Ask five people at the same company for the ICP and you get five adjectives: "mid-market SaaS, growth-minded, modern stack". Nothing in that sentence can be queried, swept for, or falsified - which means the actual targeting happens downstream, improvised by whoever builds each list. The fix is to treat the ICP as an instrument rather than a slide: derived from evidence, expressed as executable filters, validated on a schedule like any other model. This article is the method, and it is the front door to the whole signal-based machine - every sweep in that system starts by asking "matching what, exactly?"
The slideware problem
The traditional ICP exercise produces a persona document: a fictional VP with a name, stock-photo face and "pain points". The document has real costs beyond the workshop that produced it. It cannot drive automation - no sweep can query "growth-minded". It cannot be wrong - unfalsifiable claims never get corrected. And it silently diverges from reality - the market moves, the document does not, and eighteen months later the team is targeting a segment that stopped converting a year ago. The instrument version fixes all three by being made of observable, queryable, falsifiable attributes.
Start from closed-won evidence
The data you need already exists: your closed-won accounts, your closed-lost, your churned. The exercise is enrichment plus comparison - enrich all three populations with the observable attributes (size, industry, geography, tech stack, hiring shape, funding stage, and the signal families from the intent data guide where history allows), then compute which attributes actually separate won from lost. The results reliably surprise: the attribute everyone "knows" matters often shows no lift, while an unglamorous one - a specific tool in the stack, a hiring pattern, a size band nobody targeted deliberately - separates cleanly. This is a sandbox computation, not a workshop: run it over the full history, exactly.
Express it as executable filters
| Slideware | Instrument |
|---|---|
| "Mid-market SaaS" | B2B software, 50-500 employees, US/UK/EU |
| "Modern stack" | Runs [specific CRM] and [specific analytics]; no [disqualifying platform] |
| "Growth-minded" | Hiring in GTM roles within 90 days, or raised within 12 months |
| "Feels the pain" | Posting text mentions [pain phrases]; or intent surge on [topics] |
Every row on the right is a query some data source can answer - which is the entire point. Weight the filters by the conversion lift the evidence step measured, sum into a fit score, and the ICP becomes what it should have been all along: the scoring function at the front of the pipeline, the one the postings sweep and every other mission consumes.
The negative ICP: who to refuse
The higher-ROI half of the exercise is the one teams skip: the attributes that predict loss and churn. Bad-fit pipeline is expensive everywhere it touches - sales cycles that stall, discounts that close them anyway, churn that follows, case studies that never materialize - and it enters through targeting that had no refusal rules. The churned-account comparison usually names them plainly: a size band that never retains, an industry whose requirements you cannot meet, a stack that fights your integration. Write them as hard filters. A sweep that returns fewer, better accounts is the machine working, and the discipline pairs naturally with evidence-first outreach - both are bets on quality compounding over volume.
Quarterly revalidation
The ICP in motion: how the machine consumes it
The payoff of the instrument format is that everything downstream stops improvising. The weekly signal sweeps filter on it; account scoring weights by it; routing reads its fit score; even the AI advertising media plan's audience definitions inherit it. One definition, versioned, consumed by every mission - which also means one place to improve: when the quarterly revalidation shifts a weight, the entire machine retargets the next morning. That property - targeting as configuration rather than tribal knowledge - is quietly one of the largest gains agentic go-to-market delivers, and it is available to any team willing to replace adjectives with queries.
Frequently asked questions
What makes an ICP "executable"?
Every attribute is a query some data source can answer - size bands, geographies, named technologies, hiring patterns, signal thresholds - weighted by measured conversion lift into a fit score that sweeps and scoring consume directly.
How do you derive an ICP from data?
Enrich closed-won, closed-lost and churned accounts with observable attributes, then compute which attributes separate the populations. Weight by lift. The surprises - attributes everyone believed in showing none - are the value.
What is a negative ICP?
The refusal rules: attributes that predict loss or churn, written as hard filters. Bad-fit pipeline costs through the entire funnel, so excluding it at targeting is usually worth more than another positive segment.
How often should the ICP be revalidated?
Quarterly against the trailing period’s actual conversions, plus off-cycle when win rates drift or great customers arrive from outside the definition. An unvalidated ICP becomes fiction at the pace the market moves.
Can this work without a data team?
Yes - the lift analysis, enrichment and scoring are exactly the sandbox arithmetic an agent mission computes across your CRM history, and the executable format is what lets every later mission consume the result automatically.
Sources
- Gartner - B2B buying journey research
- Ehrenberg-Bass Institute - category buyer research behind in-market share assumptions
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.