The agent calculates a close probability for every active opportunity using a model built from historical deal outcomes and current deal signals: stage, days in stage, contact engagement level, buying committee coverage, competitive context, deal size relative to historical wins, and activity recency. Each deal gets a model-predicted probability score alongside the stage-based probability the CRM assigns by default. The gap between the two — the delta between the rep's stage commit and the signal-based probability — is the most actionable output in the forecast.

A single-number forecast is less useful than a range built from deal signal quality. The agent produces three forecast views: commit (deals with signal-based probability above 70% and a verified close date), best case (commit deals plus deals at 50–70% that could close with specific interventions), and pipeline (all deals currently active). Each view is calculated from deal-level signal scores, not from the rep's verbal commitment. The commit number is what the data says will close — not what the rep wants to close. The difference is measurable and, over time, reduces forecast variance.

The forecast isn't just a prediction — it's a to-do list. The agent identifies which deals in the commit and best-case categories have the highest risk of slipping and what specific intervention is needed to close them in the current quarter. A deal at 62% probability that's stalled at the commercial review stage needs a different intervention than a deal at 58% that lacks an economic buyer introduction. Each flagged deal comes with a specific action: an email the rep should send, a resource to share, a stakeholder who should be looped in, or an escalation that should be triggered.

A forecasting model that isn't measured against actuals doesn't improve. The agent compares each quarter's signal-based forecast to the actual closed revenue at quarter end — by deal, by rep, and by segment. Deals that the model called at 70%+ and didn't close are analyzed for the signals that predicted the outcome (the model missed something). Deals that closed faster than expected are analyzed for what drove the acceleration. These learnings feed the next quarter's model calibration — so the forecast gets more accurate over time rather than staying at the same variance.