A6 · Evidence case · public-case-study (Clari enterprise reference, 2024)

ML pipeline forecast: trained on deal stage velocity, activity signals, stakeholder engagement, and historical close rates; produced weekly forecast with confidence intervals

Forecast accuracy: from ±22% to ±9% at 60-day horizon; deal at-risk identification 3 weeks earlier than rep reporting; 2 quarters to positive ROI — The model's value is not the forecast number — it is flagging disagreement between human and model to force earlier deal inspection. Agreement is less valuable than surfaced divergence.
moderate evidence self-reported Last updated 2026-06-18

Source trust note: Results are reported by the implementing company or a vendor's reference customer. Direction is credible; magnitude may be overstated due to selection effects.

What was built

ML pipeline forecast: trained on deal stage velocity, activity signals, stakeholder engagement, and historical close rates; produced weekly forecast with confidence intervals

Company type: Enterprise SaaS ($50M+ ARR)

Tier map

Tier 2 strategy: pipeline intelligence and planning

Tier 2 · Strategy C6 Agentic hub →

Human-in-the-loop design

CRO reviews model forecast vs rep-submitted forecast weekly; model flags deals where rep forecast deviates >30% from model prediction

Results

Forecast accuracy: from ±22% to ±9% at 60-day horizon; deal at-risk identification 3 weeks earlier than rep reporting; 2 quarters to positive ROI

Quality caveat: these results are self-reported — treat as directional signal, not precise benchmark.

Source trust

self-reported public-case-study (Clari enterprise reference, 2024)

This case is rated self-reported. The implementing company or a vendor reference customer is reporting their own results. Direction is credible — these teams built something real and measured it. Magnitude may be inflated due to selection effects (teams who had good results are more likely to share them).

Failure mode observed

Model trust erosion: when model was wrong on a high-profile deal, sales leadership discounted all model outputs for 2 quarters

Transferable lesson

Key lesson: The model's value is not the forecast number — it is flagging disagreement between human and model to force earlier deal inspection. Agreement is less valuable than surfaced divergence.

How to cite

@misc{shalvi_gtm_evidence_ai_pipeline_forecasting_crm_2026,
  author    = {Singh, Shalvi},
  title     = {ML pipeline forecast: trained on deal stage velocity, activity signals, stakehol — Agentic GTM Evidence Case},
  year      = {2026},
  note      = {Source trust: public-case-study (Clari enterprise reference, 2024). Methodology: moderate.},
  url       = {https://shalvisingh.com/gtm/evidence/ai-pipeline-forecasting-crm}
}

Singh, S. (2026). *ML pipeline forecast: trained on deal stage velocity, activity signals, stakehol — Agentic GTM Evidence Case*. GTM World Model. Retrieved from https://shalvisingh.com/gtm/evidence/ai-pipeline-forecasting-crm

← All evidence Annual Index → C6 Agentic hub →