Frequently asked questions

Where can I use GenAI Activities?

You can use GenAI Activities with the following products.

Apps

UiPath Apps has a native integration with Integration Service. You can use GenAI Activities in display controls and as events directly. You can also build apps that initiate workflows that contain GenAI Activities.

Studio Web

UiPath GenAI Activities are available in Studio Web. Use the activity search pane to find the UiPath GenAI Activities category or search for individual activities by name.

Studio Desktop on version 2023.10 or newer

UiPath GenAI activities, like all Integration Service activities, are available in Studio, in the Activities panel, in the Available section, as part of UiPath.IntegrationService.Activities. There is no need to manually manage the package. The most recent version of the activities is automatically applied to your dependencies when you add an activity to your workflow. If you open a new project and add the activities in it, the latest version is applied.

Autopilot for developers

GenAI Activities are incorporated into the Autopilot for developers index. If you describe an automation that requires an LLM call or a specific activity like Categorize, that activity is generated by your natural language description of your automation.

Action Center

Automation developers use Action Center to incorporate a human-in-the-loop for validation purposes. This is a crucial step for automations containing GenAI. You can and should incorporate the validation of GenAI outputs in addition to your normal debugging and testing. You can easily set up conditions under which human validation is required using If statements (deterministic) or by executing a separate prompt activity to evaluate the output. Under the appropriate conditions, you can send the output to Action Center using one of the following options:

• Action Apps – Action Apps allow you to build a custom application that accepts inputs and input/outputs from workflows that are paused until a human reviews something before proceeding. With Action Apps, you can use the full capabilities of UiPath Apps, including custom user interfaces, Integration Service API calls, and the ability to call other workflows. Finally, you can approve and send the confirmation back to the paused workflow.
• Tasks – Tasks are simple steps sent to Action Center. They are not as easy to customize as Action Apps, but can be used to approve or exit a workflow depending on custom-defined inputs and outputs.

Autopilot for everyone

Autopilot for everyone and UiPath Assistant, which it is based on, allow you to reference workflows that may contain GenAI Activities. In this way, you can interact via an intelligent chat interface and invoke processes that have GenAI activities woven into them.

Document Understanding and Communications Mining

Document Understanding and Communications Mining offer extensive functionality and design-time control for dealing with both documents and communications intelligently. Both products bring standard machine learning and generative features to increase confidence and exercise greater control over any use case involving documents or communications.

You can combine Document Understanding and Communications Mining processes with other processes that use GenAI Activities. For example, you can ingest a document, digitize it, split it into different sections, extract different key pieces of information via Document Understanding, and then translate and summarize sections using GenAI Activities.

Why do I need to connect to UiPath GenAI Activities if they are a first-party service?

GenAI Activities use Integration Service to allow you direct access to UiPath-managed models via the UiPath AI Trust Layer. Integration Service requires a connection to execute authenticated API calls and to provide governance around which connections the robot uses at run time. GenAI Activities are based on Integration Service, thus require a connection. You are asked to create this connection only once.

What value do the GenAI Activities serve?

GenAI Activities give you access to managed subscriptions for popular foundation models from a variety of vendors. This means that you don’t need to commit to cloud service providers or manage infrastructure. Instead, you can focus on reaping the benefits of GenAI decisioning directly into your workflows and optimizing for the highest quality results.

GenAI Activities offer quality-tested, prompt-based activities with simple inputs and outputs to accomplish common RPA tasks like categorization, summarization, image comparison, etc.

All calls are routed through the UiPath AI Trust Layer, so you benefit from governance controls and audit capabilities. In the AI Trust Layer Automation Ops policy, you can choose to disable the GenAI Activities for a particular tenant. This ensures no user or robot is able to execute requests. You can also set up an Integration Service Automation Ops policy to hide the entire UiPath GenAI Activities for tenants, groups, or users in your organization.

Are GenAI Activities available on-premises?

No. You can use on-premises robots that have access to Orchestrator in the Cloud, but GenAI Activities, AI Trust Layer, and Context Grounding are currently available only in Automation Cloud.

Why can’t I see GenAI Activities in Studio Web or Studio Desktop?

If you don't see the GenAI Activities in Studio Desktop or Studio Web, make sure you:

• Don't have an Automation Ops policy that prevents you from using GenAI Activities.
• Use Studio Desktop version 2023.10 or newer.

Why are GenAI Activities failing in my automations?

Errors may occur for the following reasons:

Rate limiting: UiPath has instituted safeguards to ensure that GenAI Activities are not abused or misused. You may encounter a rate limit error, calculated based on the total tokens (roughly corresponding to word count) submitted within a certain time interval. If you encounter this error, reconfigure your workflow to include multiple GenAI calls with smaller context windows, separate GenAI Activity executions across multiple workflows, or add delays in the workflows. If you consider this error has occurred without reason or is blocking your adoption of GenAI Activities, contact your account management team.

Automation Ops policy: An AI Trust Layer Automation Ops policy can block your tenant from executing activities. Contact your administrator and make sure the policy is disabled in order to use the activities.

Context window exceeded: All models supported in the GenAI Activities have very explicit context windows, which means that only a text string of a certain length is supported. Check the LLM vendor documentation to see what the individual model limits are (expressed in tokens, but can easily be converted to word count). Consider the length of your inputs expected at run time. Another option is to use Context Grounding to analyze content that may exceed the context window for a single request.

Context Grounding errors: See the Context Grounding Frequently asked questions and ensure that you are following best practices, particularly regarding the supported file types.

Why isn’t a certain model available in Content Generation or Image Analysis?

UiPath gives you access to quality-proven, enterprise-ready models in a provider-agnostic manner. The Content Generation and Image Analysis activities allow you to select models that UiPath has licensed and validated for GenAI Activity use. If there is a model that you and your organization would like supported, reach out to your account team or contact us via the UiPath Community forum.

Why am I getting a suboptimal output from my GenAI activities?

For the Content Generation or Image Analysis activities, where you control the prompt directly, the most likely cause for a suboptimal output is that the prompt is not worded appropriately or critical variables/arguments are not correctly populated. Use a model to help you better word your prompt by explaining your use case, your expected output, and the prompt you’re using. Poorly worded or unclear prompts can result in unexpected results.

Ensure you test your workflows extensively, particularly so with GenAI Activities. Make sure to log messages and that variables and arguments are correctly displayed.

If you're using Context Grounding, make sure the index that you are referencing has been properly ingested. If you're using a file, make sure the file is in a supported format.

Your prompt may be cut off due to hitting a context window limit.

Why can’t I select a model for activities like Categorize, Named Entity Recognition, or Object Detection?

Activities such as Categorize, Named Entity Recognition, or Object Detection are designed to deliver the best output for a given task using a pre-tested prompt and model combination. All of these tasks can be accomplished using a custom prompt as well, if you require more control over model configuration.

What are the licensing requirements to use GenAI Activities?

GenAI Activities consume AI Units on a per-execution basis. Each GenAI Activity execution, regardless of token-size, is charged 1 AI Unit against the available allotment.

If you're using Context Grounding, irrespective of whether an existing index or file, the execution costs 2 AI Units. So each GenAI Activity execution (1 AI Unit) + Context Grounding (1 AI Unit) = 2 AI Units per execution where Context Grounding is in use.

The AI units consumption associated to the use of GenAI Activities is displayed in Automation Cloud Admin, on the Consumables tab of the organization- and tenant-level Licensing page.

Reach out to your account management team to acquire AI Units.

How do I use Context Grounding in GenAI Activities?

Context Grounding is currently available in the Content Generation activity. You can use an index (a repository of documents upon which the prompt is grounded) or upload a file for a just-in-time Context Grounding. It works by taking all of the data submitted in the prompt and searching through the index or file submitted for contextually relevant data to ground the prompt in.

Check out the Context Grounding Frequently asked questions and Best practices for the latest information.

How can I incorporate human-in-the-loop to validate GenAI Activity outputs?

With GenAI Activities you can save on development time and build what would otherwise be very complex, deterministic workflows. There is a trade-off, however, in that all GenAI outputs are inferences. You can optimize inferences with better instructions (prompting) provided to the model, higher quality inputs/variables, and grounding your queries and tasks in an appropriate context.

However, we recommend you institute validation steps whereby a robot loops in a human to ensure that the GenAI output was accurate. This doesn't have to happen for all workflows. In many cases you can incorporate a score or evaluation threshold that would warrant a human-in-the-loop (perhaps by using a GenAI activity for evaluation). For human validation, UiPath has a variety of options available:

• Action Apps – Action Apps allow you to build custom app interfaces using UiPath Apps that offer the full breadth of Apps capabilities (including Integration Service activities/API calls, etc.). You can configure where and how GenAI activity outputs are displayed, action options that would dictate the next steps that the robot takes, etc. There are Marketplace templates available to get you started. By inserting Create App Task and Wait for App Task and Resume activities, you can pause a robot in the middle of its operation until an assigned business owner or user validates the output. The robot remains in a waiting state until the Action App task is completed.

• Action Center Tasks – You can also use simple Action Center Tasks. Note that they don’t offer as many functionalities or customization options as Action Apps.

Can I use my own instance of a model through a cloud service like Azure OpenAI?

You can use the Integration Service third-party connectors to connect to popular cloud Services like Azure OpenAI, OpenAI, etc.

What kind of use cases can I accomplish using GenAI Activities?

Generative AI can make very complex tasks relatively simple for RPA developers to accomplish. With GenAI Activities, working with LLMs in the context of Studio, Studio Web, Apps, etc., becomes simple. Check out the following list to see what you can do with GenAI activities:

• Ticket categorization in Service Desk automation: Using the Categorize activity combined with popular Integration Service connectors like ServiceNow, Freshdesk etc., you can manage ticket categorization and routing.
• Meeting minutes summary: Using the Summarize Text activity, you can summarize meeting transcripts and then send the summary to team members using popular email connectors or messaging apps, like Slack or Microsoft Teams.
• Automated follow-up for unresolved customer issues: With the Generate Email activity, you can feed custom instructions and email body requirements directly and output an email in formatted HTML or text. You can then send this email using a Gmail or Microsoft Outlook 365 Send Email activity.
• Rewriting technical documentation for different audiences: Use the Rewrite activity to simplify technical documentation for non-technical users and teams.
• Internal policy translation for global teams: With the Translate activity you can translate policies and documents for one of the languages supported or a custom language.
• Detecting customer query language for multilingual support: Detect the language of incoming customer queries using the simple Detect Language activity for routing to the appropriate team.
• Contract drafting with internal data integration: Have legal teams reference a corpus of contracts already written to match style, tone and reference relevant context by using the Content Generation custom prompt activity, with the Existing Index option for Context Grounding.
• Product description with image integration for e-commerce: Use Image Analysis to generate a custom product description of an image for use in e-commerce apps or for inventory management systems.
• Inventory management and quality control: Detect defective items from a production line using custom criteria in the Detect Object activity.
• Detecting differences in product versions for compliance: Compare multiple product images to determine differences and violations of compliance using the Image Comparison activity.
• Document type classification in HR automation: Review scanned copies of documents and classify them according to custom criteria using the Image Classification activity.
• Customer feedback sentiment analysis for product improvement: Review customer feedback and reviews to build recommendations for product feedback using the Sentiment Analysis activity.
• Detecting duplicates in R&D proposals: Compare multiple proposals and determine where there are duplicates using the Semantic Similarity activity.