AI Decisioning for measurable predictions.

Train and validate models on business data, then use the predictions in connected tools, decision APIs, and Agentic Workers.

Explore models and decisions

Next 30 days

Churn probability

Queue a retention review for accounts with the highest estimated risk.

Account risk distributionIllustrative data
  • Account C 18%
  • Account B 42%
  • Account A 76%
Data inputs
Account IDs, product activity, billing events, support history, and past cancellations or non-renewals.
Prediction target
Define churn and the prediction window. Score only accounts that are active at the start of that window.
How to validate
Compare with recent inactivity rules on a later period. Review how many flagged accounts actually churn and how many churned accounts the queue misses.

More model templates

Fraud detection

Flag activities for human review based on patterns in confirmed cases.

Transaction timestamps, event characteristics, and reviewed outcomes. Track false positives and sample cases below the alert threshold.

Next best campaign

Rank eligible campaigns by expected incremental value.

Prior campaign exposures, outcomes, and costs. Compare with a holdout and apply channel consent and suppression before activation.

Next best offer

Select the best eligible incentive after cost, with the option to make no offer.

Offer history, acceptance, customer context, and eligibility. Evaluate net value against an untreated group where possible.

Product recommendations

Rank relevant items for each customer or account.

Catalogue availability and timestamped customer-item interactions. Compare ranking quality with a popular-items baseline and exclude unavailable products.

Customer segmentation

Group customers by similar behavior or value for different treatments.

Customer identifiers and agreed activity, purchase, or value attributes. Check that segments are stable and large enough to serve.

Next best channel

Choose among eligible channels based on the expected business outcome.

Past contacts and responses by channel. Apply consent and frequency caps, then compare with the existing channel policy.

Send-time optimization

Choose an approved send window for the intended response.

Historical send and response timestamps and local time zones. Compare with the usual schedule in a holdout and honor quiet hours.

What you can do

Define what to predict

Choose an entity, such as a customer, lead, or product. Define the outcome and prediction window with a model template, a predictive query, or a SQL label rule.

Train more than one model

Create separate prediction tasks for different outcomes and horizons. Churn and purchase propensity estimate likelihood, lifetime value and demand predict amounts, and next best action ranks eligible options.

Compare with a baseline

Evaluate each model on data from a later time period. Compare it to a simple rule or current process and check for changes in coverage, error, and score distributions over time.

Explain the signals

Review which factors influence the predictions and in which direction. Use those explanations to check model behavior, interpret results, and investigate surprises.

Make predictions usable

Publish scores to profiles, send eligible audiences to connected destinations, or return an action through the real-time decisions API. A worker can use a score to prioritize follow-up, while business rules determine what can actually run.

Setup and requirements

Steps

  1. Connect the tables and define stable identifiers and event timestamps.
  2. Choose the target outcome and prediction horizon.
  3. Train and evaluate the model against a baseline.
  4. Publish the output and compare decisions with observed outcomes.

What you need

  • Historical records linked to stable business identifiers
  • A defined outcome and enough examples to evaluate it
  • A destination and an owner for the resulting decisions

Build an Agentic Worker for your team.

Tell us which workflow you want to hand off.

Book a demo