21.6 GenAI Activities: Generative AI Inside Workflows

21.6 GenAI Activities: Generative AI Inside Workflows

Core Concept: GenAI Activities let a workflow use large language models for common tasks, such as summarizing, classifying, translating, extracting entities, or writing text, without building prompts and API calls from scratch. They run through UiPath's AI Trust Layer, where administrators govern models and data handling.

---

The Activities

---

Designing with GenAI Activities

Choose the narrowest activity

A dedicated activity such as Categorize or Summarize Text gives more predictable results than a general Content Generation prompt. Use Content Generation when no dedicated activity fits.

Constrain outputs

• For Categorize, provide a closed list of categories and handle unexpected answers.
• For Content Generation, ask for a specific format, such as JSON with named fields, and validate the result before using it.

Ground answers in company data

Answers about company policies or products should be grounded with a context grounding index, so the model uses retrieved passages rather than general knowledge.

Keep people in the loop

For decisions with consequences, such as refunds or legal replies, route the AI result to a person, for example with an Action Center task, before acting.

Protect data

Use PII Filtering before sending customer text to other systems or logs, and follow the tenant's AI Trust Layer policies on which data may be sent to models.

---

A Typical Pattern: Email Triage

1. Detect Language, then Translate non-English emails into English for the team.
2. Categorize into Complaint, Question, Order Change, or Other.
3. Sentiment Analysis flags strongly negative emails for priority.
4. Named Entity Recognition extracts the order number and customer name.
5. Context Grounding Search finds the relevant policy passage, and Generate Email drafts a reply using it.
6. The draft goes to an agent for review through an Action Center task.

Each step uses a focused activity, and the workflow validates each output before the next step.

---

GenAI Activities Versus Other AI Options

---

Testing and Monitoring

• Build a test set of real inputs with expected outputs, including hard cases, and re-run it when prompts or models change.
• Log the category or decision and a short reason, not full sensitive texts.
• Handle failures: wrap calls in Try Catch or Retry Scope, and decide what happens when the service is unavailable.
• Watch for drift: if categories start to go wrong, review prompts and category descriptions.

---

Exam-Style Scenarios

Scenario 1: Support managers want a short summary of each long complaint in the ticket. Use Summarize Text, and store the summary in the ticket rather than the full text in logs.

Scenario 2: A process must answer employee questions using the HR handbook. Use Context Grounding Search on an index of the handbook, then Content Generation grounded on the retrieved passages, and send uncertain answers to an HR specialist.

Scenario 3: Customer emails arrive in five languages, but the team reads only English. Use Detect Language and Translate before categorizing, and keep the original text for the reply.

Scenario 4: A legal team needs names, dates, and company names from contracts listed for review. Use Named Entity Recognition, and route the list to a reviewer before it is used in any decision.

---

Key Terms

• Context grounding index: a searchable store of company documents used to ground model answers.
• DeepRAG: an analysis over documents whose result is retrieved with Get DeepRAG Analysis by ID.
• AI Trust Layer: the governance layer through which UiPath AI features reach models, with policies for data and model use.

---

Common Traps

• Using Content Generation for tasks a dedicated activity handles better.
• Trusting generated output without validation, especially numbers and IDs.
• Sending personal data without PII Filtering where policy requires it.
• Forgetting to refresh an index with Update Context Grounding Index after the source documents change.