Learn / AI and automation
Building an AI cold email campaign, start to launch
Everybody is talking about Claude right now and almost nobody is using it right. They prompt and reprompt all day until talking to the model takes longer than the work would have. So this is the opposite: one campaign built in order, from a blank page to a loaded sequencer, in a market I knew nothing about going in.
Six steps: persona, pain, offer, list, copy, load. The one that decides whether the other five matter is step two, reading where the market already complains, in their own words. Watch the build below, or read the method underneath it.
The order is the method
Most people start at the copy, because copy is the fun part, then work backwards and invent a persona to justify the email they already wrote. Here every step is the input for the next one.
| Step | What the model does | What you still decide |
|---|---|---|
| 1. Persona map | Returns the roles and company types worth emailing | Which segments you can reach at volume |
| 2. Pain mining | Scrapes where those roles complain, maps quotes to personas | Which complaints your offer can honestly answer |
| 3. Offer | Returns front-end variations built on the mined complaints | Which one you can deliver without going broke |
| 4. List | Builds the company universe and the contact layer, exclusions included | Whether the companies are the ones you meant |
| 5. Copy | Applies proven frameworks to every segment and angle | The angles worth a test slot, and the CTA friction |
| 6. Load | Creates the campaign, leads, sequence, schedule and settings | Whether to press launch |
Step 1: the persona, before a single email exists
Finding companies to target is the easy half. The half people skip is who inside those companies owns the problem you solve. Run ad creative for e-commerce brands and the person you need is the CMO. Send the identical offer to the CFO and it gets ignored, not because the offer is weak, but because none of it is that person's job.
So drop the target company's website into Claude and ask, in plain language, who the relevant personas are at the companies you want to reach. For this build I picked a market I did not know: an AEO platform, answer engine optimization, which is SEO for LLMs. Specifically their agency partnership offer, since cold email is usually a better partnership tool than an individual sales tool.
It came back with two groups. Agencies already selling SEO or AEO, where this is an easy add-on. And agencies with no search practice at all, ad shops and PR firms, where it becomes a service they can sell without owning the fulfillment. Then a second split by headcount, because job titles at a 12-person shop and a 200-person shop are not the same titles. Most outbound skips both splits, scrapes every lead it can and blasts one email at all of them. That is why those campaigns feel generic.
Step 2: mine the pain where they already complain
This is the step worth stealing even if you ignore the rest of the page. Ask a model to describe your prospect's pain points cold and you get the textbook version: the same three bullets already sitting in that prospect's inbox, written by everyone else who asked the same question. Nobody is moved by a paraphrase of themselves.
Go instead to the threads where these people vent about the problem in real time. Reddit, LinkedIn, niche forums. Their words, dated this month, which is more recent than anything the model was trained on. I run it as a repeatable job rather than a browsing session: a demand mining tool wired to Apify scrapers, pointed at the personas from step one. The full run cost about $130.
What came back was a pain-to-persona map plus 3 strategic findings:
- The dominant emotion in that market was impostor anxiety, not opportunity. These agency owners were not excited about a new revenue line. They were quietly afraid of being caught not knowing something their clients had started asking about.
- There was live skepticism about the vendor itself, not only the category. People were arguing about whether that platform's numbers could be trusted.
- They said out loud what a tool has to show them before they believe it. That is the offer, handed over on a plate, by the people you are about to email.
Finding one alone rewrites the campaign. You do not sell upside to somebody scared of being exposed, you sell cover. Finding two means the copy has to work against distrust that already exists with your name on it. And the mapping matters as much as the findings: every takeaway stays attached to the persona who said it, so each segment gets copy backed by its own evidence instead of a blended market average.
The mining is the cheap part. Skipping it is what costs you.
Step 3: the offer comes out of the mining, not out of your service
Pitch your service and ask for a quick call and you get the same silence everyone else gets. The offer has to make ignoring you feel like the mistake, which means handing over something real first: a deliverable, a free trial, a guided pilot, work you did before they asked for it.
Prompt for it with steps one and two loaded in. Four variations came back: a done-for-you artifact, an audit or diagnostic, a lead magnet or data asset, and a trial or guided pilot. I am rarely a fan of the audit, since a diagnosis from a stranger is just an opinion and it leaves the prospect with homework, but for a platform that already sells free trials the pilot writes itself. What made these usable was the positioning, which leaned on the distrust the mining surfaced. The full build for that step is in the front-end offer breakdown.
Two rules keep the output honest. Most value in the least time, with the lowest possible effort from the prospect. And it should feel like something they would normally pay for.
Step 4: build the list around the segments
Two ways to get this wrong. Go 100% intent-only and you narrow the list until you cannot reach enough of the right people to matter, which is why the agencies doing that are the ones going under. Go pure volume and you email the wrong people and damage your own deliverability. The boring answer is volume paired with relevance.
Feed the persona work into your list tool through its API and it builds the company universe for you. When a segment has to be exact, like agencies that already sell search versus agencies that do not, get the company layer from a directory instead of guessing. Clutchbreaks agencies down by the percentage of their delivery in each service category, a cleaner split than any keyword filter, and it scrapes fine. The same run pulled the platform's existing customers into an exclusion file, so the list never included anybody already paying them.
Run a sample before the full pull. Mine came back split into 5 segments: established AEO, boutique AEO, traditional SEO, traditional SEO at 51 to 200 employees, and general enterprise agencies. It also told me contact quality was high but that titles like head of AEO and head of SEO were far rarer than expected, which is the useful part. You write to the list you have, not the one you drew on a whiteboard. Then verify every address before it goes near a sequencer, because bounces are a deliverability problem and Google's sender guidelines are explicit about that. The step-by-step version is in the Claude Code list build.
Step 5: copy written from your own data
With the offer and the personas locked, the copy nearly writes itself, as long as you are not asking a model to invent what a good cold email looks like. I ran an analysis across the last 1,500,000 cold emails we had sent and pulled out the frameworks that actually surfaced as winners. That analysis became a prompt, so the model applies proven structures to this offer and these personas instead of guessing.
The output is a matrix: segments down one axis, frameworks across the other, every cell a test slot. With a new client the priority is testing fast and cutting what loses, and a matrix gives you something to cut. Which model to hand this to is in the AI copy test.
Two of the angles it returned are worth studying. The first:
{{first_name}}, if I ran one of your clients through ChatGPT, Perplexity and AI overviews and showed you exactly where they are losing ground to competitors, free, before your next client call, would that be worth a look? Takes us a week, costs you 5 minutes.
It never mentions the software, which matters, because naming the tool quietly reduces the value of the thing you are doing. It sells the deliverable, states the effort on your side, and states the cost on theirs. The second pokes the bear with a complaint straight out of the mining:
When a client asks why traffic is down, what do you show them besides "it is Google's fault"? We hand you the chart: which competitor is winning inside ChatGPT and AI overviews, and why. Want this run free on one account?
Step 6: hand the loading over
Loading a campaign by hand is at least half an hour of copy and paste, the least valuable half hour in the process, and it still has to happen. Across dozens of clients it eats a day that should have gone to strategy and closing.
So I asked Claude Code to load the test campaign into our sending tool and watched it name the campaign, import the leads, drop in the first sequence, set the schedule, set the settings to plain text with new leads prioritized, and load the sender emails. Launch stays a human decision. The other five workflows we run this way are in 6 ways we use Claude Code for cold email.
What the same order produced on the next build
I ran this sequence again on fractional executives, a market we had never sold to ourselves, and got from a blank page to a launched campaign in under 2 hours. First week: 3,200 emails to 3,200 prospects, 40 replies at a 1.24% reply rate, 13 interested responses, and 5 of those booked. The first meeting landed the same day it went out.
That reply rate is unremarkable, slightly low if anything. The number that matters is interested-to-booked, and it came from the offer, which came from the mining. Across the last year the same order has booked 3,000+ qualified calls for clients and opened more than $45 million in pipeline.
One caveat before you copy any of it: none of this builds the machine underneath. Domains, warmup, sending accounts, and the system that handles replies are still their own builds, and a campaign loaded onto bad infrastructure just fails faster. AI compresses the campaign, not the foundation.
tl;dr: persona, pain, offer, list, copy, load, in that order. The mining step is what makes the other five worth doing, and it costs about $130.
FAQ
Can AI build a cold email campaign from scratch?
It builds every piece except the judgement. In one sitting a model will map the personas, mine the pain, draft the offer, pull the list, write copy per segment, and load it into the sequencer. What it cannot decide is which segment is a real business, which offer you can afford to deliver at volume, or when to press launch.
How do you find cold email pain points with AI?
Do not ask the model what your prospects care about, because it hands you the textbook version already sitting in their inbox. Point it at Reddit and LinkedIn threads where those exact roles complain, scrape them, then map each quote to the persona who said it. Their language, dated this month, beats anything the model was trained on.
How long does it take to build a cold email campaign with AI?
Under two hours from a blank page to a launched campaign, once you have run the order a few times. The first build is slower because you are still deciding what good looks like at each step. What stays slow is everything AI does not touch: domains, warmup, and handling replies once they land.
Is AI-written cold email copy good enough to send?
Only if you feed it proven frameworks and real prospect language first. From a cold prompt you get generic copy that reads like every other AI email. Given the offer, the persona map, the mined complaints and frameworks pulled from campaigns that already worked, editing takes minutes instead of an afternoon.
What order should you build an AI cold email campaign in?
Persona, pain, offer, list, copy, load. Each step is the input for the next, which is why the order is the method. Most people start at the copy because copy is the fun part, then invent a persona to justify the email they already wrote. That campaign reads generic no matter how often you reprompt it.
If you would rather have the whole thing built and run for you, offer, lists, copy, infrastructure and someone on the replies, that is what we do: 3,000+ sales calls booked for clients in the past year. Or keep building it yourself. Every training, grouped by topic, is on the trainings hub, and the full walkthrough for this one is on YouTube.
PS - the copy prompt is not the thing that makes this work. A $130 scrape of people complaining is.
