Ideal customer profile (ICP)

How an ideal customer profile works

An ICP turns a fuzzy idea of a good customer into rules a team can apply to a list. Salesforce describes it as a detailed description of a buyer who is a perfect fit for a product, most often used in B2B 1. Qualtrics describes it as a usually fictitious company with the budget, scale, location and needs to be the dream lead 2.

The input is the current customer base. Start with the customers that bring the most value, then look for what they have in common 1. The usual attributes are these.

• Industry and the products or services the company sells.
• Location and any regional limits on where the seller can operate.
• Employee count, revenue and budget.
• The technology stack the company already runs.
• The pain points the product solves.

Most of these are firmographics, and the stack is technographics. Salesforce lists pain points as a core part of the profile and calls an ICP that ignores them counterproductive 1.

The output is a short written profile plus a set of rules. Some rules are filters. Headcount between 50 and 500 in software in one country maps straight onto search parameters 3. Other rules are judgments. Whether a company has an in-house data team, runs a channel sales motion or sells into regulated healthcare lives in a company description and a careers page, not in a field 3. Those rules need enrichment and a reader.

An ICP feeds most of the go-to-market stack. It sets the account list for account-based marketing, the fit half of lead scoring and the filters for list building. It also says who to leave out, which saves money on every downstream step.

Filter rules vs judgment rules

Applying an ICP the same way across thousands of accounts is harder than writing it. Filter rules are the easy part because they map to fields. RichAPI's people_search, for example, accepts employee_size_start and employee_size_end next to industry and location filters 4. A request with those values returns only people at matching companies.

Judgment rules break at scale. A person can read three browser tabs and decide in ten seconds whether a company sells into regulated healthcare 3. Repeated across a large list, that reading time adds up and the answers drift from one reader to the next. Teams then either drop the rule, which loosens the ICP, or hand it to a model.

A model can apply a judgment rule if it gets facts first. RichAPI's ICP playbook pairs two calls for this 3. The first call, enrich_company, takes a LinkedIn company page URL and returns industry, headcount, specialties and description at 1 credit per call 3. The second call, ai_enrich, runs a scoring prompt over those fields and returns structured JSON at 2 credits per call 3.

The playbook insists on a third answer. Each rule returns yes, no or unknown, because a rule that can only say yes or no tends to say yes when it has nothing to go on 3. Unknown should route to a person rather than to the no pile 3. Calibrate on 100 accounts the team already has an opinion about before running the prompt on the whole list 3.

Rules that a filter can express should stay filters. The playbook says a rule such as headcount between 50 and 500 in software belongs in people_search, with no AI step at all 3. Save the model for the rules that need reading.

Example of an ICP in practice

A company sells claims software to hospitals and clinics. Its best customers are healthcare providers and health-tech vendors with in-house integration work, and the buyer is usually an operations or IT lead. The written ICP reads like this.

Industry: Hospital and health care, health-tech software
Headcount: 50 to 500
Region: United States and United Kingdom
Must have: In-house integration work (HL7, EHR, claims)
Exclude: Agencies, resellers and staffing firms

The first three lines are filters. The fourth is a judgment rule, and the fifth is an exclude list. The team pulls 1,000 candidate accounts with the filters, then runs the two-call check from the playbook on each one 3. At 1 credit for enrich_company and 2 credits for ai_enrich, the check costs 3,000 credits for the full list 3. The prompt asks one question with three possible answers, at temperature 0 with a JSON output schema, so code can branch on the verdict 3.

Suppose 240 accounts come back yes 3. The playbook uses the same figure and notes that running email_finder at 5 credits per call on all 1,000 accounts would have cost more than the filter 3. Contact search and email finding then run on the accounts that passed. The unknowns go to a person for a quick read, and their answers become test cases for the next version of the prompt.

If most accounts pass, the rule is too loose to earn its cost. The playbook says to skip the AI step when a rule passes about 90% of accounts 3.

ICP vs buyer persona

In Qualtrics' framing, the ICP describes the company that the buyer persona works for 2. Teams need both, in that order. The ICP picks the accounts, and the persona shapes who to contact inside them and what to say.