Building an ICP From Your CRM Instead of a Guess

A founder sent me his ideal customer profile last month. One slide, very clean. Series A to Series B SaaS, 50 to 200 employees, North America and UK, VP Sales or CRO as the buyer.

Then I asked him to export his closed-won deals. Eleven of his fourteen customers had fewer than 40 employees. Nine of them were bootstrapped. The buyer was the founder in eight cases, not a VP Sales, because most of these companies had never hired one.

His ICP slide described the company he wanted to sell to. His CRM described the company that actually pays him. Those were two different businesses, and his outbound had been aimed at the wrong one for seven months.

To build an ICP from your CRM, export every closed-won account with its employee count, industry, deal size, sales cycle length and the reason the deal opened. Then find the attributes shared by the deals that closed fastest and the customers who stayed. Three attributes that repeat beat a twelve point persona document.

Why guessed ICPs survive so long

Nobody defends a made up ICP on purpose. It survives because nothing ever tested it.

The typical origin story: four people in a room, a whiteboard, a template downloaded from a content marketing site. Everyone contributes the customer they most enjoyed working with, or the logo they most want on the website. Then it sits in the sales deck for two years while the business drifts somewhere else.

Nobody goes back and checks it, because checking feels like admin rather than strategy. So the document ages quietly while the CRM fills up with evidence against it.

Here's what most people get wrong: they treat the ICP as a targeting decision made in advance. It isn't. It's a pattern you discover after the fact, and then aim at deliberately. You don't decide who your best customer is. You find out, and then you go get more of them.

The export that answers it

You need fewer columns than you think. Pull your closed-won accounts from the last 18 months with these fields:

• Employee count at time of purchase, not today
• Industry, in your own words rather than the CRM picklist
• Deal size and whether it expanded afterwards
• Days from first touch to signature
• Whether they're still a customer
• What was happening at the company when the deal opened

That last one is the column nobody exports, and it's the one that matters most. Firmographics tell you which companies could buy. The trigger tells you which ones were ready to. A 60 person logistics firm is a suspect. A 60 person logistics firm that just lost its only salesperson is a buyer.

That field isn't a tidy dropdown anywhere in your CRM. It's in the notes of the first discovery call, and you'll have to read them. Budget an afternoon. It's still the highest return afternoon in your quarter.

How much data you actually need

Every analytics vendor writing about this puts the threshold at 50 to 100 closed deals before segment numbers mean anything, which is fair if you're doing statistics. It's also useless to most founders reading this, who have eleven customers and need to make a targeting decision on Monday.

With a small sample you're not doing analysis, you're doing forensics. Eleven deals won't give you a confidence interval. They will tell you that eight of your buyers were founders, which is a finding you can act on tomorrow.

Segment win rates once you have the volume for it. RAIN Group's Center for Sales Research puts top performing sales organisations at roughly 62% win rates against about 40% for everyone else. Read that as directional, not settled fact: RAIN sells sales training, and nearly every published B2B win rate figure comes from a vendor with something to sell. Your own segmented win rate is the only number here you should fully trust.

Build the negative list at the same time

The half people skip: export closed-lost and churned accounts too.

Look for what they share. In our own work the pattern was blunt: every client who churned inside four months had nobody internally who owned the follow up. Right size, right sector, funded, and the meetings we booked went nowhere because no one on their side picked up the phone. No firmographic filter catches that. One question on a qualification call does: who owns the follow up, and what else is on their plate?

Write it down as a disqualification rule and enforce it. A negative ICP saves more pipeline than a positive one, because it stops you spending three months learning something your last twelve months already recorded.

Turning the pattern into targeting

A finished ICP fits in four lines and becomes filters someone can run.

Not: mid market technology companies experiencing growth challenges.

Instead: UK and Ireland, 25 to 80 employees, field service or logistics, founder led, and either hiring a salesperson right now or having lost one in the last 90 days.

The second version becomes a list: headcount band, geography, sector, hiring signal. Someone can build that in an afternoon with a data provider and a job board scrape. The first version becomes a slide.

We rebuilt targeting this way for an HVAC services campaign last year and the reply rate landed at 6%, against roughly 1 to 2% on the broader list we'd started with. Same offer, same inboxes, same sequence. The only thing that changed was who received it. Most outbound problems founders bring me are targeting problems in a copywriting costume, and the pilots that fail in the first 60 days usually fail here.

When your CRM can't answer the question

Sometimes the export comes back as noise. Two possibilities.

You have no ICP yet, because you're eight customers in and every one arrived by a different route. Run the exercise as a hypothesis generator instead: pick the segment with the shortest sales cycle and aim there for a quarter.

Or your CRM data is too poor to read, which usually looks like blank close reasons and opportunities logged after the deal was already signed. Then the ICP isn't your first problem. That's a RevOps problem, and no amount of targeting work survives contact with a CRM nobody maintains.

What to do this week

Export closed-won, closed-lost and churned for the last 18 months. Read the first call notes on your ten best accounts and note what was happening at each when the deal opened. Find the three attributes that repeat. Build one list against those attributes only, and run it against your current list for two weeks.

If you'd rather someone else ran the export and told you what it says, that's most of a free GTM audit. Bring CRM access and 45 minutes. Pricing is published if you want the numbers first.

FAQ

How do I build an ICP from my CRM?

Export closed-won accounts from the last 18 months with employee count, industry, deal size, cycle length, retention status and what triggered the deal. Find the attributes that repeat, then write the ICP as filters you can build a list from.

How many customers do I need before the analysis is worth doing?

Around 50 closed deals for statistically meaningful win rates by segment. With ten to fifteen you're looking for obvious repeats rather than significance, which still beats a whiteboard session.

What is the difference between an ICP and a buyer persona?

The ICP describes the company worth selling to: size, sector, geography, situation. The persona describes who signs. Get the ICP right first, because the wrong company with the right job title still doesn't buy.

Should I include churned customers in my ICP analysis?

Yes, as a separate output. Churned accounts define your negative ICP. If a segment converts well but leaves inside six months, it belongs on the disqualification list.

What if my best customers are too small to be profitable?

That's a pricing problem, not a targeting problem. Raise the price or narrow the offer before you abandon the one group that reliably buys from you.

How often should I rebuild the ICP?

Once a year, or after 20 new closed deals. Pricing and product changes move the pattern faster than founders expect.

Can I do this without a proper CRM?

Yes, if you have invoices and an email archive. Build a spreadsheet of every customer, when they bought, how much, and what was happening at their company. Slower to assemble, same output.

Does an AI SDR change which ICP I should target?

It changes the economics of reaching a segment, not which segment is right. Cheap outreach makes wide lists tempting, which is exactly how cost per booked meeting quietly doubles.