How to write an ICP an AI agent can actually qualify against

Most ideal customer profiles are written for a pitch deck: “B2B SaaS, 10–200 employees, growth-minded teams”. A person can work with that because they fill the gaps from experience. An AI agent can’t. It has a company’s website and your words, nothing else. If your words describe things the website doesn’t show, the agent guesses, and every guess becomes an email to someone who was never going to buy.

This guide rewrites an ICP into criteria a model can check, puts the disqualifiers first, and shows how to test the result on 20 companies before it spends your budget.

Why the ICP decides your reply rate

Relevance is the main reason people ignore cold email. In Gong’s research, 87% of buyers say the sales emails they get don’t address a relevant challenge (Gong). In Hunter’s 2026 survey of decision-makers, 61% named lack of relevance as the reason they don’t reply (Hunter). Better copy can’t fix that. Better targeting can.

The effect also shows up later in the funnel. In a 2019 TOPO survey of more than 150 account-based practitioners, organizations with a strong ICP reported 68% higher account win rates (TOPO, via SalesTechStar). It’s a self-reported survey, but the direction matches what most founders learn the hard way. Lenny Rachitsky’s interviews with early-stage founders found that most got their first ICP wrong (Lenny’s Newsletter).

ICP, persona and qualification are three different jobs

• The ICP describes the company. HubSpot defines it as the “perfect company for your product or service” (HubSpot).
• The buyer persona describes the person inside that company you write to, and how to talk to them.
• Qualification decides whether one specific company meets the ICP. It’s the step an agent can do for you.

Sales frameworks like BANT, MEDDIC (created at PTC in 1996, per MEDDICC) and CHAMP are built for deal stages and need a conversation. Nobody publishes their budget or decision process on their website. Before the first email, you can observe need and, roughly, size. Everything else waits for a reply.

Write criteria the agent can observe

Clay defines a buying signal as “an observable business event that changes a company’s odds of buying”, with a source you can point to (Clay). Use the same test for every line of your ICP: could someone point to a page that proves it? If not, rewrite it or drop it.

Use more than one signal. Common Room’s advice is that the strongest insights come from stacking signals (Common Room). Every founder in Lenny’s interviews ended up with at least three attributes. Three to five observable criteria is usually enough. Write the problem signal as a gap you can see: “help content is a single FAQ page”, not “needs a help center”. That gap is also what your first email will open on.

Write the disqualifiers first

Fit criteria tell the agent what to look for. Disqualifiers stop it from talking itself into a yes. “ICP fit… eliminates accounts that can never convert,” as Common Room puts it (Common Room). The two errors don’t cost the same. A company wrongly rejected is one row you never see. A company wrongly qualified is a real email to a real person, a likely ignore and maybe a spam report, which hurts your deliverability.

Good disqualifiers are just as concrete as the fit criteria:

• Business types that never buy: agencies, consumer apps, marketplaces, competitors.
• Already solved: “runs a full help center with dozens of articles”. Say what “solved” looks like on their site.
• Out of range: a team page with hundreds of people, or no paid plan at all.
• Dead or not real: no updates in a year, a personal side project, a parked domain.

End with a rule for missing evidence: if the site doesn’t show enough to decide, don’t qualify. Without it, a model fills gaps with plausible guesses, and hallucination in a qualification step is an email to the wrong company.

A template you can copy

A good prospect for [product]:
- [What they sell and to whom — checkable on their homepage]
- [The sign they have the problem — a page, a gap, a tool you can see]
- [The sign they can pay — public paid plans, a price floor]
- [Size or stage — from the team page, careers page or pricing tiers]

Not a fit:
- [Business types that never buy: agencies, consumer apps, competitors…]
- [Companies that already solved it — and what that looks like on their site]
- [Anything too big, too small or out of region]

If the site doesn't show enough to decide, say so and don't qualify.

Here is the same template filled in for Helpgrove, a fictional help-center tool for small SaaS teams. Every line points at something on a prospect’s own website.

A good prospect for Helpgrove:
- A B2B SaaS with a public pricing page — they have paying customers who need answers
- Help content is missing or thin: a single FAQ page, a Notion or Google Doc, or docs that are just a few pages
- A small team (roughly 2–50 people) without a dedicated support or docs person

Not a fit:
- Agencies, consultancies and services businesses
- Open-source or developer tools whose docs live on GitHub
- Consumer apps
- Anyone already running a full help center (Intercom, Zendesk, HelpScout Docs) with dozens of articles

Test it on 20 companies before you scale

Clay’s advice for research agents holds here too: limit each agent to one task, “name the pages it should check”, and test on a handful of rows before you run at scale (Clay). For qualification, the loop looks like this:

Use 15 to 25 companies per batch, enough to see both fits and misses. Good reasoning names its evidence (“help is one /faq page with 9 questions”); vague reasoning means the agent couldn’t find what you asked for. A false positive usually points to a missing disqualifier. A false negative usually means a criterion is too literal, or asks for something the site can’t show.

Expect to keep refining. PostHog’s rule of thumb is that five paying customers who look the same is a good sign and ten is very strong, and that getting the ICP right takes three months to a year (PostHog). The criteria you write in week one are a starting hypothesis.

Five ways criteria go wrong

1. Adjectives instead of evidence. “Fast-growing” and “sophisticated” mean nothing to a model reading a homepage.
2. Data the site doesn’t have. Revenue, funding stage and churn are rarely on a company’s own website. Leave them out unless you also supply the data.
3. Too many ANDs. Seven required criteria means almost nothing qualifies, and what does is often a misread.
4. Unstated assumptions. If you only sell in the US or only in English, say so.
5. Describing your product instead of their situation. “Would benefit from AI-powered help docs” is your pitch. “Has no searchable help content” is their situation.

Doing this in PineLead

In PineLead

Criteria in plain English, verdicts you can read

PineLead finds new companies every day and researches each one’s public website: what it sells, its pricing, its help and blog pages, and any published contact address. Your criteria are then handed to the qualifier as written. It judges only on that research. When the research is missing something a criterion depends on, it says so and leans towards not qualified.

1. Open Configure → Limits → Analysis and paste your criteria into What makes a good prospect for this project?
2. Set Prospects per day low for the first run (10 to 15), and turn on Analyse prospects automatically.

1

2

3

3. Open Prospects. Every company has a status and a score from 0 to 100. Hover the score to read the reasoning, or filter to Rejected to check the disqualifiers.

1

2

3

4. For the full trail on one company, press the history icon on its row. You’ll see what research found, the verdict, and the draft it led to.

1

2

3

A qualified company with a deliverable address becomes a lead and gets a draft. A qualified company with no address stays under Qualified for you to pick up by hand. To push one company through right away, use ⚡ on its row. For a quick test batch, use Limits → Usage → Run now. Research costs 10 credits and a verdict 2 (pricing), so a 20-company test costs about $2.40. The scoring scale is calibrated: 85 and above only when every criterion has evidence, below 40 for a clear miss.