What moved the score, in the use case's own terms.

A score nobody can account for does not get acted on. Every use case gets an Explainability tab showing which factors raise and lower its predictions, and by how much, in the unit the use case is measured in.

What you can do

In the right unit

Probability points for a likelihood, the predicted value's own units for a value model, relative chance for a ranking. Not an abstract importance number you have to translate.

Per class and per label

Where a use case picks one of several classes, it shows what raises each class. Where it can apply several labels at once, it shows what drives each label.

Four views of the same model

Which factors matter and in which direction, how a factor moves the output across its range, where in the score range each factor matters, and how the picture after training compares with the latest prediction.

Quick, whatever the run's size

It reads a sample of the customers a run scored, so it costs about the same on a small run and a large one, and every panel says how many customers it read against how many were scored.

About the use case, not one customer

It explains how the model behaves across the customers a run scored. Individual prediction rows, tables and exports carry no factor columns, so this is a tool for understanding and defending a model rather than for telling one customer why.

Queryable, not just viewable

Each run also writes its explainability tables into the workflow's database, so you can dig into them in SQL or ask AI Studio about them.

Protected columns never appear

Columns an admin has excluded or tagged as personal are left out. A value shown is read from the customers' own rows, never free text, and only where your organisation stores source data.

Setup and requirements

Steps

  1. Train a use case. The tab appears when training finishes.
  2. Read what drives the score, and switch class or label where there is one.
  3. Run a prediction. The tab refreshes and compares the two.
  4. Query the explainability tables where you want to go further.

What you need

  • A trained use case
  • Source-data storage on, for factor values to be shown

Two limits worth knowing. These factors are read from the model, not from the world: they say what moved the model's score, which is not the same as what causes the outcome, and every panel says so. And they explain the use case, not the individual — there is no per-customer list of reasons on a prediction row, table or export today.

Build an Agentic Worker for your team.

Tell us which workflow you want to hand off.

Book a demo