What is ICP scoring with Decision Mode?
ICP scoring with Decision Mode evaluates prospect evidence against a defined ideal customer profile and returns calibrated probabilities for possible answers. Teams can use those results to keep, reject, or review rows before later work. The rubric, source context, thresholds, and treatment of uncertain rows remain decisions for the team.
Build a useful rubric
Start with criteria you can observe in the data, such as whether a company sells to businesses or has a stated product relevant to your offer. Enrich its website first, give the decision model the relevant context, and specify what counts as a strong match or insufficient evidence. A probability helps compare options; it is not proof that a company will buy.
Decision Mode uses typesafe-ai/jev-1.14.0 for calibrated, non-generative evaluation. Its base fee is $1.00 per 1,000 rows plus 1.25 times input token cost, approximately $1.01 per 1,000 in the reference scenario. Actual token use changes the total. The model can support three-way keep, reject, and unsure routing when the Flow is configured accordingly.
| Outcome | Example treatment | Next step |
|---|---|---|
| Keep | Strong evidence of fit | Continue to selected downstream work |
| Reject | Evidence contradicts the rubric | Skip the selected downstream work |
| Unsure | Evidence is missing or ambiguous | Hold for a human review |
Where should the checkpoint sit?
Place the evaluation after the evidence it needs and before expensive optional enrichment or personalization. In a full Flow run, configured filters can prevent rejected rows from reaching downstream steps. A test run marks the verdicts but does not stop rows at gates, so downstream test steps can still run and incur charges.
For example, 10,000 succeeded website lookups cost $60 at the default $0.006 rate. The reference Decision Mode estimate is about $10.10 for 10,000 rows. If 4,000 rows are rejected before a $0.01 personal LinkedIn lookup, the gross avoided lookup spend is $40, before accounting for the gate cost and any misses. This is an illustration, not a forecast.
How should a team validate the result?
Review a sample of clear matches, clear non-matches, and uncertain cases with the people who own the ICP. Compare the source evidence with each outcome, adjust the rubric and thresholds, then test again. Do not turn a model score into an automatic high-stakes decision or treat it as a substitute for human judgment.
Explore the Flow gallery for a sample rubric and use the pricing calculator to model different rejection rates.