How to Use Claude AI for Business: The Six Levels, and Where to Stop

The video version, 13 min 30 sec. Watch on YouTube

• 0:00The ladder
• 1:19Level 1: talk to Claude
• 3:23Level 2: give it your world
• 4:59Level 3: make real things
• 6:04Level 4: make it know you
• 7:33Level 5: connect your apps
• 9:13Level 6: it works without you
• 11:07Above level 6: Claude Code
• 12:15The verdict

Most people who use Claude at work are on level one. They type a question, read the reply and close the tab. That works, but it leaves most of the product unused, and the useful parts are not hidden behind a paywall.

This guide walks through all six levels with one running example: a small coffee roastery called Third Window. Following a single business the whole way makes it easier to see what each level adds. Swap in your own company, a client or a campaign as you read, because the levels behave the same way.

A note on shelf life. Menus move, so follow the reasoning rather than hunting for a button in exactly the place described. Prices, plan limits and model names were checked against Anthropic's own pages in September 2026. Treat the structure as stable and the numbers as something to re-check.

What is Claude AI?

Claude is an AI assistant made by Anthropic. You write to it in plain language and it replies, reads documents you give it and produces files. On paid plans it can also act on connected apps and folders without you watching. It runs in a browser, as a desktop app and on phones.

The chat box is where everyone starts, and it is the smallest part of the product.

Level 1: talk to Claude

Learning the interface takes five minutes. Asking well takes longer, and it explains most of the difference between a generic answer and one you can use.

Give every request three parts:

• Context: what Claude cannot know on its own. Who you are, what you sell, who buys it.
• Task: the specific job. Not "write some copy" but "write the bag description for the Ethiopia Guji".
• Format: the shape of the answer, including length and tone.

Ask for coffee copy with none of that and you get words that would suit any roastery anywhere, because nothing in the request narrowed it down. Add the three parts and the reply is close to printable. The model, the plan and the price stay the same.

The second habit saves more time. When a reply is nearly right, answer it rather than opening a new chat. Everything you have already said stays in play, so corrections build on each other: make it half the length, keep the last line, drop the third paragraph. Treat the first reply as a draft.

For a team this is the cheapest fix there is. Put the three parts into your brief template and everyone's first drafts improve the same day.

Which Claude model should you pick?

Model names change every few months, so remember the tiers rather than the list. Haiku is the fast one. Sonnet covers most everyday work. Opus and Fable are the slower, pricier ones, worth using when a mistake would be expensive to fix. We compare all four in which Claude model to use.

There is also an effort setting, which controls how long Claude reasons before it answers. Turn it up for problems with many steps that depend on each other.

Both draw on a usage allowance that refills on a cycle, so running out on the free plan costs you time rather than money. A sensible point to start paying is when the limit stops you twice in the same week.

Level 2: give Claude your material

At level one Claude knows only what you typed. At level two it reads your material directly, which saves you summarising it and removes the mistakes that creep in when you do.

One piece of housekeeping first: keep each chat to one job and start a fresh one when the subject changes. Long threads drift off topic, and they use more of your allowance than several short ones.

Documents are the obvious win. A wholesaler's price list is the kind of PDF that normally means a calculator and twenty minutes. Asking which option is cheapest per kilo, and where the catch is, took seconds. It found the condition in the small print: the cheapest lot came with a sixty kilo minimum order.

Images work the same way. Photograph a product on a table and ask what is working against it. Here the background was competing with the product for attention. Screenshots, handwritten notes, contracts and bank statements can all be read.

Search is free and shows where each claim came from. Open at least one source whenever the answer matters, with this tool or any other.

Research is the paid version. It reads well over a hundred pages and returns a report with every source cited. It also uses allowance faster than anything else here, so save it for decisions with money attached, like changing supplier or signing a contract.

Level 3: make real things

Up to now, everything Claude produces is text in a chat. Level three produces files you can open, send and keep.

Ask for a spreadsheet to track roasts, with eight sample rows and a margin column, and you get a real spreadsheet file. The margin column holds a working formula rather than typed numbers, so changing a cost updates every margin. Documents, slide decks and PDFs come out the same way, and paid plans add dedicated tools in Claude Docs and Claude Slides.

The second kind of output is built live beside the conversation. A landing page for the roastery appears while it is being written, and you revise it by asking: warmer background, bigger headline. You can then take it somewhere you had not planned, such as turning the page into a quiz that recommends a roast and clicking through it yourself.

That took one sentence and about two minutes.

Level 4: make Claude know your business

Every level so far carries a hidden cost. Each new chat treats you as a stranger, so you re-explain the same background before any work starts. Level four removes that, and for most teams it is the biggest single improvement on offer.

A project is a shared workspace. Several chats live inside it and use the same material. You upload files once and write instructions once. For the roastery that meant the supplier list, the research report and a three-line brief: small roastery, plain warm voice, customers who brew at home. Writing good instructions is a skill of its own, and Kurt's breakdown of a 65-line instruction file built from Karpathy's complaints is a useful worked example.

After that, requests get short. "Give me three angles for the launch email" comes back already aware of the products, the tone and what is in stock, because the context lives in the project rather than in every message.

For a studio or an in-house marketing team, the natural setup is one project per brand or client, holding the brand guidelines, tone of voice notes and a few past campaigns. New briefs then start from the brand instead of a blank page.

Memory works from the other direction. Claude keeps things it notices while you work, without being asked, and carries them across the web app, desktop and phone. You can read what it has stored, correct it by telling it, or delete it.

If you would rather nothing was kept, the ghost icon opens an incognito chat. It is not saved, adds nothing to memory and is not used for training.

Level 5: connect your apps

This is the first level where Claude works on live data rather than a copy you pasted in.

Connectors cover Calendar, Drive, Gmail and Slack and need nothing beyond your authorisation. After that you ask in plain language. "Where did my hours actually go last week?" gets answered from the calendar itself, with nothing exported or pasted.

Read the authorisation screen properly. It lists exactly what you are granting, and you can revoke each item later.

MCP covers everything else. It is an open standard that works like a universal adapter, letting any tool make itself available to Claude, and you connect one by pasting in a server address. Expect two things: the first authorisation attempt often fails, which is normal, and the free plan allows one connection, so choose it carefully.

Paid plans can also run Claude inside Chrome, so it can work on the page you already have open.

Skills complete this level. A skill is a written procedure for one recurring job: the steps, their order and the standard to hit. Install it once and it is available in every later conversation, called by name with a slash or picked up by Claude when the job fits. The same request on the same deck comes back noticeably better with the right skill installed, because the skill spells out the standard your prompt left unsaid.

Level 6: Claude works without you

Until now you have been present for everything. Level six is where you stop supervising.

This launched separately as Cowork and has since been folded into Claude, where it sits next to Chat on web, desktop and mobile. It needs a paid plan.

You start by giving it access to one folder, and nothing else on the machine is visible to it. Then give it something tedious, such as a folder of supplier invoices with an instruction to pull the date, supplier and amount from each and compile them into one sheet.

Before it runs, you choose how closely to supervise: approve every action, carry on but pause for anything risky, or never pause. Run the first job on the strictest setting however simple it looks. Watching how it handles your files once teaches you more than any description of what it should do.

It works through the files and saves the finished sheet into the folder as a real file, not a message you have to copy out.

Scheduled tasks are the same thing on a timer. Define the job, choose how often it runs, and trigger a test run so the permissions are sorted before the first real one. The Monday report can be ready before you are.

Small businesses also get a library of ready-made jobs instead of a blank page: a Monday brief, closing the books, proposals, lead follow-up and marketing. They connect to software small companies already pay for, including QuickBooks, Shopify, Stripe, Xero and HubSpot, with roughly thirty more available.

Beyond level six: Claude Code

The name puts off the people most likely to benefit from it. Claude Code uses the same capability as level six, pointed at building software instead of sorting files, and you talk to it in ordinary English. It writes code because software is made of code; you never have to read any of it.

A site design handed to it became a working website in about twenty minutes. Two parts arrived broken, the images and the email signup, and describing the problem was enough to get both fixed.

The more interesting example: at level six, Claude sorted a folder of invoices once. Asked to build a tool instead, it made one. Add a coffee to the spreadsheet, double-click, and printable labels come out. That tool runs without opening Claude at all, so there is no chat window and nothing comes out of your allowance.

It goes a long way past labels. I used it to make a seven-minute animated film from a one-line brief. Most teams do not need to climb this high yet, though.

Which level should your team stop at?

Level four, for reasons that have nothing to do with what the higher levels can do.

Levels one to four improve work people already do every day, whatever the job. Levels five and six only repay their setup once you have a task that is repetitive, clearly defined and dull. Most teams have not found one yet, so they build the automation first, go looking for something to feed it, and it sits unused.

Get the team to level four and stay there for a month. If five and six are worth it for you, the need will show up without you looking for it.

Where to start this week

1. Write one request with all three parts, context, task and format, and compare the reply with how you asked yesterday.
2. Take the dullest document you have (a lease, a policy, a bank statement) and ask Claude to explain it back to you.
3. Create one project, add two files and write three lines of instructions.

The third takes longest and pays back the most. I stayed on level one for months before doing it, which is most of the reason this guide exists.

Frequently asked questions

Sources

Everything here was checked against Anthropic's own pages in September 2026. Prices, plan requirements and model names move, so keep the shape and check the current figure.

• Claude
• Claude pricing and plans
• Anthropic, model overview
• Kurt Loy, Claude AI tutorial for beginners (the companion post to the video)