How I Use AI at Every Stage of Cold Email (And Where Most People Get It Wrong)

AI is not a magic wand. I keep saying this because people keep treating it like one, buying tools, generating emails, blasting them out, and wondering why nothing books. The workflow I'm going to walk you through has generated close to $2 million in closed revenue for our clients this year alone. The difference between that result and zero is structure: knowing exactly where in the campaign process to apply AI, and giving it the context it needs to do its job well.

Here's how I break it down.

Start With Deep Market Research (Most People Skip This)

Before I pull a single lead or open a single tool, I spend serious time on research. This is the step that separates campaigns that book meetings from campaigns that get marked as spam.

I use AI to build a complete picture of whoever I'm running campaigns for. That means their core offering, pricing, deal size, recent company news, leadership changes. But it also means going deep on their prospects: the buyer personas, the job functions, what those people are dealing with day-to-day, what keeps them up at night. Then I go wider: competitive landscape, market trends, regulatory changes, tariffs if we're in manufacturing. Anything macro or micro that could influence why a prospect would want to change their current solution right now.

The more specific the prompt, the better the output. I'm not dropping in a vague two-sentence brief. I'm feeding AI a detailed, structured prompt that covers every dimension of the business and the market. That research becomes the foundation everything else is built on.

One thing I'm particularly focused on during this stage: finding hidden intent signals. Not the intent data Zoom Info sells. That stuff is fine, but it's available to everyone and it's often stale. I want the signals that actually scream "I'm ready to buy right now" but aren't sitting on a platform. That could be someone posting on LinkedIn about how frustrated they are with their current email marketing platform. If my client sells a competing product, I can monitor those posts, catch them in real time, and reach out with something genuinely relevant. That's a very different conversation than a cold blast.

Source Data That's Actually Targeted

Once the research is done, I move to building the lead list. Targeted means something specific here: if we're reaching out to people who have no reason to hear from us, they mark us as spam, our sender reputation tanks, and deliverability dies. So the list has to be right before anything else matters.

Three tools I'm using right now:

Ocean.io for lookalike companies. You plug in your best customer's website URL, and it uses AI to scrape the web, analyze keywords and descriptions, and surface companies that are mirror images of that customer. You can do the same thing at the prospect level. It's a fast way to find high-fit accounts you wouldn't have found manually.

Exa AI is newer, and it's genuinely impressive. It runs AI agents that can build highly specific, niche lead lists that would otherwise take a full week to compile manually. To give you a concrete example: I was building a list for a client that sells equipment leasing solutions. Their best customers are churches, specifically ones that are expanding to new locations and would need new hardware but don't want to buy it outright. Trying to find actively expanding churches through Google is a nightmare. I gave Exa a single prompt: "I want a list of churches expanding to new locations or going through major renovations in 2025." It returned verified matches, confirmed the expansion criteria, identified the decision makers, and pulled their emails. That's a list that would have taken days, done in about an hour.

Clay's AI agent (Claygent) for finding contacts at target companies. Apollo and similar tools have a real problem: even with clean filters, roughly 10% of the contacts they return aren't actually working at those companies anymore. Claygent has access to ChatGPT, Perplexity, and Claude. I give it a detailed prompt specifying exactly what job titles I'm looking for and where to look, and it returns results with a confidence score. It's meaningfully more accurate than legacy list-building tools.

Write Copy That Speaks One-to-One at Scale

Now I have a targeted list and deep research. The next step is feeding all of that into a copywriting AI agent inside Clay.

The key here is what I'm feeding it. I'm not just handing it a list of names and asking for personalized emails. I'm giving it macro context (what's happening globally in this industry), micro context (what's happening at this specific company: new leadership, recent funding, hiring activity), and the full research output from step one. The AI copywriter has real ammunition.

Then I give it a specific structure to follow: open with a pain-engaging question that pokes the bear, casually introduce the company using the research to make the pitch feel relevant, include social proof tied to a client that closely resembles the prospect, and close with a soft, peer-like call to action.

The emails that come out of this process are long by some standards. But a long email that's genuinely relevant to the person reading it will outperform a short vague one every time. When the email references tariffs squeezing margins for a manufacturing company, or mentions a specific challenge that company is publicly navigating, it doesn't read like a blast. It reads like someone did their homework.

📥 Best Email Warmup Tools

The 6 warmup tools that work — ranked by an agency managing 20,000+ inboxes.

Get it here →

Close the Loop With AI Inbox Management

Once campaigns are running in tools like Smartlead or Bison, responses start coming in. You can handle those manually, and we do sometimes. But we've also built AI agents inside Claude that act as inbox managers.

Each agent is loaded with full context about the client we're responding on behalf of. It generates replies, and if a response isn't something I'd actually send, I correct it in Claude. Over time, the agent gets better. At the scale we're operating, this is how we respond to prospects within 10 minutes of their reply, across more than a million emails sent. Manual response at that volume isn't realistic. AI inbox management is what makes it work.

Key Takeaways

• AI needs structure and context to produce good results. Vague prompts produce vague outputs.

• Deep market research isn't optional. It's the foundation that makes every downstream step better.

• Hidden intent signals (social posts, expansion announcements, trigger events) outperform purchased intent data.

• Ocean.io finds lookalike companies; Exa AI builds niche lists fast; Claygent finds contacts more accurately than Apollo.

• Feed AI your research before asking it to write copy. The more context it has, the more relevant the output.

• Long, targeted emails beat short, generic ones. Relevance is what drives replies.

• AI inbox managers inside Claude can handle responses at scale and improve over time with correction.

Frequently Asked Questions

Why does market research matter so much before building a campaign? Because everything downstream, your targeting, your copy, your angles, depends on understanding what actually motivates your prospects to change. Without that research, you're guessing. AI can only work with what you give it, and if you give it shallow context, you get shallow emails.

What makes Exa AI different from just using Apollo or a standard list-building tool? Exa uses AI agents to scrape and verify highly specific criteria that standard databases don't track. If you need a list of churches expanding to new locations, or companies that recently changed leadership in a specific industry, Exa can build that. Apollo can't. The tradeoff is that Exa is best for niche, hard-to-build lists rather than broad volume pulls.

Are longer cold emails actually working right now? Yes, when they're genuinely relevant. The length concern comes from generic emails that ramble. When an email references specific things happening at a prospect's company or in their industry, length signals effort and relevance. That's the opposite of spam.

How do you handle AI inbox management without sending bad replies to prospects? You correct the agent when it gets it wrong. Inside Claude, if a generated response isn't one I'd send, I flag it and adjust. Over time the agent calibrates to your voice and your client's context. It's not perfect out of the gate, but it improves fast, and at scale it's the only way to respond quickly enough to matter.