Creating apps

Automation apps transform your document processing requirements into reusable, shareable solutions. The app creation process guides you through defining document types, specifying data extraction needs, implementing quality controls, and packaging everything into a deployable app.

Users with developer permissions or higher can create projects and apps.

The app creation workflow

Building an automation app follows a structured development process.

1. Working with projects — Start by creating an automation project and uploading representative documents. Projects function as blueprints that outline your document processing requirements.

2. Extracting data from documents — Define document classes, or types, and specify the data fields you want to extract. Configure different extraction methods for text, tables, lists, and complex reasoning tasks.

3. Cleaning results — Transform extracted data into the format you need through cleaning operations like reformatting dates, standardizing text case, or applying custom processing functions.

4. Validating documents — Implement validation rules to ensure data quality and flag questionable results for human review. Set confidence thresholds and create custom validation logic.

5. Creating or versioning apps — Package your completed project into a reusable app that can be shared, deployed, and managed through version control.

Getting started approaches

There are several ways to get started with automation apps, from pre-built apps for common document processing needs to fully customized apps.

Customize an existing app — Start with a customization-enabled app similar to your use case. Optionally modify the underlying project to suit your specific needs.

Use existing schemas — Accelerate development by copying classes and fields from other projects in your organization, then customize as needed.

Create a blank project — Create a blank project and build your document processing logic step by step with your own documents and requirements.

Agent mode and legacy mode

Processing modes control how files are transformed into actionable insights. The mode you use determines which AI runtime powers your project, and which features and configuration options are available.

Agent mode uses next-generation models that typically deliver better accuracy for extraction and complex reasoning tasks. Agent mode projects use AI runtime 2.x versions and offer simplified configuration.

Legacy mode uses traditional processing methods. Legacy mode projects use AI runtime 1.x versions and offer digitization controls, OCR confidence scores and validation rules, and field-level model selections.

New projects use agent mode. Existing projects in legacy mode can update to agent mode to take advantage of the latest product functionality.

Updating existing projects to agent mode

Migrate legacy projects to agent mode for improved extraction accuracy and simplified configuration. Test with app versions before and after migration to measure performance changes.

1. (Optional) Copy the project you want to update.

   Pros — Preserves your original legacy mode prompts so you can test agent mode without losing your working legacy configuration.

   Cons — Ground truth datasets and accuracy tests don’t transfer with copied projects. You must recreate datasets for the copied project and then compare accuracy test results across two different apps.

2. Switch to agent mode in the project header.

3. Update fields and prompts for agent mode.

   • Simplify verbose prompts, especially in extraction fields.

   • Convert reasoning fields used as extraction workarounds to extraction fields with the appropriate field type.

   • Add explicit instructions to remaining reasoning fields, specifying data source, logic, and output format as applicable.

   • Review custom functions for broken references due to modified field names, field order, or OCR confidence or validation rules.

4. Create an app (if you copied a project) or app version.

5. Validate results.

   • If using accuracy testing: Run accuracy test and compare against your baseline.

   • If testing manually: Review extraction results for key fields.

   • Iterate on prompts and configuration as needed, creating new app versions to test changes.