How to Find the Warm Prospects Already in Your LinkedIn Network

The first time I ran my LinkedIn network through an ICP check, it came back with two warm prospects. Two, out of 25,033 connections.

The second run found 435. Same network, same tool. The only thing I changed was how I wrote my criteria.

Those 435 people are owners, CEOs, MDs, COOs and heads of operations at the kind of company I sell to. They had been in my network the whole time. I never looked, because looking properly would have cost me almost 70 hours.

Here's how the process works, what went wrong the first time, and what each step really costs.

Why is your LinkedIn network a pipeline you haven't looked at?

If you've been active on LinkedIn for a few years, you probably have thousands of connections. Some came from posts, some from events, some clicked "connect" for reasons nobody remembers. Somewhere in that pile are buyers. You just don't know which ones.

The manual way to find out is simple and miserable. Open a profile, read the title, check the company, decide. At 10 seconds a profile, my 25,033 connections would have taken almost 70 hours. So I never did it. I suspect you haven't either.

That means the list stays unsorted, and the pipeline inside it stays invisible. Meanwhile most of us put our outreach effort into cold lists full of people who've never heard of us.

A connection who fits your ICP isn't a deal. But it's a better starting point than a stranger. They accepted your request once. Some of them read your posts. That's reason enough to sort what you already have before you go looking elsewhere.

What does an AI ICP check on 25,033 connections look like?

The idea is to give a model one narrow question per person: does this person fit my ICP? Nothing else. No research, no summary, no essay.

I used Jev by TypeSafe, which I was testing at the time, called through OpenRouter. For each connection it saw only four things:

• Job title
• Company
• Headline, where available
• Whether the person engaged with my posts

It never saw a name. Names, profile links and emails stayed on my own computer.

My criteria were short:

• Role: owner, CEO, MD, COO or head of operations
• Location: company in or near the Netherlands
• Size: likely 50+ employees
• Plus: engaging with my posts

The whole batch finished in 71 seconds, for under a dollar in model costs. Every connection landed in one of four buckets:

Roughly two out of every three people I'm connected to will never buy what I sell. Useful to see in black and white, and it stings a little. Years of networking, mostly pointed in the wrong direction.

But 435 warm prospects plus nearly 2,000 maybes is real pipeline. It just needed sorting.

Why did my first run find only 2 warm prospects?

Because I told it to.

In the first version of my criteria, engaging with my posts was a requirement. It sounded smart. Someone who reacts to my content is warmer than someone who doesn't. But only 205 of my connections had engaged with my posts in the last 90 days. By making engagement a must-have, I threw out everyone else before the model even looked at their role.

The second mistake was quieter. LinkedIn's export doesn't include company size. The model has to infer it from the job title, company name and headline, the way a person skimming the list would. In my first setup, "I can't tell how big this company is" counted as a no. So anyone at a company the model couldn't size dropped out.

For the second run I changed two things:

1. Engagement became a plus, not a requirement.
2. Unknown company size counted as unknown, not as a no.

That took me from 2 to 435.

The lesson is boring and important. An AI filter does exactly what you write. If your criteria are stricter than your real buying signals, the output looks clean and is wrong. Two results out of 25,000 was an obvious red flag. A smaller miss would have been much harder to spot.

How do you write ICP criteria an AI can apply?

Writing criteria for a model is different from writing them for a sales deck. These are the rules I'd follow now.

Split must-haves from signals

Decide what really disqualifies someone. For me, that's role and rough company fit. Everything else, like engagement, is a signal that makes a borderline fit warmer. Keep the must-have list as short as you can.

Only ask for what the data can show

The export gives you a position and a company. Location and company size aren't in it, so the model is guessing them. That's fine, as long as your criteria admit it. "Likely 50+ employees" is honest. A tight employee range asks for precision the data doesn't have.

Make unknown its own answer

When a field is missing, the model needs to know what to do. Say it explicitly: unknown is not a no. If you don't, a missing field easily turns into a rejection. In my first setup, that's exactly what happened.

Give doubt somewhere to go

Four buckets beat two. With only yes or no, every doubt becomes a no and disappears. A "maybe warm" bucket keeps those people visible. Write one sentence per bucket that describes who belongs there, so a half-fit lands in the right place.

Why ask one question per connection instead of batching?

With 25,000 records, the obvious shortcut is to put many connections in one request. Fewer calls, less overhead.

I tested that. Batching was about four times faster and half the price. But only 78 to 84 percent of the batched answers matched the answers from one request per connection. Roughly one in five people got a different verdict.

When a full run costs around a dollar, saving fifty cents isn't worth that. A wrong "warm" means you reach out to someone who isn't a fit. A wrong "not relevant" is worse. A buyer quietly drops off your list and you never find out.

So: one question, one person, one call. You run hundreds of them in parallel, which is how 25,033 connections fit in 71 seconds.

How do you run this on your own network?

I put the tool I used on GitHub as the LinkedIn ICP Sorter. Everything runs locally on your machine, so names, connection data and results never leave it.

Here are the steps, with the honest cost of each one.

Step 5 is the one people skip. Don't. Start with the newest 100, read what comes back, fix your criteria, then run everything. My first setup would most likely have returned zero warm prospects on a 100-record test. That's a clear signal, and it costs nothing to fix at that point.

If a terminal isn't your thing, give the README to a colleague who is at home in one. It explains each step.

What should you do with the warm list?

A sorted list isn't pipeline yet. It's a shortlist.

Review the results before you message a single person. The model is guessing location and size, and it only sees a few fields. Read the warm list yourself. For each person, ask whether you'd actually want a conversation with them. Remove the ones that slipped through.

Then look at the maybes. At 1,982 people, that bucket is too big to read in one sitting, but it's where your criteria are least sure. Skim a sample. If you keep finding people you'd want to talk to, your criteria are still too strict. Adjust them and run again. At under a dollar per run, rerunning is cheap. Your reading time is the expensive part.

What you say to these people is a separate question, and it matters more than the sorting. But at least your outreach effort now goes to the right people instead of a guess.

What are the limits of this approach?

A few things to be clear-eyed about:

• It's a side project. I built it to test Jev. We don't use it live at Arcgent, and I share it as is.
• Size and location are inferred. The export doesn't contain them, so the model reads them from title, company and headline. Some calls will be wrong.
• Headlines and engagement aren't in the standard export. LinkedIn's own file has no headline column, and knowing who engaged with your posts takes a separate step. Without that data, those signals are simply unknown.
• Your criteria are the product. The model is fast and cheap. Knowing who your buyer really is still has to come from you.

That last point is the whole story. The model sorted 25,033 people in 71 seconds. It took me two runs to tell it who I was actually looking for.