How to define and size an ICP with an AI agent

Write filters, not a persona

A persona says “growth-minded operations leaders at scaling companies.” A filter says: third-party logistics providers with 51 to 200 employees in the United States and Canada, buyer titles of VP Operations or Director of Warehouse Operations, excluding existing customers and named competitors. The second version can be reviewed, sized, sourced, and rejected. The first can only be admired.

Actually Agentic’s discovery prompt asks the agent for exactly these portable fields: industries, employee and revenue ranges, company and people geographies, technologies, company keywords, exclusions, titles, seniorities, departments, and a rationale for each choice. Portable matters. The same filters should make sense in a statistics table, a search engine, and a reviewer’s head. [1]

Exclusions and rationale do half the work

Exclusions are where an ICP becomes honest. Existing customers, open deals, competitors, partners, segments you cannot legally or practically serve, and companies too small to buy should all be listed explicitly. An agent will not infer them reliably, and a reviewer should see them beside the positive case rather than discover them after launch.

Ask for a specific reason behind every filter. “They would benefit” is not a reason. “Their public careers pages list roles our product replaces” is. Reasons give the approver something to verify, and they make it easier to notice when the agent has stretched the definition to reach a bigger number.

Keep it narrow. A tight ICP produces a smaller estimate, but it also produces more relevant email and a cleaner test. You can widen it later with evidence; you cannot un-send a poorly targeted campaign.

Size it from public statistics, then triangulate

Start from counts someone else has already done. In the United States, the Census Bureau’s Statistics of U.S. Businesses reports the number of firms and establishments, employment, and payroll by industry and enterprise size, nationally and by state, metro area, and county. County Business Patterns reports establishments by industry and employment size. Both cover businesses with paid employees, so they miss companies with no payroll, and both are published with a lag, so check the reference year before relying on a count. [2][3]

Mind the unit. The Census Bureau defines an establishment as a single physical location, while a firm consists of one or more establishments under common ownership or control. A company with ten branches can be one firm and ten establishments, so establishment counts can overstate the number of companies you could reasonably email. [4]

In Canada, Statistics Canada publishes Canadian Business Counts twice a year: business locations with employees, by employment-size range and NAICS industry, for Canada and each province and territory. The same caution applies, because locations are not companies. [5]

Then triangulate. Compare the statistical count with a second method, such as a bottom-up count from an industry directory or association, a sample of search results extrapolated with stated assumptions, or the share of companies that visibly use a relevant technology. Multiply by the number of relevant buyers per company, not by headcount. When the methods disagree, say so and widen the range.

Report a range with a confidence label

A single number invites false precision. Ask the agent for a low, midpoint, and high estimate, the kind of source it relied on, a confidence label, and a basis that states what was counted and how. Actually Agentic’s prompt requires exactly that and tells the agent to label estimates as estimates. [1]

A worked example keeps the method visible. Suppose a public table shows about 4,000 firms in the target industry and size band across the target geographies. If a sample suggests 40 to 60 percent match the rest of the ICP, and each has one or two relevant buyers, the range runs from about 1,600 to 4,800 contacts, with a midpoint near 3,000. The numbers are illustrative; the point is that every step can be challenged.

Confidence should reflect the weakest step. An exact official count multiplied by a guessed buyer ratio is still a low-confidence estimate, however authoritative the first number looks.

Three numbers that are easy to confuse

Search terms describe an ICP; they do not enumerate one. Actually Agentic’s native discovery starts from seed companies the agent supplied and extracts only narrowly relevant business contacts published on those companies’ own sites. It works where the target market can be represented by real company domains and the relevant contacts are publicly disclosed or supplied in a reviewed list. It is deliberately not a people database. [6]

So each number shrinks the one before it. A market of 5,000 plausible contacts may yield 150 named seed companies and far fewer verified addresses. That is normal, and a quote should say so rather than price the first number as though it were the last. [1]

How Actually Agentic uses the estimate

The estimate sets the scale of a recommendation, not a promise. The landing page caps the audience at 50,000 contacts, calculates monthly capacity per mailbox from the current default planning assumptions of 20 messages per mailbox per day, 20 sending days a month, and about 2.2 touches per lead, and recommends enough mailboxes to cover the estimated audience in roughly six months. Mailboxes are grouped three to a domain by default. At those defaults one mailbox supports about 181 new leads a month, so a 5,000-contact estimate produces a recommendation of five mailboxes across two domains. [1]

The customer can dial the scale down before checkout. The quoted figure is monthly discovery and verification capacity, not a claim that the sources contain that many candidates or that the addresses will verify, and it carries no guarantee of inbox placement, replies, or meetings. The service is prelaunch and limited to organizations and recipients in the United States and Canada. [1][7]

For the arithmetic behind mailbox and domain counts in more detail, see the companion guide on sizing sending infrastructure.