Most ICP definitions are built from intuition, last quarter's closed-won list, or whoever sales remembered to mention in a workshop. Markets shift. Buyer patterns change. And the profile you're targeting today is already stale. Logloop's Causal ML Agents build your Ideal Lead Profile from four hard signals — and subtract the ones that look like wins but aren't.
Most ICP models stop at conversion. Logloop's Causal ML Agents use four progressive signals — each one a stronger indicator of real value than the last — and apply a detractor class that removes false positives your pipeline is currently full of.
The Causal ML model assigns weights progressively — each downstream signal carries more information about long-term account value than the one before it. Detractor signals are subtracted, not ignored.
Weights are illustrative of the model's relative ordering. Actual weights are calibrated per customer from your CRM data during onboarding and update continuously as new signals arrive. The detractor penalty is applied multiplicatively against an account's positive score — a strong conversion match that also carries churn-after-use attributes will score low overall.
The ICP you built 18 months ago reflected a real set of wins — at that time, in that market. Most companies never recalibrate it. The Causal ML Agents run continuously, so the profile updates as signals accumulate rather than waiting for a quarterly planning cycle.
Most companies have the data — it's in their CRM. Sales accepted or rejected leads. Deals qualified or stalled. Customers expanded or churned. The signal is there. What's missing is a model that reads all four signals together, weights them correctly, and subtracts the detractors before the profile reaches scoring and targeting.
Standard scoring models find correlations — attributes that appear frequently in converted accounts. Causal ML goes further: it identifies which attributes actually caused the expansion outcome, not just appeared alongside it. This distinction matters when churn-after-use accounts share many surface-level attributes with your best customers.
Every account in your CRM receives a continuously updated ICP score as a contact/company property. Sales sees it. Marketing targets from it. Nurture sequences are prioritised by it. When a new signal arrives — a churn, an expansion, a rejected lead — the model updates and the scores propagate.
ICP defined in a workshop, built from memory and closed-won anecdotes
Conversion is the final signal — expansion and churn never feed back into the profile
Detractor accounts look like ideal accounts at the MQL stage — no model catches them
Sales reject leads that look "ideal" — marketing can't explain why
ICP updated annually at best — market drift goes undetected for months
Churn-after-use accounts are treated as CS failures, not ICP signals
ICP built from four progressive signals — acceptance, pipeline, conversion, expansion
Revenue expansion is the highest-weighted signal — defines who the real ICP is
Detractor class explicitly subtracts churn-after-use attributes from the ICP score
Sales acceptance and rejection patterns feed directly into the model — friction becomes data
ICP score updates continuously — market shifts surface in weeks, not quarters
Churned accounts become a training class — their attributes are removed from targeting
The ICP score isn't a report. It's a CRM property that marketing, sales, and nurture all act on — updating in real time as signals accumulate.
30 minutes to map your four signal sources, identify your detractor class, and scope the Causal ML model against your CRM data. You'll know within two weeks whether your current ICP is targeting the right accounts — or a profile that drifted.