[object Object]

Most “AI in CRM” pitches are wrappers. Zia Prediction Builder is a real classifier that trains on your historical records and produces a calibrated probability per row. It’s also one of the most under-used features in the platform.

When to Use It

Right fit:

  • You have at least 12 months of closed deals (won and lost).
  • You have at least 200 records in each outcome class.
  • The signal lives in fields the rep fills in (not free-text notes).

Wrong fit:

  • New product where your historical data is from a different motion.
  • Outcomes that depend on external context (economy, regulatory) more than CRM data.

Data Prep Beats Model Tuning

Spend more time on the inputs than the model:

  • Drop fields that leak the answer (a “Lost Reason” field tells the model nothing useful — it only exists after you know the outcome).
  • Drop fields with >40% null rate.
  • Standardize picklists (already covered if you’ve done schema discipline).

The model is mostly automatic. The inputs are 90% of the quality.

Training and Evaluation

Prediction Builder splits your data 80/20 train/test and reports accuracy, precision, recall, and F1. Don’t ship a model with F1 below 0.65 — it’ll predict no better than a coin flip on the cases where you most need it.

Accuracy:  0.81
Precision: 0.74
Recall:    0.69
F1:        0.71
-> Acceptable to ship; monitor monthly

Re-train every 90 days. Models drift as your sales motion evolves.

Surface It Where Reps Actually Look

Don’t dump the prediction in a side panel. Put it on the Canvas top strip with a clear meaning:

  • 80%+ -> green “Likely Win”
  • 50-79% -> amber “Coachable”
  • <50% -> red “At Risk”

Wire a Sales Signal when a deal drops a tier (“This deal moved from green to amber today”). That’s where prediction becomes action.

Avoid the “Black Box” Pushback

Reps will ask why the model says 60%. Use Zia’s “Reason for Prediction” to expose the top three factors. If you can’t explain why, reps won’t trust the score and will route around it.

Don’t Use It for Forecast

A prediction is not a forecast. The model tells you per-deal probability. The forecast aggregates across the team with stage-based categories. Mixing them produces “AI says we’ll hit $4.2M” reports that fall apart at the first board meeting.

What to Do This Week

  1. Confirm you have enough historical data for both outcome classes.
  2. Build a candidate model in Prediction Builder; check F1.
  3. Surface the score in the Canvas top strip with tiered colors.
  4. Add a “tier change” Sales Signal so reps see movement in real time.
[object Object]
Share