Ideal customer profile: build it from evidence, not AI guesses

Ask a chatbot to describe your ideal customer and you will get an answer in seconds. It will have a job title, a company size, a list of pains and perhaps a name and a favourite podcast. It will read well. Almost none of it will come from your customers.

That is the problem this article is about. An ideal customer profile (ICP) describes the kinds of organisations where you win, deliver well and keep the customer. You build it from your own records: deals you won and lost, customers who stayed and left, accounts that were hard to serve, and what customers said in their own words. AI can help you sort and compare that evidence. It cannot supply it. Anything your records do not support stays marked as unknown.

ICP, persona, buying committee and intent are different things

These terms get used interchangeably, and the confusion causes real targeting mistakes. They answer different questions.

An ICP says nothing about whether a particular account is ready to buy today. A company can fit your profile perfectly and have no reason to change for three years. Keep fit and timing separate, or your pipeline will fill with good-fit accounts that are not in the market.

Gather a balanced set of evidence

The instinct is to look at your best customers and describe what they have in common. That produces a flattering portrait and a misleading one. If your best customers are all mid-sized, but so are your lost deals and your churned accounts, size is not what makes a customer good. You only learn that by looking at both sides.

A useful evidence set includes:

• Won and retained accounts: the ones you would happily sign again.
• Won but churned accounts: where the sale worked and the fit did not.
• Lost deals: especially those lost to “no decision” rather than a competitor.
• Hard-to-serve accounts: customers who stayed but cost far more effort than expected.
• Customer language: interview notes, call transcripts, support tickets and reviews, in the customer’s own words.
• Commercial constraints: what you can deliver, where you can sell and which deal sizes make sense for you.

You will not have perfect data. That is fine. Record what you have, note where it came from and be honest about how thin it is.

Separate who they are from why they buy

Firmographic details (industry, size, region, technology) are easy to collect and easy to over-trust. They describe the account, not the reason it needed you. For each good customer, try to answer four further questions:

1. The problem: what was going wrong before they bought?
2. The trigger: what changed that made them act then, not a year earlier?
3. The value: what did they get that they could not get from their alternative?
4. The ability to succeed: what did they have in place that let them use the product well? A team, a process, clean data, an executive sponsor.

The fourth question is the one most profiles miss, and it often explains churn. A company can have the problem, the trigger and the budget, and still fail with your product because nobody owns it after launch.

Where AI helps, and where it starts inventing

AI is useful once you have records. Give it anonymised deal notes, interview summaries or a spreadsheet of accounts and ask it to:

• group accounts by stated problem and trigger
• list which traits appear in won, lost and churned accounts
• find counterexamples: accounts that have a trait but did not succeed
• quote the source record behind every claim, using a record ID
• list the questions the evidence cannot answer

The last two instructions matter most. Require a record ID beside every statement, so a person can check it. Ask for counterexamples explicitly, because a summary naturally keeps the common pattern and drops the exceptions, and the exceptions are often where the insight is.

What AI should not do is fill the gaps. If you ask a model to role-play your customer, it will produce fluent, plausible answers built from general patterns in its training data. That can be a useful way to prepare interview questions. It is not research about your customers, and it should never enter the profile as evidence. For the detailed interview-analysis method, see how to run voice-of-customer analysis with AI.

A worked example with a small, labelled sample

This example is illustrative. The company, accounts and counts are invented to show the method; twelve records demonstrate a process, not a statistically representative result.

A fictional company sells reporting software to operations teams. It reviews twelve accounts from the last two years: six won (four still customers, two churned), four lost and two it now considers poor fit. An AI step organises the anonymised notes into candidate traits with record IDs. A person then fills in the rest of the table.

Two plausible traits get rejected. Company size looked sensible, and the AI’s first summary listed it as a defining feature because half the wins had it. The contrary column shows half the losses had it too. “Fast-growing” was worse: no record mentioned growth. The model inferred it from hiring pages, which is a guess about a guess.

The trait that matters most was not in anyone’s original description of the ideal customer: a named person who owns reporting. It showed up only because churned accounts were in the sample.

The provisional ICP then reads like this:

Operations-led organisations that report from several disconnected systems and have a named person who owns reporting. A recent change of operations or finance leadership may be a trigger (low confidence). Poor fit: teams with no reporting owner, whatever their size. Unknown: whether industry matters; we have too few records outside logistics to say.

It is short, it states its confidence, and it says what would change it. That is more useful than a page of adjectives.

Test the profile before you build on it

Take the draft to the people who talk to customers. Sales will tell you whether the profile matches the conversations that go well. Customer success will tell you which accounts struggle after launch. Ask each of them for an account that fits the profile but went badly, and one that does not fit but went well. Those cases tell you where the profile is wrong.

If you have little data, perhaps because the company is early or has just entered a new market, do not wait for more. Write the profile with low confidence everywhere, list the questions, and design the next few customer conversations to answer them. A profile that says “we do not know yet” is honest and useful. One that fills the gaps with assumptions is neither.

When to revisit it

Review the ICP when the evidence changes, not on a calendar alone: a new product or market, a run of churn in one segment, a change in what your best deals look like. Keep the evidence table as the working document and the short profile as its summary, so the next review starts from records rather than memory.

A clear ICP is the starting point for the next decisions. Positioning depends on knowing who you are for (B2B positioning: turn customer evidence into a clear message), and so does a useful content plan (B2B content strategy). If you already score individual leads, how to evaluate AI lead scoring covers the next level down.

Blank ICP evidence table

Copy this into a spreadsheet. Add rows for poor-fit traits and unknowns as well as positive traits.