AI Agents FAQ

AI agents in Appian automate complex tasks using large language models (LLMs), tools, and process models. The following questions cover common topics about building and running AI agents.

Getting started with AI agents

What do I need to enable AI agents on my site?

• Appian Cloud only: AI agents are supported in Appian Cloud environments only. You cannot use AI agents in self-managed environments.

• Supported region or cross-region inference: Your Appian Cloud site must either be hosted in a region that supports Sonnet 3.7 or later models, or be opted in to the cross-region inference feature. See Regional availability for AI agents for the full list of supported regions.
• AI model enabled: You must enable at least one AI model that supports AI agents in the Admin Console. AI agents automatically use the latest enabled Sonnet model.

For more on what AI agents are and how they work, see About AI Agents.

What's the best way to get started with a new AI agent?

First, decide which agent type fits your use case. Choose Process Agent if the agent will run autonomously within a process model, or Chat Agent if end users will interact with it through a conversational interface. Then begin with a short, simple prompt and a minimal set of tools. Test the AI agent to see where it fails. Then add more complex instructions or tools to address the identified issues. This iterative approach helps you build prompts that handle real-world complexity without over-engineering from the start.

For detailed guidance, see Best Practices for Designing AI Agents.

What's the difference between a process agent and a chat agent?

Process agents and chat agents differ in how they run and who they interact with:

• Process agents run autonomously inside a process model using the Execute AI Agent smart service. They complete multi-step workflows without human intervention and are best for unattended automation such as case triage, data enrichment, and document processing.
• Chat agents run inside an interface using the agent chat component. End users interact with them through a conversational interface in real time. They are best for virtual assistants, knowledge base Q&A, and guided form-filling.

You select the agent type when you create the agent, and it cannot be changed later. For more details, see Agent types.

Designing AI agents

Should I give an AI agent all available tools?

No. Provide only the tools required for the AI agent's defined tasks. Unnecessary tools consume more AI actions and can cause the large language model (LLM) to select incorrect tools or exceed the context window limit. For complex processes, break down large AI agents into multiple smaller, narrowly-scoped AI agents that each perform a specific task well.

For guidance on tool selection, see Best Practices for Designing AI Agents. For multi-agent design patterns, see the modular AI agents section on that same page.

Does naming convention matter for the AI agent's understanding?

Yes. Use consistent, semantically meaningful names for variables, inputs, and tools. For example, use customerInvoiceDoc instead of docID. The large language model (LLM) is trained on semantic meaning, so clear names help it understand how to use data and pass it between tools. Also provide well-written descriptions for tools directly within the associated design object.

For naming guidance, see Best Practices for Designing AI Agents.

How can I improve the consistency and accuracy of my AI agent's responses?

Use clear, structured prompt engineering with precise action verbs and specific output formatting instructions. Break complex tasks into manageable subtasks. For numerical ranking, use smaller ranges (such as 0-10) and provide qualitative definitions for each score with examples. You can use Chain of Thought or provide static few-shot examples to enhance consistency.

For prompt writing guidance, see Best Practices for Designing AI Agents.

AI agent tools and data

How can I prevent my AI agent from repeatedly calling the same tool?

Set a tool call budget in your AI agent's instructions and give the AI agent clear behavioral guidelines. A tool call budget and behavioral guidelines help the AI agent act consistently and stop before wasting AI actions.

Example instruction: "To answer the customer's question, search the knowledge base. Limit your search to three tool calls. If you can't find a definitive answer after three attempts, stop searching and escalate the case to a human supervisor."

For information on tool call limits and system cancellation, see Troubleshooting AI Agents.

How can I optimize an AI agent to pass data to a process model after querying?

Consolidate data querying within the process model itself, especially when you already know what data to query. Instead of using a separate expression rule tool to fetch data, perform the query inside the process model using a script task. This approach reduces the number of large language model (LLM) interactions and saves AI actions.

For guidance on process model tools, see Create and Configure an AI Agent.

How can I ensure the AI agent stops if a required input is missing?

Add a validation step in your process model tool. The validation step checks all required inputs and returns a clear validation result with suggestions for correction if needed. The large language model (LLM) can then retry and supply the missing or corrected inputs before continuing.

For troubleshooting input issues, see Troubleshooting AI Agents.

How can I use the output of one action as input for another?

In an AI agent run, outputs from previous actions are automatically available as context for subsequent actions. The AI agent is aware of everything that has happened and what data it has access to. If something needs to be passed out of the entire AI agent run, define it as an output variable at the AI agent level.

For details on configuring inputs and outputs, see Create and Configure an AI Agent.

Documents in AI agents

What's the best way to extract structured data from documents like PDFs?

Use DocCenter, which is designed for accurate information extraction from documents with over 95% accuracy. AI agents can use models configured in DocCenter by passing document IDs to a process model tool.

You must explicitly instruct the AI agent to use the DocCenter tool for a specific document input. If you do not explicitly instruct the AI agent to use the DocCenter tool, the AI agent may select one of the system document tools at runtime.

For detailed guidance on document strategies, see Working with Documents in AI Agents.

How can I integrate web content as context for an AI agent?

Set up a web search integration in a process model tool. The process model calls the external integration, and the AI agent uses the process model as a tool to retrieve web content.

For guidance on process model tools, see Create and Configure an AI Agent.

Why are queries returning a limited number of columns?

The AI agent uses dynamic projection to speed up queries by reducing the number of fields returned. The system balances performance with the amount of data retrieved. To get more consistent results, specify the desired fields explicitly in your prompt instructions.

For details on record type query behavior, see Troubleshooting AI Agents.

What's the best way to upload documents for testing?

Documents must be stored in the Appian repository for use in test cases. Upload your test documents to Appian before running tests in the AI agent designer.

For testing guidance, see Create and Configure an AI Agent.

AI agent performance and cost

Why is my AI agent's AI action count unusually high?

AI action usage includes the AI agent's goal, instructions, inputs, outputs, tool definitions, and conversation history within the current run. More tools directly lead to higher AI action consumption.

To optimize AI action usage:

• Offload programmable tasks: Move data queries outside the AI agent's run where possible, then pass only the relevant information as context.
• Use expression rules: Query data with expression rules instead of having the AI agent directly query a record type. The AI agent only needs to provide filters for inputs rather than understanding the entire schema.
• Avoid Read Full Document: Do not use the Read Full Document tool to load an entire document into the AI agent's context window unless your use case requires it. The Semantic Search tool consumes fewer AI actions.

For cost considerations, see About AI Agents. For document tool cost tradeoffs, see Working with Documents in AI Agents.

When should AI agent calls be synchronous versus asynchronous?

For most scenarios, use a synchronous AI agent call. A synchronous call ensures the AI agent completes its tasks before proceeding and provides consistent output. For tasks that can run in parallel (such as sending an immediate email confirmation while other processing continues), asynchronous calls reduce wait time for the end user.

Appian AI agents primarily support synchronous process model calls. You can implement asynchronous human-in-the-loop processes using an async subprocess call within a process model tool.

For integration details, see Use AI Agents in a Process.

Context window management

How does the AI agent handle its context window?

Two practices help AI agents manage context effectively:

• Semantic Search (RAG): Instead of loading all documents into the AI agent's context, store information externally and use the Semantic Search tool to retrieve only the most relevant content. This allows an AI agent to access large knowledge bases without exceeding context limits.
• Task decomposition into sub-agents: Break complex goals into smaller, independent tasks. Each sub-agent handles a focused task with a fresh context window. The output of one AI agent becomes the concise input for the next. This prevents context contamination and keeps each AI agent focused.

For document tool guidance, see Working with Documents in AI Agents. For multi-agent design, see Best Practices for Designing AI Agents.

AI agent governance and compliance

How can I ensure human oversight for AI agent outputs?

Implement human supervision for AI output. The AI agent's output should explain its steps and reasoning for review by a human. This is important for auditability and compliance with regulations that categorize AI systems as assistants rather than final decision-makers.

For guidance on designing for human oversight, see About AI Agents.

AI agent configuration management

How frequently should I update my AI agent's model?

Appian AI agents use pre-trained models (the latest Claude Sonnet) and do not perform custom fine-tuning with your specific data. The AI agent gains context for each run from the tools and inputs you provide. Changes in AI agent behavior result from modifications to your configuration (goals, instructions, tool definitions, input data) or the non-deterministic nature of the large language model (LLM), rather than from model updates.

How can I revert to a previous version of my AI agent?

The AI agent design object includes built-in versioning. Each time you update the configuration, add or modify test cases, and click SAVE, the system automatically creates a new version. To revert to an earlier version, open the Versions list, select the desired version, and click SAVE in the new tab to make it the latest version.

How can I track AI agent processing time?

Check the Monitor tab in the AI agent designer to see the total duration for each AI agent run.

For monitoring details, see Monitor AI Agents.

Feedback