How to automate sales for a company with AI

A sales email that could have gone to anyone gets treated like one, and deleted. This guide shows how to automate sales for a company with an AI agent that does the opposite. It finds businesses worth contacting, learns one real thing about each of them, and sends a short email with a small demo made just for that business.

You can run it for your own business, or set it up for another company and charge for it. The starting point is Harvey, a free, open-source sales agent you clone from GitHub and change to fit how you sell.

How an AI agent automates sales

A sales agent does the slow, repeated parts of selling so a person can spend their time on calls. Harvey runs in a loop every 15 minutes and works through these jobs:

• Finds businesses that match who you sell to
• Reads each business's website for useful facts
• Finds the person who decides and checks their email address
• Writes a short sequence of three emails
• Waits for you to approve each email, then sends it
• Reads the replies and drafts the next message

It runs on your own computer and uses your Claude or ChatGPT subscription, so there is no separate bill for AI usage.

Why one personal email beats a template

A template sent to 500 businesses gets ignored, and enough of them will land your domain in spam. A personal email starts from something true about the reader: they have no online booking, their blog stopped two years ago, or they are paying for Google Ads that send people to a slow page.

Harvey's writing rules are strict about this. Emails stay under 75 words, start with the reader rather than with you, ask one question, and never invent facts. Add a demo built for that business and the email shows what you can do instead of only claiming it.

Step 1: Map how the company sells today

If you are setting this up for another company, start with a conversation, not a tool. Write down how they sell right now:

1. Who they sell to, and how they find those people
2. What the first message says, and who sends it
3. What happens after someone replies
4. Where deals stall or get forgotten

Then mark each step: the AI can do it, the AI drafts it and a person checks it, or a person keeps it. Finding, researching and first drafts suit the agent. Pricing, calls and closing stay with a person.

Step 2: Clone Harvey and let Claude set it up

You need Python 3.11 or newer, a Claude Pro or Max plan (or a ChatGPT plan with the Codex CLI), and an email account on a separate domain used only for sending.

git clone https://github.com/ethanplusai/harvey.git
cd harvey
claude

Then type "set up Harvey for me". The repo ships instructions that turn Claude Code into a setup guide. It checks what is missing, installs it, asks for what it needs, and trains Harvey on your product.

Step 3: Teach it what you sell and who buys

Point Harvey at the website of the business it is selling for. It reads the site and writes notes on the product, pricing and the problems it solves.

harvey train https://your-company.com

Then open harvey.yaml and fill in the ideal customer: industries, job titles, company size and location. Next, choose the facts Harvey looks for. It knows 23 of them, and it collects none until you confirm them.

harvey signals --confirm free

Pick the facts that make your offer obvious. If you build websites, a business with no website is the best lead you can find. If you sell booking software, look for businesses with no online booking.

Step 4: Find and research the businesses

Harvey estimates the cost of a search before it spends anything. The default source, OpenStreetMap, is free. Paid sources cost under a dollar per thousand businesses and cover trades that the free map misses.

harvey discover --estimate
harvey discover

For each business it then reads the homepage, sitemap and team page with plain web requests, with no AI involved. That is where the useful details come from: which agency built the site, whether they run ads, when the blog last changed. Each email address is marked verified, risky, guess or invalid, and only deliverable ones are ever sent to.

Step 5: Add a demo made for each business

Harvey does not build demos out of the box. This is the part you add, and it is what makes the email worth opening. Match the demo to the fact you found:

• No website: a one-page site with their name, services and area
• No online booking: a booking page in their colours
• A blog that stopped: three draft posts on topics their customers search for
• Missing search basics: a one-page check of what is broken and how to fix it

A coding agent can build each demo from the notes Harvey already collected. Our guide to making a website with AI covers the build and the free hosting. For search checks, see how we run automated SEO audits with background agents.

To connect it to Harvey, open Claude Code inside the repo and ask for the change in plain words:

After the scout scores a prospect, build a one-page demo for that
business from its saved facts, publish it, and save the link with the
prospect so the writer can use it.

One rule to change: prompts/writer.md allows no links in the first email and one link across the whole sequence. Decide where the demo link goes, either the first email or the second, and edit that rule to match. Keep it to one link.

Step 6: Approve every email before it goes

By default nothing sends without you. Open the dashboard and work through the outbox one email at a time.

harvey dashboard   # http://localhost:5555
harvey run

Press A to approve and R to reject. Before anything leaves, a fixed check with no AI involved blocks emails that are too long, use banned phrases or still have blank name fields. If something looks wrong, harvey sending pause stops all sending at once. Only turn off approval once you have read enough emails to trust it.

Step 7: Keep your domain out of spam

Most cold email fails because of the domain, not the words. Before your first real send:

1. Buy a separate domain for sending, such as getyourbrand.com
2. Set up SPF, DKIM and DMARC records on it
3. Warm it up for two to four weeks
4. Start at 10 to 20 emails a day per inbox and grow slowly
5. Stop and clean the list if more than about 3% bounce

Add your business address and a simple way to opt out. If anyone asks whether they are talking to an AI, the answer is yes.

Sell this setup to other companies

Once it works for you, it is a service you can offer. Most small companies know they should follow up more and never find the time. Map their sales steps, set up Harvey on their own domain, rewrite skills/email_frameworks.md and skills/objection_handling.md in their voice, and hand over the dashboard.

Charge for the setup and a monthly fee to keep the lists fresh and the demos coming. For more services you can sell with AI, read how to make money with Claude Opus 5.5. For more guides like this one, visit the tellnova blog.