New: Draft, publish, and roll back changes to your models

Changes to classification and extraction models no longer go live the moment you make them. Every edit is now saved in a draft, and nothing reaches your analytics or automations until you publish it. A full version history lets you look back at every published version and return to one if you need to. Model drafts are available now.

What is new

• Drafts for every change. Renaming a label, rewriting a description, or adding an entity all land in a draft. Your published model keeps processing live messages the whole time.
• Publish when you are ready. Select Publish to put your changes live in one step. A colored dot on the button shows whether there are unpublished changes.
• Compare before you commit. When a model has a draft, the model accuracy shows the draft's figure next to the published one, so you can check that a change is an improvement before it goes live.
• Estimate the new volumes. On a classification model, start a volume estimate from the Monthly volume column to see how your changes shift the share of messages each label gets.
• Version history. Open History from the model's menu to see every version of the model and revert to an earlier one.

Why it matters

Editing a live model used to mean your changes hit analytics and automations straight away, which made it risky to work on a model during business hours. With drafts, you can experiment freely, check the effect on accuracy and volumes, and publish only when you are confident. If a change does not work out, you can discard it or roll back to an earlier version.

Getting started

Open any classification or extraction model and make a change. You will see a Draft badge next to the model name and an amber dot on the Publish button.

For more on how models work, see What is a classification model? and What is a model draft?.