Pipeline Health by signal, not by score: what an agent watches that your CRM dashboard misses

A lead score is a lagging indicator that looks like a leading one. By the time the score crosses your threshold, the moment that would have made the score move has already passed - the buyer has stalled, the champion has gone quiet, or the manager has overridden the forecast.

A signal is different. A signal is one observed event - "deal stalled in stage three for 14 days," "primary contact has not opened email in 21 days" - that is itself the trigger to act.

Why lead scores fail mid-market

Mid-market sales cycles are short enough that the half-life of useful information is days, not weeks. Scoring models smooth the signal out of existence. By the time a model rolls up enough behaviours to move the score, the AE has either already noticed and acted, or noticed and not had time to act.

The other failure mode is interpretability. When the score is high but the deal is dead, no one knows which input was wrong. When the score is low but the deal closes, no one updates the model. Lead scores accumulate distrust, not predictive power.

Three signals the Pipeline Health agent watches

DivetIQ's Pipeline Health agent (part of the CRM module, coordinated by the same orchestrator as the rest of the agent catalog) subscribes to the Kafka event bus and watches three classes of signal continuously.

1. Stage stall - a deal in any pipeline stage longer than the rolling 90-day median for that stage and segment. The agent does not wait for the AE's weekly update; it fires when the timer expires.
2. Contact decay - the primary contact's engagement (email opens, meeting accepts, in-app activity on the customer portal) drops below the threshold for their account tier. The agent flags it before the AE notices in the next call review.
3. Win-rate drift - the rolling win-rate for a segment, region or product moves more than two standard deviations from its trailing baseline. Used to recalibrate forecast coverage, not to blame an AE.

Each signal carries its lineage: the underlying events, the trailing baseline used, the confidence score, and the recommended play.

Wiring signals to plays through OpenAPI

The agent does not "send the AE a notification" and stop there. It calls the same OpenAPI tools the AE would call.

For a stage stall, the play is a next-best-action card on the AE's home screen, plus a draft outreach in the customer's preferred channel - email, Microsoft Teams, WhatsApp Business or in-app. For contact decay, the Marketing Performance agent is wired in alongside Pipeline Health to re-engage through nurture content. For win-rate drift, the Forecast agent recalibrates pipeline coverage and surfaces the variance in the next monthly business review.

Same RBAC. Same audit log. Same kill switch per agent and globally. Above the agent's configured risk threshold the action queues for human approval; below it, the action fires autonomously and is logged. The threshold is yours to set per role.

Lead scores are not bad. They are too coarse for the questions mid-market sales teams need to answer in 2026. Wire the signals directly. Trust the agent the way you trust your best AE: not to be right every time, but to make the decision visible.