Claude for Small Business: A Customer-Reply Assistant You Can Build Today

The Short Answer

Claude for small business can start with one job: drafting replies to your most common customer emails from your own price list and policies. You build it in an afternoon as a Claude Project on the Team plan, and most of that goes to gathering documents. You paste a customer's email in, Claude drafts a reply in your voice using only your documents, and you read it, fix it, and send it yourself. Nothing connects to your inbox.

This is for owners of 5 to 30 person service businesses with no engineers, where one person answers the same quote, scheduling, and policy emails all day. Replies get slow and the voice shifts with whoever is typing. A Project can cut the time to a first draft and steady the voice, and a ten-email test tells you whether it does for your business.

As of November 2026, per Anthropic's pricing page and before tax: Team is for 2 to 150 people. A seat is one person's login, and the Standard seat is where to start, at $20 per person per month on annual billing or $25 monthly. With the two-person minimum, that is $40 or $50 a month. Pro is $17 a month on annual billing, $20 monthly. Team is available in certain supported locations, so confirm yours at claude.com/pricing before you buy.

Why Team Instead of Pro Once Customer Emails Are Involved

Customer emails carry names, addresses, phone numbers, and often what is wrong at their property, so the plan's data terms decide where that text can end up.

As of November 2026, Anthropic's privacy documentation says it does not train its models on inputs or outputs from its commercial products by default, and its pricing page lists "no model training on your content by default" for Team. On personal plans such as Free, Pro, and Max (the pricing page lists Ultra the same way), whether chats help train the models depends on a privacy toggle. Anthropic's pages don't say whether that toggle starts on or off, so check it yourself: Settings, then Privacy, then the switch labeled "Help Improve our AI models."

Two Team details to know before you start:

• Project sharing is a Team and Enterprise feature. On Pro, the Project lives in one person's account, so your office person can't use it. Chats inside a shared Project stay private to whoever had them.
• A thumbs up or down on a response can send that whole conversation to Anthropic, stored up to five years and used for training. A Team Owner can switch that off under Organization settings, Data and Privacy, Rate chats.

Gather the Inputs

Start by listing your 5 to 8 most common email types: quote request, price question, service area, rescheduling, policy question, quote follow-up. Then collect five things. Claude drafts only from these, so anything missing here shows up as a gap in the replies.

1. Your price list, with real numbers, what each service includes, minimums, what moves the price, and what a quote needs from the customer (size, frequency, add-ons).
2. Your policies and FAQ: cancellation, rescheduling, deposits, payment, warranty.
3. Your service area, as cities or ZIP codes, plus what you say to someone outside it.
4. Five to eight past replies you liked, copied from your Sent folder into one document as voice samples. Replace names, addresses, phone numbers, and account numbers with placeholders like NAME and PHONE, because everyone with access can read this file.
5. Twelve real emails, scrubbed the same way. You test with ten and hold two back. Include one or two of each common type, plus a refund request or complaint, an angry email, and one asking about something your documents never cover.

Word, PDF, plain text, and CSV files work. Claude reads only the text, so retype a screenshot of your price list. Files can be up to 30MB each. To change a price later, swap in the new file.

Build the Project

1. Sign up for Team at claude.com/pricing under your business name and invite your office person as the second member. Anthropic's Team plan help article covers the admin screens.
2. As Owner, open Organization settings, Data and Privacy, and switch Rate chats off. Owners also control whether memory is available, and each Project has its own memory, so leave it off while you test.
3. Go to claude.ai/projects, click "+ New Project," and give it a plain name. Claude can't see the name or description, so the guidance goes in the instructions.
4. Click "+" and upload your price list, policies, service area, and voice samples.
5. Click "Set project instructions," paste the block below, replace each [bracket] with your own text, and save.
6. Click "Share project," add your office person, and choose "Can view," which lets them chat but not change anything. "Can edit" lets them change instructions and files.

Anthropic's help center says a new version of Projects is rolling out in stages, with Team to follow. If your screens differ, look for Projects in the left sidebar. The pieces are the same.

You are the email reply assistant for [Business name], a [type of business] serving [service area]. The email types I handle most: [your 5 to 8 types].

You only draft replies to customer emails I paste in. A person reads, edits, and sends every reply.

Rules:
1. Answer only from the price list, policies, and service area documents. Voice samples are for tone only; never take a price, date, policy, or promise from them. Do not fill gaps with general industry knowledge.
2. Never invent a price, date, time slot, discount, guarantee, or availability. If the email is mostly answerable and one detail is missing, draft it and write ##CHECK: what is missing## where it belongs.
3. If the price list needs details the customer did not give, draft a reply asking for exactly those instead of quoting. If you calculate a price, show the math in the Sources line and add ##CHECK##.
4. Match the voice samples: [for example: friendly, short sentences, first names]. Open with [greeting] and sign off with [sign-off]. Answer every question in order. Stay under [150] words unless they asked several things.
5. If the email is on the never-answer list, or its whole topic is absent from the documents, write only "HUMAN ONLY:" and one sentence on why.
6. After each draft, add "--- INTERNAL, DELETE BEFORE SENDING ---", then a line starting "Sources:" naming the document for each fact.
7. If you cannot tell what the customer wants, ask me one short question instead of drafting.

Never-answer list: refunds or credits; fee waivers or discounts not on the price list; complaints; anything legal (lawyers, insurance claims, chargebacks); safety, injury, property damage, or missing items; keys, door codes, alarm codes, or property access; anyone angry or upset.

Test It on Ten Real Emails

Open a new chat in the Project for each email and paste it in. Chats share the Project's files and instructions but not each other's conversation. Score each draft on four checks.

Keep a sheet with ten rows and these columns: number, email type, the four checks, pass, and what went wrong. Mark N/A where a check doesn't apply; N/A counts as a pass. An email passes only if no check fails. Record the result as a number, such as "7 of 10 passed." My suggested bar is 8 of 10 with zero invented facts. One invented price fails the whole test until you find out why.

A failure usually traces to a missing fact in the documents (add it), a loose instruction (tighten it), or voice drift (add a sample). Fix one thing, rerun the failed email in a new chat, and once you reach the bar, try your two held-back emails. Keep the sheet to rerun when a document changes.

Using It Every Day

Start a new chat in the Project and paste the customer's message. Real use can include real details, which is why you chose Team. Fill or delete every CHECK marker, drop everything below the INTERNAL line, then edit and send from your own email.

The Never-Auto-Answer List

Put this on one page, keep it by the desk, and keep it in the instructions too. Anything on it goes to one named person the same business day. If someone pastes one in out of habit, the Project should answer HUMAN ONLY.

• Refunds, credits, and fee waivers or discounts your price list doesn't offer.
• Complaints, such as "you missed a room."
• Anything legal: lawyers, insurance claims, chargebacks.
• Safety, injury, property damage, or missing items.
• Keys, door codes, alarm codes, or property access.
• Anything angry. A fast, polished reply to an upset customer can read as a brush-off.
• Any topic your documents don't cover at all. One missing detail in an otherwise answerable email gets a CHECK marker instead.

You review every reply, but busy reviewers skim, and a polished draft is easy to approve unread. The list catches what a skim would miss.

I run Claude on my own sites under three rules: a human approves anything that publishes, a second reviewer checks the work, and the scheduled reports are read-only. This setup runs on the first rule. Claude drafts and a person decides what goes out, and the emails on the list never get a draft.

What This Does Not Do

This setup drafts replies and nothing else.

• Your inbox stays disconnected, so you copy and paste. Team offers connectors to services like Gmail, but connecting one means deciding what Claude may read, whether it may send, and who approves it, which is a separate project.
• It can't see live data such as today's schedule, a job's status, or a customer's history in your CRM, the software where you keep customer records.
• Someone has to own the Project and decide what happens when that person leaves. An admin can also turn off Project sharing organization-wide.
• Old prices get quoted with full confidence, so date every file and name an owner. Record how long first replies take today, or you can't say whether it helped. The common AI integration mistakes explain why.
• One person using it for a few weeks is a pilot. A whole office needs someone whose job includes keeping it working; see how long an AI integration takes.
• Texts, web forms, and review replies each need their own instructions and their own test.

Why Not Grok Bot or Cursor for This Job

Grok Bot launched in beta in August 2026, and as of November 2026 its vendor (x.ai/bot) calls it generally available for Teams and Enterprise. It is an agent that signs in to your tools and works on its own cloud computer, which is more access than a first project needs. Cursor describes itself as a coding agent for building software. Its Teams plan turns Grok Bot on by default, which is why the names show up together.

Where 11 Mile Co Picks Up

If you'd rather build this with someone beside you, the AI Adoption Working Session is a paid 90-minute live build in the tools you already have. You leave with a working Project in your own account, the prompts or steps to keep using it, and a one-page memo.

A session can prove a pattern on one workflow. Connecting the inbox for real, live prices and schedules, permissions, and a team rollout are the larger work, and that is AI integration. For scope and cost, see what AI integration consulting is and what it costs in 2026.

Once the drafts pass your test and you want them in your inbox, used across the office, or working from live prices, that is when to talk.