Agentic AI Assistant

The experimental agentic AI Assistant is a new generation of the GoodData AI Assistant. It is designed for more guided interactions and more complex analytical tasks. It can ask follow-up questions, help narrow down ambiguous requests, and combine multiple analytical steps into one response.

Before You Start

An administrator must enable the experimental agentic AI Assistant in the Settings before users can use it.

You also need:

• AI Assistant configured for your organization
• permission to use AI Assistant in the workspace
• a supported model configured for the experimental agentic AI Assistant
• data sharing configured according to the use cases you want to enable

For the standard AI Assistant setup, see Configure AI Assistant.

Data Sharing

The experimental agentic AI Assistant supports an opt-in data sharing setting. It can be turned on in the workspace settings. For administrators who manage settings through API, the setting name is enableAiOnData.

When data sharing is enabled, the assistant can use richer skill composition for more advanced analytical tasks. This includes workflows where the assistant creates a visualization, evaluates the result, and then uses additional skills such as anomaly detection or Key Driver Analysis to continue the analysis.

When data sharing is disabled, which is the default, the assistant still supports simpler use cases, such as creating visualizations, but some more advanced multi-step workflows may not be available.

For more information about individual skills and their data-sharing behavior, see Agentic Capabilities.

Recommended Models

The experimental agentic AI Assistant is designed to work with reasoning-capable models.

Use a GPT-5.x model, such as gpt-5.2, for the best experience.

Older models used by the standard AI Assistant, such as gpt-4o, are not supported for the experimental agentic AI Assistant. These models may still remain available for the standard AI Assistant while both experiences coexist.

How It Differs from the Standard AI Assistant

The experimental agentic AI Assistant currently behaves differently from the standard AI Assistant.

Current differences include:

• it is designed for more complex, multi-step tasks
• it can combine multiple analytical skills into one response
• some advanced workflows depend on the current data sharing setting
• however, it currently supports fewer features

For straightforward questions and the broadest current feature coverage, the standard AI Assistant may still provide a more familiar experience.

Use the Experimental Agentic AI Assistant

The experimental agentic AI Assistant is useful when your request needs clarification or when the answer requires several analytical steps.

Clarify Ambiguous Requests

If your question can mean several different things, the assistant may ask you to choose what you want before it creates a result.

For example, if a business term can map to several objects in your semantic model, the assistant may explain the ambiguity, offer a small set of choices, and continue only after the intent is clear.

Guide the Next Step

Instead of returning a generic response, the assistant can guide you through what it needs next.

For example, when creating a visualization, it may ask:

• what to measure
• how to break it down
• what time period to use

Suggest Alternatives

If a result is empty or a requested output is not supported, the assistant may explain the issue and suggest a better option.

For example, it may:

• suggest a wider date range when a query returns no data
• explain when a requested visualization type is not supported and offer alternatives

Ask for Missing Context

If the assistant does not have enough context, it can ask for the missing detail before continuing.

For example, if it cannot determine which person, metric, or period you mean, it may ask you to specify that detail.

Combine Multiple Skills

The main difference from the standard AI Assistant is that the experimental agentic AI Assistant can combine multiple skills in a single request.

For example, it can:

• create a visualization
• check the result for anomalies
• run Key Driver Analysis for detected anomalies
• prepare a final summary based on the combined result

This can help with tasks such as understanding why a metric changed, identifying unusual behavior, and summarizing the likely explanation in one response. If the assistant already has enough information, it can perform this analysis directly without asking for confirmation first.

API Changes for Integrations

The chat API for the experimental agentic AI Assistant is new and differs from the earlier AI Assistant chat API.

If you integrate the assistant through HTTP API, check the current API reference in your GoodData environment:

/apidocs/?urls.primaryName=Gen-AI+%28v1%29

The earlier chat APIs continue to work for the previous AI Assistant experience while that experience is still supported. They will be deprecated separately with the earlier chatbot.

If you maintain an existing integration, review the new API before enabling the experimental agentic AI Assistant.

Examples

Clarify a business term
The user asks: What is the total ecom spend for Q1 2025?
The AI Assistant then:

• explains that ecom may be ambiguous
• offers a few possible mappings
• continues after the user selects one

Guide a visualization flow
The user asks: Create a new visualization.
The AI Assistant then:

• asks what to measure
• asks how to break it down
• asks for an optional time period

Recover from an unsupported request
The user asks: Create a heatmap for my top industries.
The AI Assistant then:

• explains that the visualization type is not supported
• suggests an alternative supported output

Start a multi-step analysis
The user asks: Help me understand why revenue dropped last month.
The AI Assistant then:

• creates the relevant visualization
• checks the result for anomalies
• runs Key Driver Analysis for detected anomalies
• returns a summarized explanation in one response

Current Limitations

This feature is still evolving. At this time:

• behavior may differ from the standard AI Assistant
• some standard AI Assistant features are not yet available
• some requests may return fewer interaction shortcuts than users expect
• unsupported requests may require a simpler prompt or a different workflow
• advanced skill composition depends on the current data sharing setting
• the experimental agentic AI Assistant requires supported reasoning-capable models
• the experimental agentic AI Assistant uses a different chat API than the earlier AI Assistant