Agentic Workflow AI Agents
Audience: Administrators, Developers, Solution Architects
Purpose: Complete reference for the AI Agent action types available in Kizen Agentic Workflows, including how to call a language model directly, extract structured data from file fields, and transcribe audio content into text that downstream steps can consume.
Overview
AI Agent actions bring large language model capabilities directly into an Agentic Workflow. Each action invokes a configured LLM provider against a specific input (a prompt, a file, or an audio asset) and stores the result in an Agentic Workflow variable that subsequent steps can reference.
Note: For error handling, see Agentic Workflow Error Handling.
Why AI Agents Matter
Much of the data that drives business processes arrives in formats that traditional Agentic Workflow cannot read: PDFs, scanned documents, images, voice notes, recorded calls, and freeform text. AI Agent actions close that gap. They let an Agentic Workflow read a contract, pull values out of an invoice, summarize a transcript, or transcribe a customer voicemail, then route that output into ordinary downstream steps like field updates, messages, or branching logic.
The result is an Agentic Workflow that operate on the full range of content your business actually receives, not just the structured fields already in the system.
System Behavior of AI Agents
Before configuring an individual AI Agent action, a few behaviors apply across all three action types.
LLM provider configuration is required: Every AI Agent action depends on at least one LLM provider being enabled in your business settings. If no providers are configured, the Choose Model dropdown displays No Options and surfaces a Configure LLM Providers link. AI Agent actions cannot run until a provider is enabled.
Output is stored in an Agentic Workflow variable: Each AI Agent action writes its result to a variable that downstream steps reference by name. Plan variable naming before configuring the step so later actions can consume the output reliably.
Source compatibility depends on context: File Extraction and Audio Transcription draw their Choose Source options from file-type fields on the context Record or its related Objects. If no compatible field exists, the dropdown displays No Options. Confirm the context Record exposes the right file field before adding these actions.
Supported file formats are fixed: File Extraction supports PDF, JPG, JPEG, and PNG. Audio Transcription supports MP3 and WAV. Files outside these formats cannot be processed and will produce an error at runtime.
Prompts shape output quality: File Extraction and Audio Transcription both ship with a default prompt that works for general cases. Customizing the prompt with specifics about what to extract, how to format it, or how to handle ambiguity will improve accuracy on specialized content.
Model selection is per-action: Each AI Agent action chooses its own model independently. Different steps in the same Agentic Workflow can target different models based on the task, cost, or latency profile that fits best.
Types of Agentic Workflows Actions
Call LLM
The Call LLM action sends a prompt to a configured language model and stores the response in an Agentic Workflow variable for use in downstream steps.
Use the Choose Model dropdown to select the language model to invoke. Available models are populated based on the LLM providers configured for your business.
If no providers have been configured, the dropdown displays No Options and a Configure LLM Providers link is shown below the dropdown. Follow that link to enable one or more providers in your business settings before using this action.
File Extraction
The File Extraction action uses a configured language model to extract structured data from a file attached to or associated with the context Record.
Choose Source: Use the Choose Source dropdown to select the file field or attachment source to extract from. Sources are populated based on file-type fields available on the context Record or its related Objects. If no compatible sources exist, the dropdown displays No Options. Supported file formats are PDF, JPG, JPEG, and PNG.
Choose Model: Use the Choose Model dropdown to select the language model to use for extraction. Available models are populated based on the LLM providers configured for your business. If no providers have been configured, the dropdown displays No Options and a Configure LLM Providers link is shown. Follow that link to enable one or more providers in your business settings before using this action.
Prompt: A freeform text field where you can provide additional context or instructions to guide the extraction. A default prompt is provided to you but customizing it can improve accuracy in cases where the file content requires disambiguation or specific extraction behavior.
Audio Transcription
The Audio Transcription action uses a configured language model to transcribe audio content from a file attached to or associated with the context Record, converting spoken content to text for use in downstream steps.
Choose Source: Use the Choose Source dropdown to select the audio file field or attachment source to transcribe. Sources are populated based on file-type fields available on the context Record or its related Objects. If no compatible sources exist, the dropdown displays No Options. Supported file formats are MP3 and WAV.
Choose Model: Use the Choose Model dropdown to select the language model to use for transcription. Available models are populated based on the LLM providers configured for your business. If no providers have been configured, the dropdown displays No Options and a Configure LLM Providers link is shown. Follow that link to enable one or more providers in your business settings before using this action.
Prompt: A freeform text field where you can provide additional context or instructions to guide the extraction. A default prompt is provided to you but customizing it can improve accuracy in cases where the file content requires disambiguation or specific extraction behavior.
What's Next
Continue to SmartConnectors to learn how to move data into and out of Kizen on a recurring schedule, including how to configure imports, map fields, and handle errors across large data sets.
Last updated
Was this helpful?