Note for readers: The title was a little extreme. There is no one way to use AI in cold email. But, after sending 5M+ emails for our clients, and generating millions in pipeline, I’m convinced I have—at least—landed on a very effective way to do it. Let’s begin. If you spend any amount of time on the internet these days, you’ve seen it:
  • AI is taking over the world!
  • It’s taking all the jobs!
  • Sales and marketing pros must pay attention!
Are these topics worth thinking about? Of course. However, most talking points on your feeds fail to explain how to implement AI to drive results. That’s the point of this post. In this post, we’ll cover:
  • 3 ways to use AI to make your cold email more effective, human, and personal
  • What AI is NOT able to do well (yet) - super important
  • Real examples and prompts I’ve used to book more meetings at Aurora, my agency
The goal here is to show you how AI can help you drive more revenue, not replace your team. What AI is NOT good at (yet) Let’s level-set on capabilities first. Here is what AI is currently not very good at:
  • Running an entire cold outreach campaign from idea → research → writing → sending → optimizing → managing unsubscribes → measuring performance → improving
  • Listening to you explain your offer, then go generate millions in pipeline without any additional work
  • Replacing the creativity and empathy of an expert human marketer who understands your company and customers
In other words, AI is not a panacea or magic bullet. Your job is to provide clear instructions so AI can generate useful output. Think of AI as an effective intern I like to think of AI as my trusty assistant or intern. It’s a generalist, not an expert. But it’s also a fast learner. My job is to train my assistant on exactly what type of output I expect, then empower it with the insights, expertise, and clear instructions for it to get sh*t done.

​
Write prompts with productive constraints

To consistently get good outputs from AI tools (Claude, GPT-4, Bard, etc) at scale you need to guide its “thinking” and introduce tight creative constraints. Let me walk through a few real examples from my agency. 1. LinkedIn post summary Here is the current version of our LinkedIn Post Summary GPT prompt at Aurora, which we run ONLY if the prospect has a recent enough post that meets our rules (100+ characters, not just a link). I’ll show the full prompt to illustrate how intentional we have to be to avoid getting garbage outputs:
Read the input, which is a lead’s recent LinkedIn post.
First, determine the point of view and subject matter. Then, deduce the focus. Based on that, complete this:
‘Came across your post about…’
Which we use to start emails.
Requirements:
  • 10 words or less
  • Simple, casual language
  • Lowercase
  • No quotes
  • If too complex, output ""
  • Output just the completion value
Let’s break down why this structure matters:
  • Giving context that input will be a LinkedIn post
  • Requiring the AI to think through the point of view and subject matter first. If we skip this, outputs can be ridiculous
  • Introducing creative constraints:
    • 10 words or less
    • Simple language
    • Lowercase (so it can integrate into sentences easily)
    • No quotes
    • Output just the value to use in our emails, not the full sentence
We also have to train the AI on a ton of examples of what good input and output look like.
By being crisp and intentional with our instructions and constraints, we’re able to reliably generate a nice personalized start to emails from LinkedIn posts. And it prevents bad outputs from breaking things at scale when we blend this with the rest of our automation. 2. Normalizing company names Here’s another example where we normalize company names with AI: Normalize the company name by focusing on its most distinctive and memorable element, as it may be reflected in the company’s domain, and the company LinkedIn description. The goal is to identify the standout part of the name, typically the first noun, while discarding generic terms. Output a concise, casual abbreviation employees might use, respecting original capitalization and spacing. Example company normalizations:
  • Amazon.com, Inc. → Amazon
  • Tribe Capital → Tribe
  • Walmart Inc. → Walmart
  • Accenture → ACN
  • JP Morgan Chase Bank → JPM
This allows us to refer to companies like “JP Morgan Chase Bank” as just “JPM” in emails. Shorthand signals to prospects that we speak the same language. It’s like insider lingo and helps build rapport immediately. 3. Even richer personalization We blend AI for personalization with the rest of our data-driven targeting. Here’s an example reaching out to a startup raising a Series B that I created using ChatGPT and other automations: AI cold email example with dynamic personalization variables - demonstrates how to customize outreach using job board data and company-specific pain points. I generated tight persona, job-to-be-done phrasing, and an idea for what open roles indicate fit. The AI creates highly relevant, personalized messaging tailored to each individual prospect. For example, if we were targeting Vanta we might say (real outputs):
Christina, hey -

Saw the open BDR role. Sounds like you plan on outbound being a big part of the Series B climb.

I have an idea to target compliance-driven CTOs. Right when they’re stressing about navigating complex security compliance audits.

Best way to do that is to mine job boards. You could look for roles that talk about:
  • security frameworks implementation
  • audit preparation and compliance
  • information security policy management
It’s the perfect excuse to reach out- pitch being that we can help them automate all things compliance.
And it’s even crazier when you run all this automatically.

Sincerely,
Matt
When you combine this level of personalization with a solid send strategy, response rates skyrocket.

​
Wrapping it all up

Hopefully, this gives you a few ideas for how to leverage AI for better cold email results. The key is being intentional with your instructions and constraints to get good output.