Microsoft has published a four-step guide for building a knowledge-grounded voice and avatar agent in Foundry Agent Service. The Community Hub guide, dated October 30, 2026, explains how to connect a Foundry IQ knowledge base, select Personal Voice and a custom photo avatar, then start a live conversation in the portal.
The playground path requires no local application code. Microsoft handles the interactive experience there, so builders can try the combination without writing their own speech or avatar interface.
This is an implementation guide for an existing capability, not a new model or general-availability announcement. Microsoft’s Voice Agent repository identifies the service as public preview, with availability dependent on the project’s models, voices, region and access.
The practical value is the connection between three separately prepared resources: what the agent knows, how it sounds and whose likeness it presents. Getting those resources ready is the substantial prerequisite behind Microsoft’s quick portal workflow.
Four Steps in the Foundry Portal
Microsoft’s guide starts with prepared assets in a supported Foundry project. Builders need a working Foundry IQ knowledge base and, for the personalized version, an authorized Personal Voice and custom photo avatar. Platform-provided voices and avatars are also available as a simpler starting point.
The accompanying setup README specifies a Foundry project, not a hub-based project. It also recommends testing the knowledge base before attaching it to the agent.
Create a Voice Agent
In the Foundry portal, open the project and navigate to Agents → Build an agent. Enter a name, choose Voice as the interaction mode, select the appropriate use case, and create the agent with its playground.
Review the selected model and default instructions. Those settings determine the conversational behavior; selecting a voice and face alone leaves the agent’s purpose underspecified.
Connect the Knowledge Base
Expand Knowledge in the agent configuration, select Add, and choose the prepared Foundry IQ resource from the available knowledge bases.
Give the agent a narrow purpose, such as explaining product features from approved documentation. Microsoft recommends concise spoken responses, using conversation context for follow-up questions, and explicitly acknowledging when the connected materials lack an answer.
That last instruction matters for a branded assistant. A confident delivery can make an unsupported answer sound authoritative, so the knowledge boundary belongs in the agent’s instructions from the start.
Choose the Voice and Avatar
Under Voice, choose the prepared Personal Voice through Custom → Personal. Enable the avatar and select the prepared custom photo asset.
These choices control presentation. They do not expand the knowledge base or grant permission to use someone’s identity. Authorization and consent must already be in place.
Builders can also select platform-provided assets while preparing personalized ones. The sample README defaults to Andrew Dragon HD and the Standard Harry / Business avatar, avoiding personal voice recordings and likeness uploads for that initial configuration.
Save and Start a Live Session
Save the configuration and start a conversation in the playground. Try a factual question supported by the documents, a follow-up that depends on the previous answer, and a question outside the knowledge base.
Microsoft describes the four-step assembly process as taking minutes once the assets are ready. That timing excludes document preparation, permission changes, consent collection and asset processing.
Grounding Needs a Separate Validation Pass
Foundry IQ supplies reusable knowledge bases. In the sample’s portal workflow, builders upload authorized documents, wait for processing, and inspect the results in the Knowledge playground. The portal handles parsing, chunking and indexing; the walkthrough does not require a separately uploaded Storage copy or a manually defined Search index.
The README describes a minimal extractive configuration using Retrieval reasoning effort: Minimal and Output mode: Extractive data. Builders must still satisfy any model requirements shown for their chosen settings.
Testing retrieval before adding personalized presentation is a useful separation of concerns. If the knowledge playground returns poor references, changing the voice or avatar will leave that problem intact. I’d start with platform-provided assets, validate retrieval and refusal behavior, then introduce the personalized identity.
Microsoft’s Community Hub demonstration uses a Theodore Roosevelt knowledge base. It shows a conservation question, a contextual follow-up and an unrelated stock-price question that the agent declines to answer from the available material. The demonstration uses Andrew Dragon HD and the preset Marcel photo avatar. Roosevelt is the subject of the documents, not the identity depicted by the assistant.
The sample README also reports 219 passes and 18 failures across 237 knowledge-covered questions, a 92.41% pass rate. This is Microsoft’s reported knowledge evaluation, not an independent measurement. Its stated scope covers questions represented in the knowledge base, so it supplies no pass rate for out-of-scope questions or complete voice-and-avatar reliability. The README provides the evaluation context.
That distinction makes the unrelated-question check consequential: answer accuracy within a corpus and restraint outside it need separate attention.
Consent Applies to Both Voice and Likeness
Microsoft’s guide requires builders to use their own voice and likeness or obtain explicit authorization and the required consent from the person represented.
The preparation requirements differ by asset:
- Knowledge: Use authorized documents and obtain approval for the upload destination, costs and permissions.
- Personal Voice: Prepare a short, clean voice recording and the speaker’s recorded consent statement.
- Custom photo avatar: Prepare an authorized photo and a consent video from the person pictured.
The setup documentation also warns against committing credentials, consent media, real resource identifiers or local output to Git.
Consent and technical availability are separate gates. Having permission to represent someone does not guarantee that Personal Voice or a custom photo avatar is available in a particular resource or region. Conversely, a selectable feature provides no authorization to use a person’s identity.
Microsoft says prepared identity assets can be reused across agents, while documents can be refreshed independently. That separation is operationally useful: updating a product manual does not require recreating the voice or avatar. Teams should still check that each new use fits the authorization they obtained.
Portal and Python Serve Different Setup Needs
Frequently Asked Questions
3 questions
1Does the Foundry Playground Require Local Code?
No. The documented portal path requires no local application code, browser proxy, microphone library or hosted container. Python is optional.
