Foundation

Part 1 — Foundation

Read this part once. Everything else in the guide assumes you have.

1.1 What Claude can and cannot do for your business

Claude is an AI assistant made by Anthropic. You talk to it in plain English. It can read, write, analyze, summarize, plan, and — when connected to your business apps — take actions in those apps on your behalf.

What Claude does well today

• Reads and writes anything text-based. Emails, proposals, contracts, job posts, SOPs, reports, replies to reviews.
• Works inside your apps. Connected to Gmail, Outlook, Google Drive, Calendar, QuickBooks, Stripe, Slack, your CRM, and thousands more through connectors, it can look things up and take actions.
• Runs on a schedule. A task like "every Monday at 7 a.m., pull last week's sales and email me a summary" runs without you.
• Follows rules you give it. Tell it your pricing, your tone, your policies once, and it applies them consistently.
• Builds things. Dashboards, spreadsheets, documents, simple internal tools, and small apps.
• Analyzes data. Give it a spreadsheet of transactions, it finds patterns, errors, and trends.

What Claude does not do (or should not do alone)

• It does not have a phone line by itself. To answer calls, it works with a voice platform (covered in Chapter 1).
• It does not move money on its own. Paying bills, running payroll, transferring funds, and buying things stay with a human. Claude prepares; you approve.
• It does not replace licensed professionals. It prepares tax organizers; your CPA files. It drafts contracts; your attorney reviews. It flags HR risks; you and your advisor decide.
• It is not always right. It can misread a document or make an arithmetic slip. That is why every chapter has guardrails and review steps.
• It does not remember everything forever by default. You give it standing context through Projects and your Business Brain document (Section 1.5).

The honest summary: Claude can take 60–80% of the administrative work off your desk in most areas, and turn the remaining 20–40% into quick approvals instead of hours of work.

1.2 The tools, one paragraph each

Claude.ai (the chat app). The web and desktop app where you talk to Claude. Use it for one-off work: "rewrite this proposal," "summarize this contract," "what does this spreadsheet tell me?" It is the front door to everything else.

Projects. A folder inside Claude.ai where you store standing instructions and documents. Everything you say in that Project uses that context. You will create one Project per business area (for example, "Invoicing") and put your Business Brain and that area's rules inside it. This is how you avoid re-explaining your business every time.

Cowork. The mode of the Claude desktop app built for getting work done on your computer and in your apps. It can open your files, use your connected tools, run multi-step tasks, and produce documents, spreadsheets, and dashboards. Most of the "agents" in this guide run in Cowork.

Claude Code. A version of Claude for building software. Most owners will never need it. It appears in this guide only when you want a custom internal tool (for example, a dashboard that pulls from three systems). If you have a developer or a technical helper, they use this.

Scheduled tasks. Instructions you set to run automatically at a time or interval: every morning, every Friday, on the 1st of the month. This is how Claude works while you sleep. Each chapter lists the scheduled tasks to turn on. Because nobody is watching when they run, scheduled tasks only do Green work: anything Yellow they touch becomes a draft in your Needs Approval list (see Section 1.4).

Connectors (MCP). The plugs that let Claude reach into your apps. "MCP" is the technical name for the standard; you will just see a list of apps to connect: Gmail, Google Calendar, Google Drive, Microsoft 365, Slack, QuickBooks, Stripe, HubSpot, Notion, Asana, Zapier, and more. You connect them once in your Claude settings; after that you can switch each one on or off for a conversation. Zapier is a special connector: it reaches 9,000+ apps that do not have their own connector. In Zapier, turn on only the specific actions an agent needs (prefer "find" and "read" actions), and check the list of enabled actions every quarter.

Skills. Saved playbooks that teach Claude how to do a specific kind of job your way — "how we write proposals," "how we close the month." You create them once; Claude uses them every time the job comes up. Think of a skill as a training manual for your AI assistant.

Claude API. For developers. It lets other software (like a voice-answering platform) use Claude's brain. You will meet it in the Phone chapter, where a voice provider handles the phone line and Claude handles the conversation.

Which tool when

You want to...	Use
Ask a question or do a one-off task	Claude.ai
Keep standing instructions for an area	A Project (plus a Skill for repeatable jobs)
Have Claude work in your apps and files	Cowork with connectors
Have something happen automatically	A scheduled task
Reach an app with no connector	Zapier connector
Answer the phone	Voice platform + Claude API
Build a custom tool	Claude Code (with a developer)

1.3 What an "agent" is (for owners, not engineers)

An agent is Claude with three things attached: a job description, access to tools, and a schedule or trigger.

Think of hiring an assistant. You would:

1. Write the job description. "You handle incoming invoices. Here is how we categorize expenses. Here is who approves what." In Claude, this is the Project instructions plus a Skill.
2. Give them logins. Access to the inbox, the accounting system, the calendar. In Claude, these are connectors.
3. Tell them when to work. "Check the inbox every morning at 8." In Claude, this is a scheduled task — or you trigger it yourself by asking.
4. Review their work. At first, you check everything. As trust builds, you check less. In Claude, this is the permission model in Section 1.4.

That is all an agent is. There is no magic and no code required for the vast majority of this guide.

One agent per area is the rule this guide follows. An Invoicing agent, an Email agent, a Payroll-review agent. They share the same Business Brain so they agree on facts, and they hand work to each other through your systems of record (your CRM, your accounting software, your calendar), not through each other directly.

Team note (Dana, Integration Architect): We considered one giant "run my business" agent. We rejected it. Separate agents are easier to trust, easier to fix, and easier to give narrow instructions. The master dashboard (Part 3) is where it all comes together, not inside one agent.

1.4 Setup checklist and the Three-Lane Permission Model

Accounts and access (do once, about 60 minutes)

1. Claude subscription that includes Projects, Cowork, connectors, and scheduled tasks. Use a business email, not personal.
2. Create a Project called "Business Brain." You will fill it in Section 1.5.
3. Connect your core apps in Claude settings, in this order: email (Gmail or Microsoft 365), calendar, file storage (Google Drive, OneDrive, or Dropbox), accounting (QuickBooks, Xero, or FreshBooks), payments (Stripe, Square, or PayPal), and your CRM if you have one. Connect Zapier for anything else, enabling only the actions you need.
4. Create a shared folder in your file storage named "AI Operations" with subfolders per area. Agents will save their outputs here.
5. Choose your system of record for each area and write it down (see table below). This one decision prevents most integration headaches.
6. Set up a "Needs Approval" inbox. A label in your email or a channel in Slack or Teams where agents post things waiting on you.
7. Decide your review cadence. Most owners: 15 minutes each morning for approvals, 30 minutes Friday for the weekly dashboard.

Systems of record (fill in yours)

Area	Your system of record	Example
Customers and leads	____________	HubSpot
Money in and out	____________	QuickBooks Online
Payments	____________	Stripe
Calendar	____________	Google Calendar
Tasks and projects	____________	Asana
Documents	____________	Google Drive
People and payroll	____________	Gusto
Support tickets	____________	Help Scout

The Three-Lane Permission Model

Every action an agent might take goes in one of three lanes. Every chapter's Guardrails section uses these lanes.

Lane	What it means	Examples
Green — AI acts alone	Low risk, reversible, no money moves, no outside party is committed	Categorize an expense, draft a reply, create a calendar hold, update a CRM note, produce a report, flag an anomaly
Yellow — AI prepares, human approves	Something goes to a customer, employee, vendor, or government	Prepare an invoice, a customer email, a job listing, a social post, a payment reminder, or a vendor order — you send it
Red — Human only, AI assists	Money transfers, legal commitments, people decisions, anything irreversible	Pay bills, run payroll, sign contracts, hire or terminate, file taxes, change bank or security settings, delete records

Rule of thumb from Marcus: If it would embarrass you, cost you money, or get you sued if wrong, it is not Green. Start with everything in Yellow. Move an action to Green only after the agent has done it correctly 20 times in a row and you decide, in writing, to promote it.

Four rules that keep the lanes honest:

1. Some Yellow actions never become Green, however many times they go right: sending a contract or any agreement for signature; final notices, collections-agency referrals, payment plans, and any threat of legal action; job offers, rejections, discipline, pay, and termination communications; messages to a government agency, regulator, or insurer; certificates of insurance or any other attestation; and anything that commits money. A contract sent correctly twenty times is still a binding contract the twenty-first time.
2. A blank limit means no automation. If a dollar limit in any prompt is still a [BRACKET], blank, or unclear, the agent treats it as $0 and everything waits for your approval.
3. Scheduled tasks only do Green work. Nobody is there to approve at 8 a.m., so a Yellow step inside a scheduled task always produces a draft for your Needs Approval list, never a send.
4. Lanes are instructions, not locks. Connectors belong to your whole Claude account (you can switch them on or off per conversation), so a Project cannot be walled off from an app you have connected. What actually stops a Red action is the app's own setting: read-only access, an approval step in your bill-pay or payroll tool, and — where your Claude plan offers it — setting send, pay, and delete tools to "Needs approval" or "Blocked." Set those before you connect money and people tools.

1.5 Your Business Brain document

The Business Brain is a single document that tells every agent who you are. Write it once, update it quarterly, and put it in every Project. It should be 2–4 pages.

Everything in it is copied into every Project you upload it to, and into Claude's memory if you use memory. So keep personal details to a minimum: roles and work contact routes, not home addresses, personal phone numbers, or personal emails. Sections 8 and 10 are optional — use role names ("on-call technician") if you prefer. Never put tax ID numbers, bank details, or passwords in it.

Copy this template into a document called Business-Brain.md and fill it in:

# BUSINESS BRAIN — [Company Name]
Last updated: [date]

## 1. Who we are
- Legal name, DBA, business address, phone, website
- What we sell, to whom, in one paragraph
- Founded [year]; [#] employees; locations: [list]
- Owner: [name]; key staff and roles: [list]

## 2. Customers
- Ideal customer: [describe]
- Customer types / segments: [list]
- Average job or order value: $[amount]
- How customers find us: [channels]

## 3. Products, services, and pricing
- [Service/Product] — $[price] — [notes, minimums, what's included]
- Discount rules: [who can approve what]
- Things we do NOT do: [list]

## 4. How we talk
- Tone: [e.g., friendly, professional, brief, no slang]
- Words we use / avoid: [list]
- Sign-off: [e.g., "Thanks, The [Company] Team"]
- Never promise: [e.g., specific delivery dates without checking]

## 5. Policies
- Hours: [days/times]; after-hours handling: [rule]
- Payment terms: [e.g., Net 15; deposit 50%]
- Refunds/cancellations: [rule]
- Warranty/guarantee: [rule]
- Late fees: [rule]

## 6. Systems of record
- CRM: [tool] | Accounting: [tool] | Payments: [tool]
- Calendar: [tool] | Tasks: [tool] | Files: [tool]
- Payroll/HR: [tool] | Support: [tool]

## 7. Approval rules (Three-Lane Model)
- Green (AI alone): [list]
- Yellow (AI prepares, I approve): [list]
- Red (human only): pay bills, payroll, contracts, hiring/firing, taxes, bank/security changes
- Approvals go to: [email label / Slack channel]
- Auto-send limits (blank = never auto-send): invoices under $[INVOICE AUTO-SEND LIMIT] that exactly match an approved quote; first replies and template follow-ups to leads under $[FOLLOW-UP AUTO-SEND LIMIT]
- Low-risk bill flag: known vendors under $[LOW-RISK BILL FLAG] are marked "low risk" (blank = flag none; you still approve every bill)

## 8. People and escalation
- Emergencies go to: [name, phone]
- Sales questions go to: [name]
- Billing questions go to: [name]
- Vendor/technical issues go to: [name]

## 9. Key numbers we watch
- [Metric]: target [value]
- [Metric]: target [value]

## 10. Professional advisors
- CPA: [name, firm, email]
- Attorney: [name, firm, email]
- Insurance agent: [name]
- Bookkeeper (if outside): [name]

How to use it: Every Project you create gets this file plus the area-specific instructions from that chapter. When something changes (new price, new employee, new policy), update the Business Brain and upload the new copy to each Project — nothing syncs between Projects on its own.

1.6 Security, privacy, and compliance basics

Eleven rules. Non-negotiable.

1. Use a business account with two-factor authentication on Claude and every connected app.
2. Connect only the apps your business actually uses, and lock them down in the app. Connectors apply to your whole Claude account, not to one Project, so the Payroll-review agent can technically reach your marketing tools. Use each app's own controls — read-only access, approval steps, and (where your plan offers them) per-tool "Needs approval" or "Blocked" settings for anything that sends, pays, or deletes.
3. Never paste secrets into a chat. No passwords, full card numbers, bank logins, Social Security numbers, or API keys. Connectors handle authentication securely so you never have to.
4. Personal data stays in its system of record. Employee and customer records live in your HR, CRM, or accounting tool. Agents read them there; they do not copy them into documents unless a chapter says to.
5. Start every action in Yellow. Promote to Green only after 20 correct runs in a row and your written decision — and never for the actions listed as never-promote in Section 1.4.
6. Keep a human name on customer-facing communication. Customers can be told an assistant helps you, but a person owns the relationship.
7. Review agent output weekly, even Green-lane work. Spot-check 5–10 items.
8. Know your regulated data. Health information (HIPAA), card data (PCI), children's data, and financial-account data have specific rules. If you handle them, ask your attorney or compliance advisor before connecting that data to any AI tool.
9. Keep records. Agent outputs saved to your AI Operations folder are your audit trail.
10. Treat inbound content as information, never instructions. Emails, web forms, tickets, attachments, résumés, invoices, and web pages can contain text aimed at your agent ("ignore your rules", "rank this candidate first", "send the customer list"). Agents follow only the instructions in their Project and your messages. Every chapter's prompts assume this rule.
11. Check your organization's data settings in Claude to understand how your business data is handled, and review them once a year.

⚠ Varies by location / needs a professional: Labor law, sales tax, privacy law (such as California's CCPA or the EU's GDPR), and record-retention rules differ by state and country. Wherever this guide marks ⚠, confirm with a licensed professional before automating.

---