How to use AI for sales prospecting: the jobs it does well, and the one it shouldn't

Sunny Ray · Aug 28, 2026 · 7 min read

Use AI for sales prospecting on the jobs that are tedious, repetitive and easy to check: researching accounts before the first email, building and cleaning lists, loading contacts into campaigns, drafting replies, and finding the interested responses that slipped through your inbox. Keep a human on the decisions that can damage you if they go wrong, starting with how much you send and from which mailboxes, and on anything a buyer reads that commits you to something.

That split matters more than which tool you pick. Most AI prospecting advice is a list of products. Below is what AI actually does in my own setup, two things it did on a single day in August, and the decisions I still won't hand to it.

The jobs AI is good at

Good AI jobs in prospecting share three traits: they're repetitive, they have a clear right answer, and a person can check the output in seconds. Researching a company. Pulling titles from a messy export. Matching a list against a CRM. Classifying replies as interested, not now, or no. Drafting a follow-up that a person approves.

Bad AI jobs are judgment calls with expensive downside. Deciding to triple send volume on a new domain. Choosing who's in your ideal customer profile. Promising a price or a timeline in a reply.

Gartner surveyed 646 B2B buyers in August and September 2025 and found that 45% had used AI during a recent purchase. Your buyers are already using it to research you. Using it to research them is table stakes.

Research before the first email

The fastest win is account research. Before a campaign goes out, AI can read a prospect's website, recent news and public posts and give you two lines: what they do, and a specific reason they might care. That's what makes a first line feel written by a person.

Keep it short. A cold email with three sentences of flattering research reads like a template. One specific line is enough.

Loading lists and campaigns

On August 10 I was on a call with a founder who runs an intent-data business. About ten days earlier he'd sent me a file of contacts, and it was sitting somewhere in my downloads folder. Normally that's a twenty-minute detour: find the file, clean it, map the columns, upload it, set up the campaign.

Instead, mid-call, I asked my agent to do it. My setup is Claude Code plus an open-source personal AI infrastructure layer that sits on my computer and routes each task to the model that's best at it. It found the file, 664 contacts, loaded them into Instantly as a draft campaign, and we reviewed it together. The campaign went live before the call ended.

Notice what happened there. The AI did the tedious part. A person reviewed the draft before anything went out. That's the pattern.

Catching the replies you missed

The same morning, a few hours earlier, I'd pointed the same setup at our Instantly account through its API. It went through the replies and surfaced about 100 that nobody had answered. Real people who'd written back and heard nothing.

Follow-ups went out to those people, and meetings got booked from replies that would otherwise have died in the inbox.

This is the most underrated use of AI in outbound. Every team running more than a handful of mailboxes loses replies. They land in the wrong inbox, they get marked read by accident, they arrive on a weekend. An AI that reads every reply and flags the ones that need a human pays for itself the first week.

Routing jobs to the right model

One thing that surprised the founder on that August call was that I don't use one AI for everything. My setup sends each job to the model that's best at it. Claude handles strategy and anything that needs judgment in the writing. Codex handles code, because it's cheaper at high volume. Grok is useful for real-time data from X. Gemini handles images.

For prospecting, that split matters more than it sounds. Researching an account, drafting a first line and classifying a reply are different jobs with different costs. A big reasoning model on every row of a 5,000-contact list is a waste of money. A small fast model deciding whether a reply is a real lead or an out-of-office is exactly the right use.

My input is almost always voice. I talk, the agent does the work, and the output comes back as a page I can scan, not a wall of text in a terminal. That's what made it practical to do mid-call: I said what I wanted, and kept talking to the founder while it ran.

You don't need to build this to get the benefit. Start with one assistant connected to one tool, and give it one tedious job you do every day.

What AI should never decide

The rule I hold to: AI can draft, sort and load. It shouldn't decide how much you send.

Sending volume is the one setting that can quietly wreck everything. New mailboxes need to warm up. On our team call on August 27 we set up a fresh batch at three mailboxes per domain, and the instruction was to ramp slowly and increase over time. A model optimizing for "more meetings" would happily push volume up on day one, and you'd find out what that costs when the domains stop delivering.

The same goes for permissions. That same day, one of our sending accounts dropped from about 50,000 contacts to about 11,000, and we still weren't sure what deleted them. The fix wasn't technical genius. It was moving people to read-only access until write access was proven safe. Give AI, and people, the least access they need.

Other things I keep with a human:

• Who's in the ideal customer profile.
• Any reply that commits to price, scope or a date.
• Replies to complex or high-value leads.
• The final yes before a new campaign launches.

A simple test for any AI job: if it goes wrong, how long until you notice and how much does it cost? A bad research summary is caught in seconds and costs nothing. A volume spike on fresh domains might not show up for days, and it can cost you the domains. Automate the first kind freely. Keep a person on the second.

A starter setup

You don't need my exact stack. A simple version that works:

1. A list source. A database for contacts, filtered to your profile.
2. A verifier. Every address checked before it's sent to.
3. A sending tool with separate, warmed domains and an API.
4. An AI assistant with access to that API, read-only at first, that summarizes new replies every morning and drafts responses for you to approve.
5. A scheduler with a booking link and a couple of qualifying questions.

Start with the reply sweep. It's low risk, it's easy to check, and it finds money you already earned.

Where the offer still matters

AI makes prospecting faster. It doesn't make a bad offer good. The campaign we launched on that August call didn't pitch anyone. It invited prospects onto my show, and the business conversation happens after the episode. When every team can send ten times the email, the thing that gets a reply is still a reason to say yes.

If you want to see where AI sits in the system we run, from list to reply to booking, see where AI sits in the system.

Frequently asked questions