ai-augmented-workflows.md

navigation_title	AI-augmented workflows
applies_to	stack serverless preview 9.3, ga 9.4+ ga
description	Combine Elastic Workflows with Agent Builder to build AI-augmented workflows that pair deterministic steps with agent reasoning.
products	id kibana id cloud-serverless id cloud-hosted id cloud-enterprise id cloud-kubernetes id elastic-stack

AI-augmented workflows [workflows-build-ai-automation]

Workflows and {{agent-builder}} are complementary. Workflows give you deterministic, auditable, event-driven automation: the steps always run in the same order and produce the same kind of result. {{agent-builder}} agents give you reasoning: the ability to interpret unstructured context, classify signals, and generate natural-language summaries. Combining the two lets you build automation that's both reliable and intelligent.

The integration runs in both directions: a workflow can call an agent as a step, and an agent can trigger a workflow as a tool.

What you can build [workflows-ai-patterns]

• Call an agent from a workflow. Use the ai.agent step to invoke any agent built in {{agent-builder}}. The agent sees the workflow's data through template variables, performs its reasoning, and returns a response that subsequent steps can act on. Refer to Call {{agent-builder}} agents from Elastic Workflows for the full ai.agent reference and examples.
• Trigger a workflow from an agent. Create a workflow tool in {{agent-builder}} and assign it to an agent. The agent can then invoke the workflow from a conversation, extracting the needed inputs from the user's message and surfacing the workflow's output in chat.
• Classify and route with AI. Pair ai.classify with switch to send each alert, ticket, or signal down a different branch based on a model's categorization.
• Summarize evidence before action. Use ai.summarize to turn gathered evidence into a concise summary for a case description, notification body, or reviewer prompt.
• Gate AI decisions on human review. Pair AI classification with human-in-the-loop to show the model's output to an analyst before the workflow takes action.
• Send structured prompts to an LLM. Use the ai.prompt step with any configured AI connector to run classification, extraction, or summarization without going through an agent.
• Compose reusable AI building blocks. Extract repeated AI step sequences into a child workflow that parent workflows invoke with workflow.execute.

When to use each direction [workflows-ai-when-to-use]

• Use agent-from-workflow when the workflow already knows what it's doing and needs AI only to reason over a specific data set. For example, summarizing an alert, classifying severity, or extracting fields from unstructured text.
• Use workflow-from-agent when the user (or another agent) is in a conversation and wants to trigger a deterministic procedure. For example, isolating a host, opening a case, or running a set of enrichment queries.

How-to guides [workflows-ai-how-tos]

Step-by-step guides for AI-augmented workflows:

• Classify and route mixed items with AI: Use ai.classify with foreach and if to send each item down a different branch based on a model's categorization, and summarize with ai.summarize.

Learn more

• AI steps reference: ai.prompt, ai.classify, ai.summarize, and ai.agent parameters and output structure.
• Composition steps: Invoke child workflows and emit outputs.
• Human-in-the-loop: Pause a workflow for reviewer input.
• Call {{agent-builder}} agents from Elastic Workflows: End-to-end guide for invoking agents from workflow steps.
• Workflow tools in {{agent-builder}}: Configure an agent to trigger a workflow.
• {{agent-builder}} overview: Concepts, tools, and agent types.