Pick one task you repeat every week, write its steps and rules into a Claude Project's instructions, test it on twenty past cases, and measure the time it saves. Claude Pro costs $20 a month ($17 billed annually); Team starts at two seats from $20 each a month billed annually, or $25 monthly. You can build and test on the free plan first.
The word "automate" can mislead here. In this first workflow Claude drafts and a person checks, which is closer to a fast assistant than a robot. That is the right place to begin, because you learn where Claude is reliable before it touches anything a customer sees. Prices and plan limits change, so confirm them on claude.com before you pay.
Choosing a task that will actually work
Most first attempts fail at the choice, not the tooling. People pick the most annoying job they have, and the most annoying job is often one that depends on judgement, politics or information that sits in somebody's head. A good first task has five traits.
- It repeats at least weekly, so you collect evidence quickly.
- The inputs arrive as text, a spreadsheet or a PDF, not as a phone call or a face-to-face conversation.
- You can write down the rules in a page, because a new hire could follow them.
- A mistake is cheap and visible before it does damage, because a person reads the output first.
- You can find twenty past examples, with the answer you gave, to test against.
Score a handful of candidates against those traits. Here is an illustrative comparison for a specialist food importer that sells to restaurants and delis, with the owner and two staff.
| Candidate task | Weekly volume | Rules writable? | Past examples? | Risk if wrong | Verdict |
|---|---|---|---|---|---|
| Replying to trade price and availability enquiries | 25 | Yes | Yes, in sent mail | Low, checked first | Start here |
| Chasing late invoices | 6 | Mostly | Some | Medium, tone matters | Second |
| Writing supplier purchase orders | 4 | Yes | Yes | High, money moves | Later |
| Deciding which customers get discounts | 2 | No | No | High | Leave with a person |
| Summarising delivery-delay emails for customers | 8 | Yes | Few | Low | Good third |
The enquiry replies win: high volume, writable rules, plenty of history, and a human reading every draft. The discount decision loses, because the rules live in the owner's head and the downside is real. If your own table produces a tie, choose the task whose twenty past examples are easiest to collect, because testing is the step people skip when it is tedious.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhich Claude plan to start on
You do not need to commit money before you have evidence. Free accounts can create up to five Projects, so the whole build and test can happen at no cost, within the usage limits that reset on a schedule. Once you are using it daily, move to a paid plan. As of November 2026, the individual and team prices on Anthropic's pricing page are these.
| Plan | Price (USD) | Good for |
|---|---|---|
| Free | $0 | Building and testing a first Project |
| Pro | $20 a month, or $17 a month billed annually | One person running the workflow |
| Team, standard seat | $25 a month, or $20 billed annually, 2 to 150 seats | Sharing a Project across staff |
| Team, premium seat | $125 a month, or $100 billed annually | Heavy daily users |
| Max | $100 a month | One person with very high usage (five times Pro) |
| Ultra | $200 a month | One person with the heaviest use (ten times Pro weekly) |
For a one-person business the choice is Pro. A Team plan needs at least two seats, so a sole trader would pay for two, which is about $40 to $50 a month, unless they have a second person who will use it. The reason to pay for Team is sharing: projects can be shared with colleagues on Team and Enterprise plans, and the plan does not train on your content by default. If three people will use the same enquiry-reply Project, Team at $20 per seat on annual billing is $60 a month, and everyone works from the same instructions. The Claude Team tutorial goes through that decision.
The API, Anthropic's pay-as-you-go service for developers, is billed separately from Pro. The pricing page now lists monthly API credits on some plans ($100 on Max, $200 on Ultra and up to $500 pooled on Team), which matters only if someone in the business builds on the Claude Platform. You do not need the API for anything in this tutorial.
Writing the instructions your Claude Project follows
A Claude Project holds two things: instructions that apply to every chat inside it, and knowledge files that Claude can read. The instructions are the workflow. Spend most of your time here. Write them as you would brief a capable new colleague on their first morning, covering the job, the rules, the tone, what to do when unsure and the shape of the output.
Below is a filled-in example for the importer's enquiry replies. Everything in it is illustrative.
ROLE
You draft replies to trade enquiries for a specialist food importer.
A staff member reads and sends every reply. You never send anything.
WHAT I WILL PASTE
An email from a restaurant or deli asking about price, availability,
minimum order or delivery.
WHAT YOU MUST DO
1. Identify what is being asked (price, stock, minimum order, delivery,
new account, something else).
2. Answer only from the files in this Project (price list, delivery
terms, account terms). Quote prices exactly as listed.
3. If the answer is not in the files, do not guess. Say what is
missing in a note to staff at the top, and draft a holding reply.
RULES
- Minimum order is 3 cases per line, delivery free over 12 cases.
- Never offer a discount or a payment extension. Flag the request.
- Prices on the price list are per case, ex delivery.
TONE
Friendly, brief, plain English. Greeting "Hi [first name]". Sign off
"[your name]". No exclamation marks. Under 120 words unless a table
of prices is needed.
OUTPUT FORMAT
STAFF NOTE: one line (what was asked, anything missing or risky)
DRAFT REPLY: the emailNotice what is in it. The role says a human sends. The job is split into steps. There is an explicit instruction for what to do when the answer is missing, which prevents the single worst habit of language models, which is making something up to be helpful. The output has a staff note at the top so that the reviewer sees risk before reading the draft.
The knowledge files
Upload the documents Claude needs to answer from: the current price list, delivery terms, minimum orders and account terms. Project files are capped at 30 MB each, which is plenty for text and spreadsheets. Keep the files short and current. A price list that is three months old will be quoted with complete confidence, so put a "valid from" date at the top of each file and add one line to the instructions: if the price list is more than 31 days old, say so in the staff note.
Testing on twenty past cases before anyone relies on it
Take twenty real enquiries from the last two months, with the replies you sent. Choose a spread: a few easy ones, a few awkward ones, one where you said no, one where information was missing. Paste each into the Project in a fresh chat and score the output against what you actually sent.
Here is an illustrative pair. The enquiry reads: "Hi, can you do the 5 litre olive oil tins at 8 cases a week? Do you deliver Tuesdays? Also is there any discount for a standing order?"
STAFF NOTE: Asks price/volume, delivery day, and a standing-order
discount. Discount not allowed by my rules: needs owner decision.
Delivery days not in the files: confirm before sending.
DRAFT REPLY:
Hi [first name],
Thanks for getting in touch. The 5 litre tin is $62.00 a case on the
current list, and 8 cases a week is well above our 3-case minimum.
Delivery is free on orders over 12 cases, so at 8 cases there would
be a delivery charge, which I will confirm shortly.
I will check which days we can deliver to you and come back about
Tuesdays. On a standing order, I will need to ask about terms and
get back to you.
Best wishes,
[your name]What you would fix: the draft is correct about the rules, it flagged the discount and the missing delivery days, and it did not invent a Tuesday run. But "a delivery charge, which I will confirm" is vague, because the delivery charge table was not in the files. The fix is to add the table to the knowledge files, not to correct the draft by hand each time. This is the pattern for the whole test: every error is either missing information (add a file), a missing rule (add a line to the instructions) or Claude not following a rule (rewrite it more plainly and put it nearer the top).
A scoring sheet that takes ten minutes
Create a sheet with one row per test case and these columns: case number, matches what I sent (yes, nearly, no), the problem, fix needed, fix type (file, rule, wording). After the twenty, count the results. A reasonable bar, illustrative, is at least fifteen yes or nearly, with no case in which the draft would have embarrassed you or cost money. Anything below that, apply the fixes and run the same twenty again. Re-running the same set is the point: you are measuring whether the instructions got better, not whether a lucky draw went well.
Measuring what it saves, with real sums
Time ten enquiries the old way. Say each takes six minutes, which includes reading, checking the price list, typing and sending. Then time ten with the Project: paste, read the draft, fix a line or two, send. Say that comes to two minutes. The sums, illustrative, for 25 enquiries a week:
- Before: 25 x 6 minutes = 150 minutes a week.
- After: 25 x 2 minutes = 50 minutes a week.
- Saving: 100 minutes a week, roughly 7.2 hours a month.
- Value at $30 an hour of staff time: about $217 a month, against a plan cost of $20 on Pro or $60 for three Team seats.
That is the optimistic version. Take your own measured times and be honest about review: if you find yourself rewriting half of each draft, the real after-time is nearer four minutes and the saving shrinks by half. There is also a point about where time goes. Saved minutes are only worth money if they get spent on something: more calls to buyers, or going home earlier. The tutorial on why AI saves time but not money covers how to capture the gain deliberately.
Moving from copying and pasting to a connected workflow
For the first month, copy and paste. It is slower than a connected setup, but it keeps every step visible. Once the drafts are consistently good, you can reduce the handling. Claude's connectors can link it to tools such as email, calendar and documents, and the Gmail connector can now draft, send, reply and forward, asking for your approval before each send, reply or forward by default. On Team and Enterprise plans an owner decides whether members may skip that approval, so check the organisation setting as well. A sensible stepping stone is a connector that creates a draft in your inbox rather than sending, so that the human review stays in the place where you already work.
Check what each connector is allowed to do before you switch it on, and keep the approval prompts on. The tutorial on connecting Claude to business apps goes through what is safe to enable first. Resist the urge to let a first workflow send anything by itself. The fully automatic version belongs after months of logs showing that corrections are rare.
When the Project quotes something out of date
Consider an illustrative kitchen fitter who set up a Project to turn site-visit notes into a customer-ready quote summary. For three weeks it worked well. Then the supplier raised worktop prices, the price file in the Project was not replaced, and a handful of summaries went out with old figures. Nobody noticed because the drafts looked right: the formatting was perfect and the wording was calm.
It surfaced when a customer queried a figure that did not match the final invoice. The cure is cheap. Put a "valid from" date at the top of every file, tell Claude in the instructions to state the date it used in the staff note, and put a recurring reminder in your calendar to refresh the files on the day your supplier prices change. A fluent draft is not evidence that the facts in it are current.
Other first workflows, to show the shape
The pattern above is not specific to enquiry emails. Here are four more that use the same ingredients: repeating input, writable rules, a human check.
- An online clothing store: return requests. Paste the customer's message and the order line. The Project holds the returns policy and drafts either an approval with a label instruction, or a polite refusal quoting the relevant rule. The staff note flags anything outside the policy window.
- A kitchen fitter: site notes to quote summary. Rough bullet notes go in, and a structured summary comes out: rooms, items, assumptions, exclusions and questions for the customer. Claude does not price. A person adds the figures.
- A craft brewery supplier: stock-alert summaries. A weekly stock export goes in, and the Project returns the ten lines below reorder level, grouped by supplier, with a draft order email for each. The reorder levels sit in a knowledge file.
- The importer again: delivery-delay notices. A forwarded carrier email goes in, and a short customer-ready update comes out, listing which accounts are affected from a customer list in the files.
Each of these starts as a copy-and-paste workflow, gets tested on twenty cases and is measured the same way. If you want help deciding between them, the tutorial on what to automate first gives a longer scoring method.
Rewriting an instruction that Claude keeps ignoring
Sooner or later one rule will not stick. Say the importer's draft keeps offering a "small discount for standing orders" even though the instructions say never to offer one. The temptation is to add capital letters. A better route is to look at how the rule is written and to change three things.
- State the rule once, positively, with the reason. Before: "No discounts!!" After: "Never offer or hint at a discount, because only the owner can approve one. If a customer asks, write 'I will check and come back to you' and flag it in the staff note."
- Give the alternative action. A rule that only says what not to do leaves a gap, and the gap gets filled with something helpful and wrong. Say what to write instead.
- Add a worked pair. Put one short example of an enquiry that asks for a discount and the correct draft into the instructions. One concrete example does more than a paragraph of explanation.
After changing it, rerun all twenty test cases, not just the one that failed. A fix for one rule can change how another behaves, and the full rerun is how you find out in the sheet instead of in a customer's inbox. Keep the old version of the instructions in a document with its date, so you can go back if a change makes things worse.
What to keep out of the Project
Be deliberate about what you paste. Business plans do not train on your content by default, whereas individual plans have a privacy setting that controls whether your chats help improve the models, so check yours in the account settings before you put customer details in. Even on a business plan, apply the principle of least data: a reply drafter needs the enquiry text and the price list, not the customer's full account history or bank details.
Keep passwords, payment card numbers and anything covered by a confidentiality agreement out of chats altogether. And do not treat Claude's output as the record. The sent email, the invoice and your order system are the record. The Project is a drafting tool.
Running it for four weeks, then deciding
A lightweight cadence gets a first workflow to a clear verdict without much overhead.
- Week one: run in parallel. Write your reply as usual, then compare with what the Project drafted for the same enquiry. No customer sees a Claude draft yet.
- Week two: send the Claude draft after review for the easiest category only, such as stock and price questions. Log every edit you make, in a sheet, with the reason.
- Week three: apply the edits as instruction changes and file updates. Extend to the next category.
- Week four: repeat the timing exercise on ten cases and redo the sums. Count the share of drafts sent with no edit.
The decision at the end is one of three. Keep and extend, if the saving is real and the edit rate is falling. Keep but narrow, if it works on one category and not the others. Or stop and pick a different task, which is a perfectly good result: you spent a few hours and a month of a $20 plan to learn that, and that is cheap compared with a failed larger project. The first pilot project tutorial has a fuller version of this go or no-go thinking.
If you would like someone to look at your week and say which task to start with, that is the sort of thing an AI implementation consultation is designed for.
Questions before you set up a first Claude workflow
Can I start on the free plan?
Yes, for testing. Free accounts can create up to five Projects, so you can build and trial a workflow without paying. The limits are tighter and reset on a schedule, which makes Free awkward once real work starts. Move to Pro when you rely on it daily.
Does Claude send emails or update my systems by itself?
Not in the basic setup. In a Project, Claude drafts text that you copy out. Connectors can let Claude read other apps and draft or send messages; the Gmail connector asks for your approval before each send by default. Check what each connector can write, and keep an approval step on anything customers see.
Is my business data used to train Claude?
Business plans such as Team and Enterprise do not train on your content by default. Individual plans have a privacy setting that controls whether chats are used to improve the models. Check the current setting in your account before pasting customer details, and leave out anything you do not need.
How do I know if the workflow is saving anything?
Time ten real cases by hand before you start, then time the same kind of case with Claude, including your own review. Multiply the difference by weekly volume and compare it with the plan cost. If review takes nearly as long as writing, the task is a poor fit.
What if Claude gets something wrong in front of a customer?
Keep a person approving anything customer-facing for the first month, and log every correction you make. Corrections tell you which line to add to the Project instructions. When a month passes with almost none, you can loosen the checks on the lowest-risk message types.
Further reads
- How to Get Started With AI in Your Small Business: First 7 Steps — Seven first steps for a business new to AI.
- Claude vs ChatGPT for Business Writing, Proposals and Reports — How Claude compares for proposals and replies.
- How to Set Up Claude Projects as a Shared Team Assistant — Share the Project you built with the rest of the team.
- How to Scale AI From One Workflow to the Whole Business — What to do once the first workflow is running well.
- AI vs Rule-Based Automation: Which Does Your Task Need? — Decide whether a task needs AI or a plain rule.
- Why AI Projects Fail in Small Businesses (It's Rarely the Tech) — The reasons first projects stall, and how to avoid them.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: claude.com/pricing; Anthropic help centre articles on Projects, upload limits and the Google Workspace connectors, all read on 5 November 2026. Check them before you buy, because plans and limits change.