A guide to Claude for business, for owners deciding how to use it. 17 min read.
Claude for business: a working guide.
Search for Claude for business and you get news and course pages. This is the working version: which jobs to hand over, how to check them, and how to keep client data safe. I run my own businesses on it.
Five jobs to hand over
With what good looks like and a prompt for each
Client data kept safe
Logins, what not to paste, one project per client
A team that uses it
A rollout that starts with one job per person
The Fit
What Claude for business is good for, and what it is not
It reads, sorts, drafts and researches faster than anyone you could hire. It also starts out knowing nothing about your business.
Here's the mindset shift. Stop thinking of AI as a magic wand. And stop using it like it's Google. Think of it as an overly enthusiastic intern. It's fast and keen, and it will have a go at anything. It also has no idea how your business runs until you tell it. You're the expert. Your job is to ask one question: what job can I give it?
Good for
Reading and sorting a pile
An inbox, a chat log, a folder of notes. After one webinar with about 650 people in the chat, I gave Claude the raw chat. It came back as a call list in three tiers, and the closer started ringing that day.
First drafts
Replies, proposals, reports, posts. It gets you most of the way in minutes. You finish it.
The boring weekly job
The report someone builds by hand every Friday. These are the biggest wins, and nobody brags about them.
Research
It reads far more than you have time to, then tells you what it found and where.
Not for
Knowing your business on day one
It knows a little about a lot and nothing about you. Until you tell it, every answer is a guess.
Designing the system
AI builds. It doesn't design. I once asked it to restructure a product's databases. My senior engineer opened the result and called it one of the worst things he had seen in his life.
The final call
Anything that goes to a client, moves money or carries your name needs a person to say yes first.
Jobs that don't need AI
Sometimes the right tool is a plain automation that does the same thing every time. Forcing AI into it adds risk for no gain.
The model doesn't get smarter on its own.
The model is the same every time you open it. What gets better is the context you give it: who you are, how you work, what you decided last month. So the setup in the next section matters more than any clever prompt. Structure beats raw intelligence.
The Setup
The three things that make it useful
What it knows about you, how you do things, and the tools it can reach. Miss one and you are back to pasting the same background every morning.
Most people open a blank chat, type a question and get a generic answer. Then they decide Claude isn't that good. The answer was generic because Claude had nothing to work with. Three things fix that, and all three are built into Claude.
What it knows about you
Memory and projectsA project is a workspace with its own chats, files and instructions. Load it with your offers, your clients and samples of your writing. Memory sits beside it. Claude saves what it learns as you chat, by topic, and you can read, edit or delete any of it.
Do this first. Make one project for the business. Add a one-page brief: who you serve, what you sell, how you sound. Write it once and every chat in that project starts from it.
Source. Anthropic on projects, Anthropic on memory.
How you do things
SkillsA skill is your method for one task, written down so Claude follows it. Claude works out when a skill is needed from what you ask, so nobody has to remember to switch it on. You can write your own and upload them.
Do this first. Pick the task you explain most often. Do it right once with Claude, then ask Claude to turn what you just did into a skill. Next time it's one line.
Source. Anthropic on skills.
The tools it can reach
ConnectorsA connector links Claude to an app you already use, such as your email, calendar or documents. It can then read your data and take actions there. It gets the same access you have in that app and no more.
Do this first. Connect email and calendar first. That's where most of an owner's day goes. Leave everything else off until a job needs it.
Source. Anthropic on connectors.
Set them up in that order: memory, then skills, then tools.
A connector with no context sends a generic email from your real inbox. Context first. Method second. Reach last. For more on how a skill is built, read our guide to agent skills.
The Work
Five jobs to hand over first
Each one comes up every week, is easy to check, and has a prompt you can copy. Start with one.
Pick jobs that come up every week and take minutes to check. These are the five we hand over first. Each prompt works in a plain chat. It works far better inside a project that already knows your business. What each connector can do is on Anthropic's page on the Google Workspace connectors and its page on Research.
Inbox triage and drafts
What it needs. Your email connected. Claude can search, read and draft, and by default it asks you before it sends anything.
What good looks like. Every email is sorted. The urgent ones have a draft that sounds like you. Nothing has been sent. Gaps are marked as questions, not filled with guesses.
Meeting follow-ups
What it needs. The transcript or your notes, pasted in or added to the client's project.
What good looks like. Each action has an owner and a date. Decisions are listed apart from discussion. The follow-up email is short enough that people read it.
Proposals and reports
What it needs. A project holding two or three past proposals or reports you were happy with, plus your offer and terms.
What good looks like. It follows your structure. Every number comes from something you gave it. Missing facts are flagged, not made up.
Content in your voice
What it needs. A project with samples of things you wrote yourself, plus your real stories and client results.
What good looks like. You would post it after a light edit. It uses your words and your stories. It has no made-up client and no made-up number.
Research
What it needs. Web search turned on. On paid plans, Research mode runs many searches that build on each other.
What good looks like. Every claim has a source you can open. It says what it could not find. It keeps what the sources say apart from its own view.
One job, done properly, beats five done badly.
In a live workshop with a client, we turned one weekly progress report into a skill. By hand it took up to four hours. The skill ran it in 22 minutes while the client watched. That's one boring job, and the hours come back every week from then on.
Enjoying the guide?
Enter your details to unlock the remaining insights. We'll also send you the complete guide as a reference.
The Method
Check the work: 10/80/10
You plan it. Claude does the middle. You make it right before it goes anywhere.
A task I used to do all by myself now breaks into three parts. I do the first 10%. Claude does the 80% in the middle. I do the last 10%. Most people skip the last part, and that's why so much AI slop goes out. Plenty skip the first part too, because they think it's a magic wand.
10% · You
Plan it
Say what the job is and what done looks like. Give it the background and the files. This is the part that matters most.
80% · Claude
Do it
It turns the plan into the thing: the quote, the document, the reply. This is the easy bit.
10% · You
Make it right
Go through it. Fix what is off. Make sure it sounds like you. Then you send it, not Claude.
Five checks before anything leaves
Would I put my name on this as it stands?
Is every fact, number and name from something I gave it, or can I open the source?
Did it do the job I asked for, or a nearby one?
Does it sound like me, or like everyone else's AI?
What did it leave out?
You're the human in the loop at both ends.
Never let AI go off and do a job on its own from start to finish. The quality of the work lives in the two 10%s you can't hand over: the plan at the front and the polish at the end. In our business nothing is sent, posted or paid without a person saying yes.
Safety
Keep client data safe
Your Claude login is a bank login now. We found that out the hard way, so you don't have to.
We ran the whole team on one shared Claude account. We knew it was risky. We never fixed it.
One day the person who ran the account was stuck in a meeting. A developer needed admin access to change a setting, so the sign-in link got handed over. Quick and easy. That developer's computer had a Trojan on it that none of us knew about. Overnight, strangers added about ten people to our plan and drained the credits. My business partner happened to be up late and saw the spend climbing. He kicked them out and set hard spend caps.
It cost us real money. Worse, what they did with the account broke the terms of service, so our whole team was locked out. All of that came from one shortcut, taken by someone more careful than I am.
Six rules we run now
One login per person
No shared account. No passing a sign-in link around. On the Team plan each member has their own seat, and admins can set spending caps and turn on single sign-on.
Lock the inbox behind it
If you sign in by email, whoever gets into that inbox gets into Claude. Put two-factor on every email account that can sign in.
One project per client
Each project has its own files and its own memory, and memory from one project isn't used in another. Keep client work inside its project and one client's details stay out of another client's draft.
Know what your plan does with your data
Anthropic says that by default it doesn't use inputs or outputs from its commercial products to train its models. There are exceptions, such as feedback you choose to send. Read its page on model training. Personal plans have their own privacy setting, so check each account.
Connect only what a job needs
Anthropic's own advice is to connect only services you trust and need. Leave Claude asking before it sends an email. On Team plans an owner can set each connector tool to always allow, needs approval or blocked. It's all on the connectors page.
Check a skill before you install it
A skill from a stranger is a set of instructions you didn't write. A bad one can quietly send your data somewhere else. Read it first, or don't install it.
Never paste these
For a one-off chat you don't want kept, Claude has an incognito chat. It isn't saved to your chat history and doesn't feed memory.
Treat the boring security step as part of the cost of running AI.
Your AI accounts are a target now, the same as your bank. A new password wouldn't have saved us. Separate logins and a spend cap would have.
The Team
Roll it out to a team
Setting it up takes an afternoon. Getting people to use it is the real job.
Now here's the important bit, and it's the part most people skip. The tech is the easy 20%. People are the other 80%. So treat getting it used as part of the job, not something you hope for. Frame it as career security for your team, because that's what it is.
Give every person their own seat
Start here, for the reasons in the last section. A team plan also puts billing and access in one place, so you can see who has what.
Pick one person to own it
Someone credible, with the job written into their role. Watch for the person already teaching themselves in the background. Being senior is the wrong filter.
Start each person on one real job
Their own job, from their own week, done hands-on. People need to use it, not watch a demo. Ask what part of their day could be handed over so they can do the rest better.
Share what works
When one person gets a job working, turn it into a skill or a shared project. On Team plans an owner can give a skill to everyone, and projects can be shared with view or edit access.
Write the rules down
One page. What never gets pasted. Which tools are connected. Who says yes before anything leaves the business.
What a team plan includes is on Anthropic's Team plan page. Sharing is covered on its pages for skills and projects.
The champion is usually not who you would pick.
I trained one agency for about a year. The founder set up an AI council, three or four people who set the AI rules for the business. One of them was a junior designer, still on probation. He had been teaching himself the whole time and bringing the founder ideas. He was the newest hire and the youngest person there, and he moved up faster than anyone.
What Comes Next
When chat is not enough
A chat window helps one person at a time. An operating system runs the same method for everyone, and keeps what it learns.
People think every AI tool is the same chatbot. It isn't. There are three ways to work with Claude, and they are as different as watching a cooking show, hiring a chef and running a restaurant. Anthropic's page on Cowork covers the middle one, and our guide to the Claude stack walks through the rest.
Watching a cooking show
Chat
You ask, it answers, you still do the work. Great for ideas. Most people stop here.
Hiring a chef
Cowork
You describe the result and Claude plans and does the task inside your files and tools.
Running a restaurant
Claude Code
It runs the whole operation: builds, scheduled jobs, the full stack. We live here.
Signs you have outgrown chat
How we run ours
Think of a phone. The apps you see sit on an operating system you don't. Ours is the same: one shared layer of memory, skills, connectors and scheduled jobs under every piece of work. I had never coded in my life before AI, and this is what my businesses run on.
Claude is the main driver. Two other models check its work, because no model signs off its own work.
A standing team of AI staff runs the day to day: a project manager, an assistant for inbox and calendar, then sales, marketing and content.
Everything we learn from meetings and calls goes into a wiki. The AI reads it before it acts.
Jobs run on their own clock with nobody at the keyboard, and flag anything that needs a person.
They write drafts only. Nothing is sent, posted or paid without a human yes.
Most AI setups reset every time you open them. Ours gets sharper every week.
The same three things from earlier are doing the work here: what it knows, how you do things and what it can reach. They are joined up and they run on a clock. For a plain definition, read what an AI operating system for business is on the Njin blog. For a worked example in one industry, read how to run a B2B education business on Claude.
Questions
Questions owners ask
Short answers to what comes up most.
Is Claude good for business use?
Yes, for the right jobs. Claude is strong at reading, sorting, drafting and research. It knows nothing about your business until you tell it, so the results depend on the context you give it and on a person checking the work before it goes out.
What should I use Claude for first in my business?
Start with a job you repeat every week and can check in a few minutes. Inbox triage and draft replies, meeting follow-ups, proposals and reports, content in your voice, and research are the five we hand over first.
Is it safe to put client data into Claude?
It can be, if you set it up properly. Give every person their own login, keep each client in its own project, never paste passwords, card numbers or ID numbers, and check what your plan does with your data. Only share client data you have the right to share.
Does Anthropic train its models on my business chats?
Anthropic says that by default it doesn't use inputs or outputs from its commercial products, such as the Team plan, to train its models. There are exceptions, such as feedback you choose to send. Personal plans have their own privacy setting, so check each account.
What is the difference between projects, skills and connectors?
A project holds what Claude knows about one area of your work: files, instructions and its own memory. A skill holds how you do a task, so Claude follows your method each time. A connector links Claude to a tool such as your email or calendar, so it can read your data and take actions there.
Do I need to be technical to use Claude for business?
No. The skill that matters is giving clear direction, the way you would brief a new hire. You say what the job is, give it the background, and check what comes back.
How do I roll Claude out to a team?
Give each person their own seat, pick one person to own it, start everyone on one real job, then share what works as a shared project or a skill. Write down what must never be pasted and who approves anything that leaves the business.
Is this guide a Claude course?
No. A course teaches you the tool lesson by lesson. This guide is about putting Claude to work in a business: which jobs to hand over, how to check them, how to stay safe and how to bring a team along.
TellTellClaudeClaudeaboutaboutyouryourbusiness.business.ShowShowitithowhowyouyouwork.work.ConnectConnectthethetoolstoolsititneeds.needs.ThenThencheckcheckeverythingeverythingbeforebeforeititgoesgoesout.out.That'sThat'sthethewholewholejob.job.
Jump back to a section
Read next
Primary sources
Want this set up around your business, with someone beside you who runs on it every day? I do that 1 on 1. We build your memory, your skills and your tools together, on the call.
See 1 on 1 Claude training with JamesClaude is a trademark of Anthropic. AI Orchestrators is independent and not affiliated with Anthropic.