Co-pilot model

The next-best-action model that picks the action per customer and learns from the outcome

What it predicts

For each customer arriving at a Co-pilot split in a workflow, which of the allowed actions has the highest expected effect on the goal of that Co-pilot: retention, reactivation, conversion or upsell. The output is a decision, not a score, and the decision changes as the model learns.

Data it needs

An action space you define (the branches of the split, for example a discount email, a personal call, a survey, or no action), a default action, a target type (the goal metric) and the customer data already in Churned: scores, segments, value, tenure, engagement and custom properties.

How it works

Two phases. In exploration the Co-pilot assigns actions at random across the audience and records the outcomes, building an unbiased picture of what works for whom. Once the minimum evaluation time has passed and the model is confident, it moves to activation and assigns each customer the action with the highest expected outcome for their profile. A permanent exploration group, typically 5 percent, keeps receiving random actions so the model can detect drift, keep learning and measure its own uplift against an unbiased baseline. The full description is on Optimise with Co-pilot.

How often it runs

Assignments are made whenever a customer reaches the split. Outcomes are collected daily and the model is retrained on them continuously.

How to read the output

The Copilots page shows, per Co-pilot, the share of customers per action, the outcome per action and per segment, and the uplift of the model's choices over the exploration group and over the default action. See Measure impact.

Where it surfaces

The Co-pilot split in the Workflow builder, the Copilots page, and the next best action on the customer page.