How to build your ICP from your customer list with AI
Learn how to use AI to analyze won and lost deals, spot patterns in your best customers, and turn those patterns into LinkedIn search filters you can run today.
Get startedShort answer
To define your ICP with AI, give a model a structured list of won and lost deals and ask it to compare the two groups on industry, company size and job title. Then map each trait it finds onto a LinkedIn search filter and test the filter on a small batch before you scale outreach.
Gather your raw deal data
AI can't find patterns in vague descriptions. Give it a structured list of who paid you and who didn't, and it works from evidence instead of guessing.
Pull 10 to 20 won deals and 10 to 20 lost deals from your CRM or a spreadsheet. If you're a founder without a customer list yet, use the best 5 to 10 accounts from a previous role, your beta users, or people who replied to your last launch. The contrast between the two groups is where the useful signal sits.
One row per deal, and keep the column names boring and consistent. The model reads your headers before it reads your values, so a column called Industry does more work than a column called Notes.
Company name
- Data point
- Lets the AI check the public profile and confirm the industry
- Why it matters
- Acme Logistics
Industry
- Data point
- Shows which niche actually converts
- Why it matters
- Third party logistics
Employee count
- Data point
- Sets the company size filter later
- Why it matters
- 51 to 200
Decision maker title
- Data point
- Names the persona you search for
- Why it matters
- Director of Operations
Outcome reason
- Data point
- Shows the trigger or the deal breaker
- Why it matters
- Needed automated reporting
Strip emails, phone numbers and anything a customer would object to seeing in a prompt before you upload. Company name, industry, headcount, title and outcome are enough.
Run the pattern analysis prompt
Most people ask an AI to define an ICP and get back adjectives: innovative, fast-growing, values-driven. That output reads like a horoscope. The fix is to make the comparison explicit and to ban generalities inside the prompt itself.
Paste your table or upload the CSV, then send this.
Run it twice. On the second pass, paste the answer back and ask which three traits appear in at least half your won deals and in none of your lost ones. That forces the model to count instead of summarize.
Paste this job
I am uploading a list of won and lost deals. Compare the two groups and find the characteristics that appear in the won deals and are missing from the lost deals. Cover three things only: industry niche, company size range, and job titles. Do not give me general advice or adjectives. For each trait, tell me how many won deals it appears in and how many lost deals it appears in.
Read the output and cut the fluff
The first draft will still contain things you can't act on. Anything you can't turn into a filter, a search term or a sentence in a message gets deleted.
Keep three to five traits. If the model hands you twelve, no single search will match all of them, and you'll spend the week arguing with filters instead of talking to buyers.
Delete this
- Companies that value innovation and move fast
- Decision makers who care about ROI
- Businesses going through digital transformation
Keep this
- 3PL and freight forwarding companies, 50 to 200 employees
- Director of Operations or Head of Logistics, not IT
- Recently opened a second warehouse or added a new lane
Turn the profile into LinkedIn search filters
A description isn't a search. Map each trait onto a filter LinkedIn actually has, and accept that some traits won't map cleanly. Where a trait has no filter, move it into the keyword field or into your opening line instead.
Be aggressive with exclusions. If the analysis showed that companies under 10 employees always churn, set your minimum size to 11. That keeps your pending invites pointed at people who can buy.
- 1
Map the industry
Take the niche the AI named and pick the closest LinkedIn industry filter. If Sustainable Packaging isn't an option, use Packaging and Containers and put sustainability in the keyword field.
- 2
Set the size bracket
Use the employee counts from your won deals to pick the range. If your wins cluster at 50 to 200, search 51 to 200 and check the first page of results by eye.
- 3
Enter the decision maker titles
Put the exact titles from the won column into the title filter. Add the two or three variants your customers actually use, since Director of Operations and Head of Ops are different strings.
- 4
Exclude what lost
Use the boolean NOT operator on titles and industries that show up in the lost column. Excluding IT and procurement from an ops-led sale saves more time than adding another include.
Write the reason for outreach
The ICP tells you who to message. The pattern behind it tells you what to put in the first line, and that first line is what gets a reply.
If the analysis found that wins cluster around operations leads at freight companies who just added a warehouse, that's your opening. Describe their situation. The product can wait until they confirm the situation is real.
Example
- You
Hi Dana, saw Northwind opened a second warehouse in March. Most ops leads I talk to at that point are holding inbound freight together with three spreadsheets. Is that where you are?
- Prospect
Pretty much. Two sites, three sheets, and nobody trusts the numbers.
- Suggested reply
That's the exact mess we fix. Want the two page version of how Acme handled it when they added their second site?
Test the profile on a small batch
Send 20 to 30 connection requests against the new filters before you commit a whole week to them. Then read the profiles of the people who accept and the ones who ignore you.
You're checking one thing: does this filter combination return people who look like your won deals? If half the results are agencies when your wins are freight forwarders, the industry filter is too broad. Fix it before you add volume, because a bad filter at scale just fills your inbox with the wrong conversations.
Write the filter combination down somewhere you'll find it again. In a month you won't remember why the headcount floor is 11, and you'll second guess a decision that was based on evidence.
If you can't describe your ICP in one sentence to a colleague, it isn't finished. Keep tightening until you can.
Doing this with Kinlead
Kinlead takes the profile you just built and finds people on LinkedIn who match it, or matches against your website, then scores each lead from 0 to 100 against that profile. Leads below the fit line aren't kept, so the list you work from is already narrowed. Lead search runs on LinkedIn's standard search, so a Sales Navigator seat isn't required.
If you'd rather catch people while they're paying attention to a competitor, the Steal Customers agent watches competitor company pages and optional founder or employee profiles, reads the comments on their recent posts, and turns commenters who look like buyers into leads. Each lead keeps the post URL, what the post was about and what they commented.
- 1
Import or find leads
Search for people matching your ICP, import a CSV of LinkedIn profile URLs, or let Steal Customers build the list from competitor post comments.
- 2
Set your limits
Choose daily invite and message limits and a send window in the lead's time zone. Each LinkedIn account sends at most one action every five minutes.
- 3
Write the opening note
Connection requests can carry an AI-written note under 200 characters that opens with why that person became a lead, which is the pattern your ICP analysis surfaced.
- 4
Choose how replies are handled
Hand the conversation to yourself at the first reply, let the AI reply until there's real interest, or let it book a meeting with your scheduling link. Replies land in one inbox, classified by intent.
Pro costs $49 a month after a 7-day free trial. A card starts the trial, nothing is charged that day, and you can cancel any time in Settings.
Frequently asked questions
Related
Get started with Kinlead
Prefer markdown? How to build your ICP from your customer list with AI.md