Generative AI Skills

What are generative AI skills?

This topic describes AI skills that use generative AI to automate many common tasks in your business processes.

Generative AI skills handle document digitization, even for scanned or image-based content. They support file types like JPEG, PNG, TIFF, PDFs, TXT, MD, HTML, and DOCX, and automatically apply Optical Character Recognition (OCR) to extract text for seamless processing.

What is generative AI?

Generative AI models create new content. Generative models take a prompt as input (such as a question or request). As the model evaluates that request, it recalls what it knows about those characteristics. As it begins to formulate a response, it uses probability to determine what comes next.

The output is entirely new content. Rather than regurgitating information word-for-word, generative models can formulate more specific answers, crafted just for that prompt. Examples of generative models include chat bots and image generators.

When to use generative AI

Generative AI is very flexible, so you can use it in multiple ways in your Appian processes. You can use generative AI to summarize content, extract information you're interested in, and detect sensitive information. You can even specify how you want those results to appear: in code, a paragraph, or a bulleted list.

Here are a few examples of when to use generative AI in Appian:

If you want to…	Use…
Quickly determine whether a form submission is a support request or a general inquiry	Text classification or unstructured email classification
Extract essential details of a legal document, such as a contract	Unstructured document extraction
Summarize action items and assignments in a meeting transcript	Text extraction
Identify sensitive, legally protected information in a portfolio of medical documents	PII extraction from documents
Evaluate if your business is a good fit for a contract opportunity	Document summarization
Create content for a new product launch, in line with your business's branding	Text generation

See below to learn more about when to use generative AI or Advanced IDP Tools for document extraction.

AI Skills using generative AI

All generative AI skills create a prompt builder AI skill with a pre-populated template based on your use case. You can adjust the prompt, control the temperature, provide examples, and test the prompt directly within the AI skill.

If you don't see an AI skill suited to your use case, you can always use the Prompt Builder. Use this skill to create a prompt to send to generative AI models. You'll tell the model what you want it to do using plain language. After adjusting a few settings and testing the prompt, you can see what the model returns based on your instructions. Use this skill to summarize long or complex text, generate new content, and more!

Classification

• Text: Identify text based on certain traits. You'll tell us about the types of text you want to classify in your business process and provide samples of each type.

Extraction

• Text or Emails: Extract data from text or emails. You'll tell us what you want to extract and how you want that information to appear in the output.
• Documents: Extract data from a document.
  • Unstructured: You'll tell us what you want to extract from an unstructured document and how you want that information to appear in the output. For example, this skill can help you extract data from large or free-flowing paragraphs of text, like contracts or from visual elements like checkboxes, charts, or images.
• Spreadsheets: Extract data from Excel spreadsheets (XLS, XLSX, XLSM). You'll tell us what fields to extract, and the skill determines the file context and returns structured JSON with both values and cell references. See Create a Spreadsheet Extraction AI Skill for details.

Generation

• Text: Create text for any purpose. You'll tell us what text you want to generate and whether you want the output to include any specific logic, format, or other requirements.

Summarization

• Text, Email Text, or Document Text: Summarize text from any source, including emails and documents. You'll tell us what you want to summarize and how you want the summary to appear.

PII Detection

• Text, Email Text, or Document Text: Extract personally identifiable information (PII) in a document, email, or text block. You'll tell us the personally identifiable information (PII) you want to extract, such as names or phone numbers.

Usage considerations

AI Actions consumption

Each execution of a generative AI skill consumes AI Actions. The Execute Generative AI Skill smart service returns the number of AI Actions consumed per execution, and the AI Skill design object test panel displays consumption during testing.

What makes a good prompt?

In the context of AI, a prompt is how we communicate with the large language model (LLM) about what we want it to do. A well crafted prompt can help you communicate with the model more effectively, and reduce the amount of time you spend tweaking or troubleshooting.

Learn how to write an effective prompt on the AI Skills Best Practices page.

File types

Generative AI skills accept PDFs, JPEG, PNG, TIFF, TXT, MD, HTML, and DOCX files as direct inputs. The way these files are processed depends on the option you select for the AI Skill. If the skill is set to Text only, only the text content is analyzed. If it's set to Visual elements & text, the skill also analyzes the visual elements of the document.

Text-based files (TXT, MD, HTML) are always processed as text input, regardless of mode selection. These file types are not supported in "Visual elements & text" mode.

DOCX files are supported in "Text only" mode, where text is extracted from the document and sent to the model.

Different requirements apply for each input type:

• Text: Only plain text is supported for text input, so file size, page count, and image limits do not apply.
• Email: EML files are supported with a maximum file size of 25 MB. Page and image size limits do not apply.
• Document: See the following table for the Document input type requirements.

Document input type requirements

File Type	Mode	Max File Size	Page Limit	Max Image Size	Notes
PDF	Text only	25 MB	100 pages	N/A	---
	Visual elements & text (Claude 4+ models)	24 MB	100 pages	N/A	---
	Visual elements & text (Claude 3 Sonnet)	25 MB	20 pages	N/A	---
	Visual elements & text (Amazon Nova 2 Lite)	18 MB	100 pages	N/A	---
JPEG, PNG	Text only	10 MB	N/A	10k x 10k pixels	---
	Visual elements	3.75 MB	N/A	8k x 8k pixels	---
TIFF (single-page)	Text only	10 MB	1 page	10k x 10k pixels	---
	Visual elements & text	3.75 MB	1 page	4k x 4k pixels	---
TIFF (multi-page)	Text only	10 MB	100 pages	4k x 4k pixels	---
	Visual elements & text	18 MB (3.75 MB per page)	20 pages	4k x 4k pixels	---
TXT	Text only	4.5 MB	No page limit	N/A	Not supported in "Visual elements & text" mode.
MD	Text only	4.5 MB	No page limit	N/A	Not supported in "Visual elements & text" mode.
HTML	Text only	4.5 MB	No page limit	N/A	Not supported in "Visual elements & text" mode.
DOCX	Text only	4.5 MB	No page limit	N/A	Not supported in "Visual elements & text" mode. Refers to the Office Open XML format (.docx). The legacy binary format (.doc) is not supported.

Excel and CSV input type requirements

The spreadsheet extraction skill has the following requirements for Excel files (XLS, XLSX, XLSM) and CSV files:

Constraint	Limit
Maximum file size	10 MB
Maximum number of cells	2 million cells
Maximum rows per sheet	40,000
Maximum columns per sheet	200
Maximum sheets	100
Maximum JSON response size	25 MB

Language support

Regional availability for the Visual elements & text option

To use generative AI skills with the Visual elements & text option enabled, you may need to configure an inference profile. If you're using an older model that doesn't use inference profiles, contact Appian Support to use this feature by sending your data to a supported region.

AI models

AI models are managed by system administrators in the Appian Administration Console. These settings control which large language models are available for use in Appian AI capabilities.

Developers select the model to use in their AI skills from the models enabled in their environment.

Model selection determines which generative AI model an AI skill uses to process prompts or documents. Each model offers different strengths. Some models prioritize speed, while others provide more advanced reasoning. Choosing the right model helps tailor your AI skill to your application's performance and accuracy needs.

You also select the model directly in the AI skill design object. From there, you can view all models available in your environment and choose the one that best fits your use case. You may see additional models in the AI skill object that do not appear on the AI Services page. These models are automatically enabled in your environment and don't require any configuration on the AI Services page.

For detailed configuration information, see the AI Services page.

Categories

Models are organized into two categories: Fast Models and Reasoning Models. Choosing the right model depends on the complexity of your task.

• Fast Models: These models are optimized for speed and are ideal for simple, high-volume tasks that require quick responses. Fast models include Claude Haiku 4.5, Amazon Nova 2 Lite, GPT 5.4 nano, GPT 5.4 mini, GPT 5 nano, GPT 5 mini, and GPT 4.1 mini.
• Reasoning Models: These models are more powerful and are designed for complex tasks that require advanced reasoning, such as analyzing intricate documents or solving multi-step problems. Using these models may result in higher accuracy but longer processing times. Reasoning models include:
  • Claude models: Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Sonnet 4, and Claude Sonnet 3.7.
  • Appian GPT models: GPT 5.4 and GPT 5.5.
  • BYO AWS GPT models: GPT 5.4 and GPT 5.5.
  • BYO Azure models: GPT 5.4 Pro, GPT 5.4, GPT 5.3 codex, GPT 5.2 codex, GPT 5.2, GPT 5.1 codex max, GPT 5.1 codex, GPT 5.1, GPT 5 Pro, GPT 5 codex, and GPT 4o.

Auto model selection

Auto model selection uses the Appian recommended model within your selected provider. Instead of selecting a specific model, you select Auto and Appian handles model selection for you.

Auto model selection is available for every enabled cloud provider: Appian, Amazon Web Services, Microsoft Azure, and custom providers. Each provider has its own Auto option. No cross-provider routing occurs—if your selected provider's models are unavailable, the execution fails rather than routing to a different provider.

How Auto model selection works

When an AI skill uses Auto mode, Appian selects a model at execution time based on the following priority logic:

• Text only operations: Fast models are prioritized over reasoning models.
• Visual elements & text operations: Reasoning models are prioritized over fast models.

Within each category, Appian sorts by model family and release date (newest first) to select the most capable available model.

Auto mode model identifier

Auto mode is identified by the model ID <provider>/auto, where <provider> is one of the following: appian, aws, azure, or custom. For example:

• appian/auto
• aws/auto
• azure/auto
• custom/auto

You can copy this identifier from the AI skill settings to use it with the Execute Generative AI Skill smart service at runtime.

Default behavior for new and existing AI skills

• New AI skills: Default to Auto mode. Appian selects the best available model within the configured provider.
• Existing AI skills: Remain on their currently configured specific model. No automatic migration occurs.

Audit logging with Auto mode

The AI audit log records the actual resolved model identifier for each execution, not "auto." For example, if Auto mode resolves to Claude Haiku 4.5, the audit log shows appian/anthropic.claude-haiku-4-5-20251001-v1:0. This provides full visibility into which model processed each request.

Extended thinking

Extended thinking enhances reasoning for complex tasks by prompting the model to work through its approach more thoroughly and explain the steps it takes. When enabled, the model breaks down intricate tasks into logical, sequential steps before generating a final response. This increases accuracy for sophisticated automation, such as post-extraction validation or complex document classification.

Extended thinking is available for the following models:

• Claude Sonnet 5
• Claude Sonnet 4.6
• Claude Sonnet 4.5
• Claude Sonnet 4
• Claude Sonnet 3.7
• Claude Haiku 4.5
• Amazon Nova 2 Lite

Extended thinking with Amazon Nova 2 Lite

Amazon Nova 2 Lite supports extended thinking. However, the model provider does not return the reasoning output. When a developer enables extended thinking and selects Amazon Nova 2 Lite, the Show Thinking tab is disabled and displays a message that extended thinking is active but the reasoning output is not available from the model provider. The Model Thinking output from the Execute Generative AI Skill smart service returns an empty value. The model still produces a final response as expected.

Extended thinking with BYO Azure models

BYO Azure models have varying levels of extended thinking support. The extended thinking toggle in the AI skill designer adapts based on the model you select.

The following BYO Azure model does not support extended thinking:

• GPT 4.1 Mini

For models that support extended thinking with the option to disable it, the toggle is enabled and you can turn extended thinking on or off. The following BYO Azure models support disabling extended thinking:

• GPT 4o
• GPT 5.1
• GPT 5.1 Codex
• GPT 5.1 Codex Max
• GPT 5.2

For models where extended thinking is always active, the toggle is turned on by default and cannot be turned off. The toggle appears disabled (grayed out), and a tooltip explains that reasoning is permanently enabled for the selected model. The following BYO Azure models always have extended thinking enabled:

• o3 Pro
• GPT 5
• GPT 5 Codex
• GPT 5 Mini
• GPT 5 Nano
• GPT 5 Pro
• GPT 5.2 Codex
• GPT 5.3 Codex
• GPT 5.4
• GPT 5.4 Mini
• GPT 5.4 Nano
• GPT 5.4 Pro

Extended thinking with Appian GPT models

Appian GPT models (GPT 5.4 and GPT 5.5) always have extended thinking enabled. The toggle is turned on by default and cannot be turned off. The toggle appears disabled (grayed out), and a tooltip explains that reasoning is permanently enabled for the selected model.

Extended thinking with Auto model selection

When an AI skill uses Auto model selection, the extended thinking toggle remains available in both the AI skill designer and the Execute Generative AI Skill smart service. However, Appian does not use your extended thinking preference to influence which model is selected. Auto mode selects the best model based on operation type and priority logic independently of the extended thinking setting. After Appian selects a model, it applies your preference when possible and silently ignores it when the resolved model is not compatible.

Temperature with BYO Azure models

The temperature slider controls how creative or deterministic a model's responses are. For BYO Azure models, the temperature slider is only available for the following models when extended thinking is disabled:

• GPT 4o
• GPT 4.1 Mini

For all other BYO Azure models, the temperature slider is hidden in the AI skill designer because these models are reasoning models that do not support temperature adjustments.

Amazon Nova 2 Lite document processing

Amazon Nova 2 Lite supports both the Text only and Visual elements & text document processing modes.

Native PDF support for Amazon Nova 2 Lite has the following limitations:

• Maximum file size: 18 MB.
• Maximum page count: 100 pages.
• LLM Context Window: Documents that exceed the model's context window limit will fail to process, even if they meet the file size and page count requirements. Dense or complex documents may reach this limit before reaching the page or file size maximums.`
• Unsupported encodings: PDFs that contain CMYK or SVG image encodings are not supported.

BYO Azure document processing

BYO Azure models support both the Text only and Visual elements & text document processing modes.

Document processing for BYO Azure models has the following limitations:

• Maximum file size:
  • Text only mode: 25MB.
  • Visual elements & text mode: 24 MB.
• Maximum page count: 100 pages.
• LLM Context Window: Documents that exceed the model's context window limit will fail to process, even if they meet the file size and page count requirements. Dense or complex documents may reach this limit before reaching the page or file size maximums.

Appian GPT document processing

Appian GPT models (GPT 5.4 and GPT 5.5) support both the Text only and Visual elements & text document processing modes.

Document processing for Appian GPT models has the following limitations:

• Maximum file size:
  • Text only mode: 25MB.
  • Visual elements & text mode: 24 MB.
• Maximum page count: 100 pages.
• LLM Context Window: Documents that exceed the model's context window limit will fail to process, even if they meet the file size and page count requirements. Dense or complex documents may reach this limit before reaching the page or file size maximums.

Availability

The available models depend on the configurations set in the AI Services page of the Admin Console and the region your environment is in. Due to this, you may see a different list of models in your development environment than in your production environment.

Appian GPT models (GPT 5.4 and GPT 5.5) appear in the model selection dropdown when an administrator has selected Appian as the cloud provider on the AI Services page and enabled these models. Appian GPT models are available in select regions.

BYO Azure models appear in the model selection dropdown when an administrator has configured the Microsoft Azure provider on the AI Services page. If the Azure provider is not configured, Azure GPT models do not appear as options.

Similarly, BYO AWS models (including GPT 5.4 and GPT 5.5) appear in the model selection dropdown when an administrator has configured the Amazon Web Services provider on the AI Services page. If the AWS provider is not configured, BYO AWS models do not appear as options.

The list of available models also changes depending on the AI skill's document processing mode. Haiku 3 supports text extraction only and does not include vision-based capabilities.

Selection considerations

Use these considerations to understand how model selection behaves across environments and what to expect when managing AI skills to ensure your automated processes run smoothly.

• Administration: Administrators can disable entire AI model providers at the environment level from the AI Services page. When a provider is disabled, Appian does not route any AI requests to that provider's models. For Appian-managed models, disabling the provider also removes those models from the automatic fallback chain. Disabling a provider also removes its models from Auto mode routing.
• Import and export: Before importing an AI skill, make sure the appropriate generative AI model is enabled in the target environment. Appian does not validate the model until the AI skill runs. The model ID must match exactly across environments. For Auto mode AI skills, the target environment must have at least one model enabled for the configured provider.
• Model deprecation: If a selected model is deprecated or disabled by an administrator, the AI skill execution will fail. AI skills using Auto mode are not affected by individual model deprecations—Appian automatically routes to the next available model within the same provider.
• Model selection: Select the latest available model. These models offer the best performance according to industry benchmarks. Alternatively, select Auto to let Appian manage model selection for you.
• Third-party cloud provider fallback: Automatic model fallback only occurs when Auto mode is selected. If you select a specific model and that model becomes unavailable, the execution fails with an error. To gain fallback protection for third-party cloud providers, use Auto mode.
• Azure provider fallback: BYO Azure models do not support automatic fallback when a specific model is selected. If an Azure model is unavailable, the AI skill execution returns an error. To gain automatic model routing for Azure, select Auto mode (azure/auto).
• Runtime model override: When you pass a model ID at runtime using the Execute Generative AI Skill smart service, Appian uses that model exactly as specified. If the model is invalid or unavailable, the execution fails without fallback. You can also pass <provider>/auto at runtime to use Auto model routing for that execution.

How to select a model

Model selection gives you a clear view of the generative AI models available for your AI Skills. The model your AI Skill is currently using won't change unless you manually select a new model and save your changes.

You can select a model when configuring the AI skill or at runtime using the Execute Generative AI Skill smart service.

When configuring an AI skill

When you create or edit an AI skill, you can select a model.

To select a model when you configure an AI skill:

1. In your application, open an existing AI skill or create a new one.
2. Open the Settings pane.
3. Under Model Selection, click the dropdown to see a list of available models.
4. Choose a model from the list. New AI skills default to Auto, which lets Appian dynamically select the best available model within your configured provider. You can also select a specific model if you want full control over which model runs.

At runtime

You can dynamically select a model within a process by using the Runtime Model input on the Execute Generative AI Skill smart service. This allows you to build more flexible and efficient processes.

The Runtime Model input accepts both specific model IDs and <provider>/auto. Passing <provider>/auto at runtime uses Auto model routing for that execution, regardless of how the AI skill is configured at design time.

For example, in a document processing workflow, you could use a fast model for the initial data extraction and then use Auto mode for a final review, letting the platform select the best reasoning model available.

Document extraction: Generative AI or Advanced IDP Tools?

When you extract data from documents, you can choose from different AI skills based on your document structure and extraction goals. Each AI skill is designed to handle specific scenarios and can be used individually or together for more accurate results.

This section explains the difference between Appian's two main document processing options and provides patterns to help you decide which is best for your use case.

• Advanced IDP Skill is a high-precision document recognition engine. It excels at identifying the text, layout, and visual elements in a document.
• Generative AI Skill is a reasoning engine. It excels at understanding the context of the text to answer questions and extract information.

When you create a document extraction model in DocCenter you can combine these AI skills by selecting the Advanced Layout option.

How Advanced IDP and Generative AI skills differ

Think of the Advanced IDP Tools skill as the tool that reads the document and the Generative AI skill as the tool that understands it.

When you send a document directly to a generative AI skill as a Text only document type, Appian performs a standard optical character recognition (OCR) step to read the text. For more complex documents, you can use the Advanced IDP skill first to perform a more detailed reading of the document's structure and content, then send that detailed map to the Generative AI skill for a more accurate analysis.

The table below summarizes the key differences.

Feature	Advanced IDP skill (The "Reader")	Generative AI skill (The "Thinker")
Primary Job	To perform high-precision OCR, converting a document image into structured text and layout data.	To understand, reason about, and generate content based on text input.
Strengths	Positional Awareness: Knows the exact coordinates of every word. Required for highlighting extracted data, detecting signatures, and extracting data from fixed forms.	Contextual Understanding: Can interpret nuanced language, infer meaning, and follow complex instructions to extract data from unstructured paragraphs.
Output	A structured map of located text, tables, and form elements with confidence scores.	New, generated text, typically in a structured JSON format as requested in the prompt.

Choosing the right pattern for document extraction

Your document's complexity and your business needs determine the best way to configure document extraction. This section describes three common patterns.

• Use only Generative AI skill
• Use only Advanced IDP skill
• Use Advanced IDP skill with Generative AI skill

Use only Generative AI skill

This approach is best for simple, text-heavy documents where the visual layout is not important. By passing the document directly to the Execute Generative AI Skill smart service, you can choose between two document processing modes:

• Text only: Use this mode when your document contains only text content. This mode uses built-in OCR to retrieve the text and supports up to 100 pages.
• Visual elements & text: Use this mode when your document contains visual elements like checkboxes, charts, images, or handwritten annotations that you need to extract. For Claude 4+ and Nova 2 Lite models, this mode also supports up to 100 pages.

Choose the mode based on your document's content type, not page count. If your document has visual elements you need to extract, use "Visual elements & text" mode. If it's text-only, use "Text only" mode for faster processing.

Use this pattern for documents like:

• Emails
• Transcripts
• Simple letters

Use only Advanced IDP Tools skill

For some use cases, you don't need the Generative AI skill's reasoning capabilities. The Advanced IDP Tools skill is effective on its own for tasks that depend on a document's structure. To use this pattern, create an Advanced IDP Tools AI Skill and call it using the Advanced IDP Tools smart service in a process model.

Use this pattern to:

• Index documents for search: Use the Text tool to extract all content from a set of documents and write it to a database field that is enabled for smart search.
• Verify signatures or checkboxes: Use the Signatures or Checkboxes tool to confirm if a document is signed or which options are selected.
• Extract from fixed forms: For highly structured documents that never change like certain government forms.
• Leverage Confidence Scores: Use the confidence scores returned by the AI skill to support automated decisions, such as straight-through processing, when the score exceeds a defined threshold.

Use Advanced IDP Tools skill with Generative AI skill

This is the most robust pattern, designed for complex or semi-structured documents where you need both high-precision OCR and intelligent reasoning. To use this pattern, select Advanced Layout when you configure your document extraction model in AI Document Center.

Use this pattern for documents like:

• Invoices with varying formats
• Complex contracts
• Utility bills

How it works

When you use the Advanced Layout option, the extraction process happens in two steps:

1. Advanced IDP Tools skill scans the document to create a detailed, structured map of all its content, including checkboxes, tables, and signatures. Appian then automatically formats this map into a clearly structured phrase that AI can understand more easily.
2. Generative AI skill receives the formatted text from the first step. Your prompt then guides the model to reason over this high-quality text to find and extract the data you need.

Highlighting extracted data

When you enable the Bounding option in a document extraction model, Appian uses the x- and y-coordinate data from the Advanced IDP Tools skill. This allows you to show users exactly where on the page each piece of data was found on the reconciliation task interface.

AI action usage consideration

This combined pattern provides the highest accuracy but also uses more AI actions than other patterns. Converting the detailed structural data from the Advanced IDP Tools skill into a formatted text string can result in a large input for the generative AI skill. However, the improved accuracy often outweighs the additional AI action cost.

Feedback