The slightly reckless guide to ABM outreach with AI
Your Linkedin feed is probably full of posts about how you can “generate 70 leads per week with this n8n and ChatGPT workflow” that the poster will share with you if you just comment ‘spammy clickbait’.
We promise, we don’t need you to comment, or repost, or follow us on Linkedin. We just want you to copy-paste the prompts we’ve shared below and test them out yourself.
Two overcaffeinated humans and a handful of GPT prompts can’t exactly replace a roomful of SDRs, but we’ve come close, especially since we’re also wearing a bunch of other hats on top of doing outreach ourselves. We’re big believers of dog-fooding our own playbooks at ESM, so we tested this ourselves in preparation for doing it with one of our clients, and since it showed pretty promising signs early on (see below for results), we’re sharing it so you can test it out, too.
1. Set up your GPT
The very first move is still the simplest: paste your product blurb into the System field, but don’t stop at the elevator pitch.
For ESM, we copy-pasted a detailed description of what we do, the exact pain points we solve, our main competitors and differentiators, and the way we price our services and structure our projects.
For you, this could look like a summary of your product suite descriptions straight from your product and pricing pages, an extract from your VC pitch deck, and examples of your competitors or your USPs.
However, that’s not everything. We also feed ChatGPT everything that we know is consistent about our tasks: the structure of answers, the tone of voice, the follow-ups. This is what that looks like for ESM (where we hate the trademark GPT ‘generic context’ intros.)
That additional layer of specificity means two things
- You don’t have to waste time working around GPT quirk to quickly digest info and iterate.
- Every follow-up prompt—no matter how weird—stays lasered on the specific facts it knows about your product and market. In practice, the model starts pulling your USP into metaphors that match your tone, weaving the price into ROI lines, even mirroring the micro-snark you sprinkle into investor updates.
2. Brainstorm your ICP without a whiteboard
Once the model is fluent in you, flip the spotlight to them: your Ideal Customer Profile.
Our prompt was: “First, let's draft a description of my Ideal Customer Profile that’s under 500 words. What information do you need from me in order to define my ICP?”
The real added-value is how granular the follow-up questions get.
Instead of “What’s the company size?” we get “What is the funding stage? How are they financed?” Rather than “What’s their marketing budget?” it asks “what is their average ACV? What is the range of their sales cycle?” These types of questions are both more specific, and easier to determine through a search.
Answering forces you to commit to numbers you’ve been hand-waving, for at least the length of one campaign. At the very least, it forces you to make one assumption and test it out.
When GPT finally spits out a polished ICP, we ask it to carve that big blob into micro-niches and argue which one will jump on our value prop fastest. And we’re talking really, really micro:
- “Series A logistics platforms automating scope-3 emissions tracking”
- “GCC based startups building “Spend-management” SaaS”
Just with that, you can already see how granular the outreach can get.
3. Make the machine do the prospecting
The next job is translating those niches into something the open web recognises. We design a “visible trigger” stack that you can throw at the machine. Below are just a few examples, but the possibilities are incredibly broad.
- Funding signal: press releases mentioning “Series A” in the past 24 months.
- Hiring signal: live jobs for “build repeatable pipeline” or “lead generation”. Marketing team under 2 people, etc.
- Tech stack signal: Free “Hubspot” branded landing pages
- Marketing signal: intermittent or absent posting on Linkedin pages
We hand those triggers to GPT and tell it to go fish, and come back with a table and .csv. You can play around with the models here–we got great results with o3, with and without deepsearch, so have at it. Either way, ten minutes later we have a 100-row CSV of companies complete with website, LinkedIn page, domain and a 15-word description the model generates on the fly (“Cloud TMS syncing Asian factories to GCC distributors in real time.”). That raw list becomes the canvas for account-specific copy.
We also got the suggestion to build a Boolean search string for Crunchbase, but haven’t tried that out yet so w're not even sure if the arguments are built right. Here’s what it looked like, if you’re curious–we thought it looked a little generic, but you never know until you test it, with these things.
("Series A" OR "Seed")
AND SaaS
AND ("manufacturing" OR "supply chain" OR sustainability OR fintech OR payments OR "voice AI" OR martech OR salestech OR "productivity" OR HRtech)
AND (Dubai OR "Abu Dhabi" OR UAE OR Saudi OR Qatar OR Kuwait OR Bahrain OR Oman)
AND ("Head of Sales" OR "VP Growth" OR "CRO" OR "Marketing Lead")
4. Turn a CSV into microscopic outreach
Here’s where the robot earns its keep. We loop over the CSV and feed each row back to GPT with a templated prompt. Below is an example of what you could use:
“Using the company description below, write a LinkedIn outreach note in 125 words or less for the [prospect job title] of [Company x]. Include a pain point that our product/service can solve for [Company x] and support this with one of the 4 case studies I have listed below. Finish with a CTA to explore further.”
The prompt will really depend on the fields you’ve chosen to focus on in the earlier “research” step. If you’ve collected ‘top pain point’ or ‘example of value we can deliver’, you can ask the machine to reference that. It can be longer, or shorter, be for email or Linkedin comments or Reddit threads.
We’re a marketing agency, so we built a series of outreach messages with highly specific examples of campaigns we would run for our target companies, based on a list of examples and options specific to that industry. We referenced only one of our case studies where applicable, but we got really specific about suggested marketing campaigns.
In fact, they were almost as specific as what we would pitch in a sales call, after the discovery. That’s not something your typical SDR can do.
That’s the secret sauce: the model isn’t just merging {company} and {pain} tokens. It’s constructing mini-stories unique to each logo, based on examples of value that you know you can deliver.
5. Deploy and check in
We queued the messages manually in Linkedin Sales Nav, AFTER a re-read and a couple of tweaks (trust issues, what can I say). They're helpful, especially if you go the extra mile with a mention of a recent company update or the latest post on the prospect’s Linkedin profile. However, I suspect the process could just as easily be run in Agent mode, if you choose to go with a less tailored message.
Results? Over three weeks we converted 9% of the total target list into meetings. That’s above the average that most Series A teams manage with a full-time SDR bench—and we did it in a few hours split between two people, a Linkedin premium subscription, and a language model.
6. Big fat qualifiers
Now before you give this workflow a whirl, we need to make a couple of qualifiers super clear:
First, amplification still costs money. We piped our AI-generated copy through LinkedIn Sales Navigator because that’s where our buyers live, and it’s easy to integrate with our CRM (Attio).
But Sales Nav isn’t free, nor are paid email-warm-up tools, enrichment APIs or any other channel that lets you hit prospects at scale without getting throttled. Budget at least a few hundred dollars a month for distribution; otherwise the prettiest AI sequences will sit in a Google Sheet collecting pixel dust.
Second—and this one separates grown-up campaigns from spam—you can’t outsource strategic thinking to the model. The tailoring works only because we fed GPT high-octane material: a razor-sharp USP, outcome-rich customer quotes and case-study links that have verified numbers in them.
If you haven’t invested in those raw ingredients, the machine will happily spin generic fluff and burn your domain reputation in the process.
And last: even with solid inputs, you still need a human sanity-check on every batch. Nonsensical suggestions slip through, cultural nuances get mangled, and the occasional joke lands like a brick. AI can draft at warp speed, but only a founder’s eye can guarantee the message feels handmade—and that’s the difference between a polite ignore and a booked call.
That’s it. That’s all we have for warning signs.
So grab a coffee, paste your blurb into the System prompt, and let the robots loose.