ICP Scoring With AI: Rank Every Account

Most ICP scoring guides end at a points table. Twenty points for headcount in range, fifteen for funding stage, ten for the right CRM in the stack, threshold at 70. Build one. It takes an afternoon and it beats working the account list alphabetically. Then a rep opens the result, finds an account at 71 sitting above an account at 68, and has no way to tell which of the two is worth the morning.

This guide builds the version that answers that. Derive the criteria from accounts you already closed, enrich the whole list from providers and from your own billing and product systems, then have AI score and rank every account against each other with a written reason attached. Scores and reasons go back to the CRM, and the top band goes into a sequence.

Define your ICP scoring criteria from closed-won accounts

Criteria should come from outcomes you already know rather than from the personas doc. Pull every closed-won account from the last 12 to 18 months. Twenty accounts is enough to see a pattern. Fifty is better.

For each one, record:

• Company domain. The identifier every later step matches on.

• What they pay, and since when. Contract value or MRR plus start date. If the CRM does not hold it, your billing tool does, and it connects as a source rather than an export.

• How fast they closed. Cycle length separates accounts that fit from accounts you dragged over the line.

• Whether they stuck. A renewal or a seat expansion is the strongest fit evidence you own, and the field most ICP work leaves out because it arrives a year after the deal.

• Who signed. Title and team. This is usually where a second segment shows up.

Then run the same pull on churned accounts and on deals that died at "no decision". If wins and losses look identical on headcount and industry, those were never really your criteria, whatever the scorecard template says.

Write what you find as a sentence, not a spreadsheet. Something like: "Seed to Series A B2B SaaS, 20 to 150 employees, sells to sales or marketing teams, already runs a CRM and a billing tool, pays us above $8,000 a year." Every clause traces back to a field you can actually enrich, which is the test that matters.

If you have no account list to score yet, building a B2B lead list from a CRM export is the step before this one.

Where a points-based account scoring model runs out

We are not going to pretend weighted scorecards are useless. Ours was a Google Sheet for the first year and it was right about the obvious accounts, which is most of them. An account scoring model built from weights hurts in three specific places.

• The weights are guesses wearing a number. Nobody derived "funding stage = 15 points" from anything. It felt like about half of headcount, so it got half.

• A missing field scores like a bad field. An account with no employee count on record loses the same twenty points as one with the wrong employee count, and 400-account lists always have holes.

• The output cannot be argued with. A rep who thinks account 71 is wrong has nothing to push against. Ideal customer profile scoring only changes behaviour when the reason travels with the number.

There is a fourth thing, and it is the one that made us build differently. A scorecard evaluates each account on its own, but ranking is relative. "Is this a good fit" and "is this a better fit than the other 399" are not the same question, and only the second tells a rep what to do on Monday.

While we are here: the quarterly ICP review deck needs to die. Six people agreeing to move the headcount band from 50-200 to 80-250 with no evidence, and then nobody updates the scoring rules anyway.

What to enrich before ICP scoring: firmographics and your own revenue data

A score is only as good as what it reads. Before scoring, fill in what is missing on every account, from providers and from the systems you already run.

The bottom three are the ones no data vendor can sell you, because that data only exists inside your own stack. They connect the same way the CRM does, matched on company domain or account ID. The top two come from a waterfall of providers ordered by cost and hit rate. You do not assemble the chain. The cheapest provider likely to answer runs first, and a miss costs nothing.

At $0.02 a credit on the $20 plan, and those figures assume every account matches, which no list ever does. The run shows its total before it starts, so an expensive enrichment can be dropped while dropping it is still free. The mechanics are on the customer data enrichment page.

Run ICP scoring across every account at once

Paste the description you wrote in the first section. That is the whole configuration. AI reads every enriched account, scores each one from 0 to 100 against the description and against the other accounts, and ranks the result.

Reading the full list at once is what makes the ranking mean anything. Ask a research agent about one company and it can tell you the company has 90 employees and raised a Series A last spring. It cannot tell you the account resembles your three best customers more closely than the other 399, because that comparison needs all 400 in view.

Our own first run covered 1,200 accounts in a few minutes, and what we did not expect was how much came back in the middle. Roughly 60% landed between 40 and 65. We had been telling ourselves we had a tight ICP. What we had was a strong top decile and a long tail of plausible companies.

Read the reasons behind each ICP scoring result

Every account comes back with a score and one or two sentences explaining it. The reasons read like notes from a junior analyst who went through the list overnight: this one matches on size and category but runs no CRM at all, that one pays like a target account but has not logged in since June.

Reps skim scores and act on reasons, which is the same argument behind lead scoring that explains the score at contact level. The auditing argument is better:

• Check the boundary. Read ten reasons either side of wherever you drew the tier line. That is where the model's judgment is weakest and where you find out whether your ICP sentence said what you meant.

• Check the top. If the top twenty reasons all cite the same attribute, your description is leaning on one clause and the rest is decoration.

• Check the empties. Accounts that missed enrichment sink to the bottom, not because the account is bad but because there was nothing to read. Filter for blank firmographics before cutting anything, fix the domains by hand, and re-run enrichment on just those.

That last one is the manual hour in this guide worth spending. It is also where you go back and rewrite the criteria sentence from earlier. Nobody gets the description right first time, and re-scoring costs almost nothing once the enrichment is done.

Write ICP scores back to your CRM and build segments

A ranked list nobody opens is a nicer spreadsheet. Map the output onto CRM properties, which are created with the right field type when they do not exist yet, and kept current by the same two-way connection that brought the accounts in.

Tier is worth writing separately from score. Views, list membership, and routing rules are all easier to build on three values than on a number with a threshold buried in the filter, and a tier survives a re-score that shifts every number by two points.

From there the top band has somewhere to go. Enrol it in an email and LinkedIn sequence built from the same enriched list, so the research that scored the account also writes the opening line of the message that reaches it. Tier 3 stays saved rather than deleted, because the next re-score will move some of it up.

Then set the cadence: monthly on the accounts you are actively working, quarterly on the rest. Companies raise rounds, buy tools, and lose the champion who would have taken the call. ICP scoring that ran once in March is describing a market that no longer exists.

Bring in your own account list and score it against the criteria your closed-won deals already prove. The free plan covers 200 credits a month and credits never expire, which is enough to run a real sample end to end and see whether the reasons change who you call first. Get started free.