Build an ICP Scorer With AI (From Your Closed-Won Data)

Your CRM already has the answer. Every closed-won deal you have ever logged is a data point describing the kind of company that buys from you. Most sales teams ignore this and build their ICP from gut feel, founder instinct, or a whiteboard session that happened three years ago.

This post shows you a better path: pull the pattern out of your closed-won data, encode it into a scoring rubric, hand the rubric to an LLM, and use the output to cut a bloated prospect list down to the accounts that are actually worth a rep's time.

---

Why Closed-Won Data Beats Intuition

When you ask a rep to describe your best customer, they will describe their favorite customer. When you ask a founder, they will describe who they built the product for. Neither answer is necessarily wrong, but neither is grounded in evidence.

Your closed-won list is grounded in evidence. It shows you:

• Which company sizes actually converted, not which ones you hoped would
• Which industries closed fastest and at the highest rates
• Which firmographic combinations appeared repeatedly in your best deals
• What the accounts that churned early had in common (if you cross-reference churn)

The goal is to surface that pattern systematically and make it repeatable.

---

Step 1: Pull and Clean Your Closed-Won List

Export every closed-won deal from the last 12 to 24 months. For each account, collect as many of these fields as you have available:

• Industry or vertical
• Employee headcount at close
• Annual revenue (estimated or actual)
• Geography or region
• Business model (SaaS, services, marketplace, etc.)
• Deal size (ACV or one-time)
• Sales cycle length in days
• Lead source
• Number of stakeholders involved
• Any technographic data you have (tools they use, stack signals)

You do not need all of these. Ten to fifteen data points per account is enough to find patterns. Clean the data so each row represents one deal and every column is consistently formatted.

---

Step 2: Find the Repeating Patterns

You are looking for clusters, not outliers. Open the export in a spreadsheet and sort by deal size descending. Then look for what the top third of deals have in common.

Ask yourself:

• What headcount range appears most in the top deals?
• Which two or three industries show up repeatedly?
• What is the median sales cycle for the fastest-closing accounts?
• Is there a lead source that consistently produces higher ACV?

Write down the three to five attributes that appear most reliably in your best deals. These become the backbone of your ICP scoring rubric.

Do the same exercise for your worst deals, meaning the ones with the longest cycles, the smallest ACV, or the accounts that churned within six months. The negative pattern is just as useful as the positive one.

---

Step 3: Build the Scoring Rubric

A scoring rubric is a simple table that maps attribute values to point scores. The LLM will use this table to evaluate any prospect account and return a total score.

Here is the format:

| Attribute         | Criteria                        | Points |
|-------------------|---------------------------------|--------|
| Headcount         | 50-500 employees                | 3      |
| Headcount         | 501-2000 employees              | 2      |
| Headcount         | Under 50 or over 2000           | 0      |
| Industry          | [Your top 2 industries]         | 3      |
| Industry          | [Adjacent industries]           | 1      |
| Industry          | [Poor-fit industries]           | 0      |
| ACV potential     | Above $[your median ACV]        | 3      |
| ACV potential     | 50-100% of median ACV           | 2      |
| ACV potential     | Below 50% of median ACV         | 0      |
| Sales cycle fit   | Typically closes in [X] days    | 2      |
| Lead source       | [Your highest-converting source]| 2      |
| Technographic fit | Uses [relevant tool/stack]      | 2      |
| Geography         | [Your best-converting region]   | 1      |

Fill in the brackets with your actual data from Step 2. A perfect score on this rubric would be 16. You might decide that any account scoring 10 or above is Tier 1, 6 to 9 is Tier 2, and below 6 is deprioritized.

The exact thresholds do not matter as much as the consistency. The rubric forces everyone on the team to evaluate accounts the same way.

---

Step 4: Apply the Rubric With an LLM

Once your rubric is finalized, turn it into a system prompt for an LLM like GPT-4o or Claude. The prompt structure looks like this:

You are an ICP scorer for [Company Name]. Use the rubric below to evaluate
prospect accounts. For each account I provide, return: a score out of 16,
a tier (Tier 1 / Tier 2 / Deprioritize), and a one-sentence rationale.

RUBRIC:
[paste your rubric table here]

Respond in this format:
Account: [name]
Score: [X/16]
Tier: [tier]
Rationale: [one sentence]

Then feed it your prospect list, either one account at a time or in batches if your tool supports it. You can run this in ChatGPT, Claude, or via API if you want to automate the process at scale.

The output gives every rep a consistent starting point. Instead of debating whether an account is worth pursuing, the team looks at the score and moves on.

---

Step 5: Use the Score to Prioritize Contact, Not Just Outreach

ICP scoring is most valuable when it connects to your actual go-to-market workflow. A Tier 1 score should trigger a specific action: which rep gets the account, what sequence they enter, whether a meeting should be booked directly.

This is where scoring and routing connect. In one B2B sales case study measured across 2,420 sales meetings and 1,281 deals, the close rate gap between the best and worst rep was nearly 30 percentage points. Routing the right account to the right rep, based on deal type and rep strengths, modeled a roughly 17% uplift in close rates on its own.

A high ICP score does not guarantee a closed deal. But pairing a well-scored account list with smart rep routing means your best opportunities get in front of your best-fit reps. You can see how that combination works in practice in our case study.

---

Download the Free ICP Scorer File

We built a ready-to-use ICP scorer template you can copy and adapt. It includes the rubric table, the LLM prompt, and a simple scoring sheet you can fill out from your own closed-won data.

---

Frequently Asked Questions

How many closed-won deals do I need to build a reliable ICP rubric?

Thirty to forty deals is a reasonable minimum to start seeing patterns. Fewer than that and you risk over-indexing on noise. If you have fewer, supplement with your top five to ten lost-then-recovered accounts and any reference customers who expanded significantly.

Can I use AI to extract the ICP pattern automatically, without doing it manually?

Yes. If you have a clean export, you can paste the data directly into an LLM and ask it to identify the most common firmographic attributes among your closed-won accounts. Treat the output as a starting point and validate it against your own knowledge of the accounts. The manual review in Step 2 catches things the model might not, like knowing that a specific deal was won because of a relationship, not fit.

How often should I update my ICP scoring rubric?

Revisit it every six months or when your win rate shifts meaningfully. Markets move, your product scope changes, and the accounts that bought in your early days may not reflect who buys now. A rubric built on stale data will score the wrong accounts highly.

Does ICP scoring replace rep judgment?

No. It removes the low-value debate about which accounts are worth pursuing and gives reps more time to focus on accounts that are already pre-qualified. The score is a starting point, not a decision. Reps should still review Tier 1 accounts before committing significant time, especially for larger deals with unusual context.

---

Stop working accounts that were never going to close. Score your list, route it right, and let your reps focus where it counts. More revenue. Same pipeline.