Give your business an AI staff.

Claude Code comes with a paid Claude plan. Pro is enough to try it. If it becomes part of your workday — and if you build what's in this guide, it will — step up to Max, which removes most usage ceilings. When you're ready to roll it out to employees, that's what Team and Enterprise plans are for: central billing, admin controls, and your company's data excluded from model training by default.

Easiest path: the Claude Code desktop app for Mac or Windows — download, sign in, done. If you're comfortable with a terminal (or your AI is walking you through it), the classic install is one command. There's also claude.ai/code for running sessions in the cloud from your browser.

# install Claude Code globally (needs Node.js — nodejs.org, LTS installer)
npm install -g @anthropic-ai/claude-code

# then start it (it opens your browser to sign in the first time)
claude

Make an empty folder — call it ai-workspace — and start Claude Code inside it. This matters more than it sounds: Claude works within the folder you start it in, so the folder is your blast radius. Drop in a few documents you don't mind experimenting on and ask for something real: "Read these three proposals and tell me how my pricing language has drifted over time."

Before Claude Code runs a command, edits a file, or touches anything outside the conversation, it asks you first. That prompt is the whole safety model, so build the habit now: read what it's asking, and approve the specific thing — never blanket-approve everything to make the prompts go away. For anything big or unfamiliar, ask Claude to use plan mode: it lays out everything it intends to do, you approve the plan, then it executes.

Gather the paper trail of how you actually operate: lease markups with your comments in the margins, past proposals, the email back-and-forth around your last five deals, the contracts you fought over. Put it in one folder and have Claude read, organize, and summarize it. This is the baseline — the difference between an assistant who just showed up and one who's read every file in your office.

The pro move: don't dump your whole inbox. Have Claude score and filter it — "go through this email export and keep only the threads where I'm negotiating terms; rank them by how much back-and-forth they contain." One of my agents distilled years of attorney correspondence down to the ~750 emails that actually mattered. Signal in, noise out.

Documents show what you did; they don't explain why, or where your hard lines are. So flip the script: have Claude interview you. By text or voice, it asks pointed questions, gives you realistic scenarios — "the other side strikes your exclusivity clause, what do you do?" — and records your answers. In my experience this is the single highest-leverage hour of the whole build.

Two rules make it stick: your interview answers override anything inferred from the documents (the corpus is evidence; you are the authority), and every answer gets saved to a dated calibration file that future sessions read first. When you correct the agent later, the correction goes in the same file. That's how it gets sharper every week instead of resetting every session.

Claude Code lets you define subagents: named specialists with a narrow job, their own instructions, and their own knowledge base. Instead of one AI trying to do everything, you get a lease reviewer, a due-diligence researcher, a financial analyst — and they can fan out in parallel: while one combs the title documents, another checks the rent roll, a third pulls comps. Each is a plain-English text file describing who the specialist is and pointing at the knowledge base you built in steps 1–2. Ask Claude to write them for you.

Last step, most important: point the team at the actual deliverable. Not "tell me about this lease" — give Claude your real diligence spreadsheet, your proposal template, your LOI format, and have it complete them for your review. Fill the model, draft the markup, produce the checklist.

Make the output a fixed contract: every run produces the same set of documents — say, a ranked issue memo, a marked-up draft, and a completed spreadsheet — every single time. Fixed outputs are reviewable outputs; you can spot in thirty seconds what's off. And hold the line from the safety rules: the agent's job is finished work for your review — never a description of work it intends to do, and never anything sent out without your eyes on it.