> ## Documentation Index
> Fetch the complete documentation index at: https://docs.snowflake.com/llms.txt
> Use this file to discover all available pages before exploring further.

# %document-ai% decommission

Snowflake is decommissioning the %document-ai% UI experience and
the `<model_build_name>!PREDICT` method, and transitioning to the [AI_EXTRACT](/sql-reference/functions/ai_extract)
function for document extraction, which is a next-generation document extraction solution powered by the `arctic-extract` model. This model offers better
scalability and accuracy, faster inference, and enhanced features such as longer output token limits and the ability to extract entities, lists, and tables
in a single API call.

The %document-ai% UI and the `<model_build_name>!PREDICT` method were decommissioned on **March 16th, 2026**.
After the decommission you no longer have access to your existing model builds in the %document-ai% UI. The %document-ai% UI and the
`<model_build_name>!PREDICT` method are not accessible after the decommission date, and existing %document-ai% models that
were not migrated to the Snowflake Model Registry are not available for inference after the decommission date.

## Behavior change

### Before the change

The workflow for extracting data from documents is as follows:

1. Create a model build.

   You create model builds in a dedicated %document-ai% UI, where you also upload documents, define values for extraction, and verify the answers that the model provides.

2. Optional: Fine-tune the model.

   You can fine-tune the model in the %document-ai% UI if the results provided by the Snowflake Arctic-TILT model are not satisfactory.

3. Run inference.

   You use the `<model_build_name>!PREDICT` method and the model build created in the %document-ai% UI to extract information from documents.

You can view the %document-ai% models in the %document-ai% UI and in the Snowflake Model Registry.

### After the change

The workflow for extracting data from documents is as follows:

- You use the [AI_EXTRACT](/sql-reference/functions/ai_extract) function to define values for extraction and extract information from documents. You no longer create model builds in the %document-ai% UI.

Previously, you had to complete a three-step workflow that involved creating model builds, but now this process is simplified to a single step which is using the
[AI_EXTRACT](/sql-reference/functions/ai_extract) function.

The models that were previously created and/or fine-tuned in the %document-ai% UI can be viewed in the Snowflake Model Registry if you migrated the models (see [](#label-bcr-2156-actions-required)).

You can use the fine-tuned model for inference, but you can't fine-tune the new version of the model.

AI_EXTRACT uses token-based billing. For more information, see the [Snowflake Service Consumption Table](https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf).

<a id="label-bcr-2156-actions-required"></a>

## Actions required

To continue running inference on the existing models that you created in %document-ai%, you must complete the following steps before **March 16, 2026**:

1. Migrate your existing %document-ai% models (both published and trained) to the Snowflake Model Registry. When prompted in the %document-ai% UI, to integrate the existing
   models into the Model Registry, follow the instructions on the integration banner.
2. Update your extraction pipelines to use the [AI_EXTRACT](/sql-reference/functions/ai_extract-document-ai) function for inference of %document-ai% legacy models.

   This ensures uninterrupted inference in production. For more information on using AI_EXTRACT with legacy %document-ai% models, see [AI_EXTRACT (Document AI legacy models)](/sql-reference/functions/ai_extract-document-ai).

3. Recommended: To continue using the %document-ai% data, export all of your existing %document-ai% model builds (which contain documents, prompts, and annotations) to a target internal stage.

## Change log

<div className="colwidths-auto">

| Update                     | Date      |
| -------------------------- | --------- |
| Initial publication        | 21-Nov-25 |
| Updated decommission date. | 20-Feb-26 |

</div>

Ref: 2156