AI
AI-Assisted Data Modeling for App Makers
Three features reduce manual work for makers and end users. Plan Designer (now GA) turns a data dictionary into a Dataverse schema without writing column definitions. In the Plan Designer data model demo (03:12), a retail review dictionary produced six tables, which a chat prompt then collapsed to four. A Prompt column applies generative AI to any table column, running tasks like PII redaction and sentiment scoring on each saved row.
For end users, the Data Exploration Agent adds a natural language search bar to model-driven apps. Type a filter condition; the grid updates immediately.
Dataverse as Agent Knowledge
“You can think of agent complexity as a spectrum.” — Julie Koesmarno
Dataverse Knowledge indexes both files and relational tables so agents can answer questions from either. Files (PDFs, documents) go through content extraction, chunking, and dual-index storage covering text and vector retrieval. Structured tables get a semantic index that resolves abbreviations, for example “WA” to “Washington,” before generating SQL at runtime. In the Returns agent knowledge demo (15:56), an employee agent checked return policy PDFs, queried tabular order data to surface a $239.95 total, applied the 50% restocking fee, and read a multiline reason column, all in one conversation.
Extending Agents with Reusable Business Logic
“Functions allow us to extend our agents with actions.” — Sean Watson
Dataverse Functions package business logic, inputs, and outputs into a reusable unit that both apps and agents can call. A maker defines the function once in the Maker Portal, referencing whatever tables and validation rules apply. Copilot Studio then surfaces it as an action with a matching input schema. In the Functions return approval demo (20:38), the agent received two of three required inputs, inferred the action, and prompted for the missing approval reason before executing the Approve Return Request function and updating the record.
Autonomous Agents with Human Oversight
Autonomous agents need a way to pause and hand off. The Document Processor Template shows the pattern: a Copilot Studio agent watches an inbox, extracts invoice fields into Dataverse, and flags uncertain records for human review. In the Document Processor end-to-end demo (23:39), a formatting error triggered a review notification; one correction moved the record to its final success state.
Two production deployments follow the same design. Velrada’s ShiftLens logs nurse absences, finds qualified replacements, and sends a one-click approval to the team lead. Litera monitors lawyer inboxes for court documents, calculates hearing dates, and stores tasks in Dataverse.
Notable Quotes
You can think of agent complexity as a spectrum. Julie Koesmarno · ▶ 7:14
Functions allow us to extend our agents with actions. Sean Watson · ▶ 18:58
The legal industry is facing a data explosion. Mike Walker · ▶ 29:01
Key Takeaways
- Plan Designer converts a data dictionary into a working Dataverse schema in one session.
- Dataverse Knowledge lets agents query structured tables and PDFs using natural language.
- Dataverse Functions share business logic between apps and agents without duplication.