don’t, unless you check. In one two-week stretch my setup had six real outages. My VA or I caught four.

An audit caught one. One job caught itself, weeks late. The monitors I’d built to catch them caught none.

Here’s what I check now. Each one cost me days.

• A task in a folder proves nothing. My daily dashboard job still had its instructions on disk, but nothing was calling it.

  Open the Scheduled page, find the task and run it once by hand.

• Silence is not health. One alert fired once, on 20 July, then stayed quiet for five weeks while the problem stayed broken.

  Bad news should repeat until it’s fixed.

• Read what ran, not what you asked for. My server jobs ran on the wrong model for seven weeks.

  The setting was written to a log and never used.

  Open a past run and read it.

• Make it admit gaps. “No fresh data today” beats a confident old number.

  Making things up is the worst failure there is. It’s worse than silence.

• One job, one path. Two jobs synced the same files with different rules.

  One of them wiped a file every night for a week.

Where should a small business start with Cowork?

One job. By hand.

Pick the weekly review. Put two snapshots of your business in a folder, run the prompt above and read what comes back.

Fix the inputs until it’s right twice, then schedule it.

Add a second job only after you’ve read the first one every week for a month. Trust is a muscle. You build it with reps, not with more connections.

My free guide to building a personal AI assistant gives you the whole thing, ready to open.

A working business dashboard with sample data. A guide to adapting it with Claude.

And the weekly review job, with the files it reads.

Questions people ask about Claude Cowork

Is Claude Cowork still called Cowork?

Partly.

On September 16, 2026, Anthropic started merging Cowork and chat into one Claude, rolling out to Pro and Max plans first, with other plans to follow.

If your message box no longer shows a Cowork option, describe the task in any conversation and Claude treats it as a task.

Does my computer need to be on for Cowork tasks?

From October 6, 2026, not for most tasks on Pro and Max plans. New tasks and scheduled tasks run in the cloud.

Anything that reads a folder on your computer, uses your browser or controls your screen still needs the Claude desktop app open.

Do I need a paid plan to use Claude Cowork?

Yes. Cowork is available on Pro, Max, Team and Enterprise plans.

Long, multi-step tasks use more of your allowance than chat, so if you plan several daily jobs, watch your usage in the first week.

Should I use Cowork or Claude Code for my business?

Use Cowork for work that ends in documents, drafts, spreadsheets and summaries.

Use Claude Code when the job is software: fixing an app, building a site, anything that ends in code.

Claude Code also keeps your folders and history on your own computer, which Anthropic suggests for work that has to stay on one machine.

Keep going.