Power Apps can provide a controlled interface for an AI-assisted task, while Power Automate handles the server-side call and records the result. That does not make the app private by default. Privacy depends on the connections, provider terms, permissions, stored data and review design.
Build one bounded assistant, not a general company chatbot.
Verified on 24 August 2026 against Microsoft Learn and OpenAI documentation.
Choose a task with a checkable output
A good first task is draft a response from these approved facts. A poor first task is answer anything about the company.
Define:
- who may use the app
- which inputs they may submit
- which source data is allowed
- what the assistant returns
- who checks the result
- which actions are forbidden
If the app uses SharePoint for prompt templates, SharePoint permissions still govern who can read or edit that list. Hiding a control in Power Apps is not security.
Suggested architecture
Use four layers:
- Power Apps: collects the request and displays a draft
- Prompt store: holds approved task templates with an owner, version and status
- Power Automate: validates input, retrieves the approved template and calls the model through an approved connection
- Audit store: records safe metadata, review status and errors without copying unnecessary sensitive text
Keep provider keys out of canvas-app formulas and SharePoint. A custom connector or controlled API layer should own authentication.
Build the first version
Create a prompt list with fields such as Title, Purpose, Instruction, Owner, Status and ReviewedOn. Only approved rows should appear in the app.
In Power Apps, let the user select a task, enter the minimum necessary context and confirm the data classification. Call a Power Automate flow with the task ID and user input.
In the flow, retrieve the approved template by ID. Reject inactive templates, empty input and text above your chosen limit. Send the request to the model and return a simple object containing status, draft text and correlation ID.
Display the response as AI-generated draft. Require the user to review it before copying, saving or sending.
Permissions, licensing and DLP
Users need access to the app, its data sources and any connection arrangement used by the flow. Premium or custom connectors can require additional Power Apps or Power Automate licensing.
Power Platform data policies govern which connectors may exchange data in an environment. Classify external AI endpoints deliberately. A policy can block a risky connector combination, but it does not validate prompt content or generated claims.
Privacy is product-specific
Do not write the API is private as a blanket rule. OpenAI states that API inputs and outputs are not used for training by default, but documents default abuse-monitoring and endpoint-specific application-state retention. Azure-hosted or other providers have different contracts and configurations.
Map the full data path and check the exact service, endpoint, region, retention controls and subprocessor terms.
Test the failure boundary
Test ordinary input, missing prompt, edited or retired prompt, hostile instructions inside user text, unsupported content, provider refusal, rate limit, timeout and a user without SharePoint permission.
Check that the app never presents an error body as an answer and that previous users' responses cannot leak through variables or shared storage.
The completed prototype proves only the tested app and flow behaviour. It does not prove model accuracy, legal compliance or that a general assistant is safe for every team.
For practical Power Apps patterns with security and handover left intact, join the Power Apps Builders Space.