Real Value Acceleration Program home

What can you build?

Explore the kinds of AI applications you can bring to discovery, including capabilities added to existing systems.

What the creator can design

Tell AI Agent Creator what you want a new or existing application to do. It helps you define the user journey and agree on a Microsoft-cloud architecture before implementation starts.

The workflow covers platform-managed agents and custom applications. It can also conclude that a bounded model call or an explicit workflow fits better than an agent.

Here, supported means covered by the design workflow. You still need to build the application; the plugin supplies no ready-made implementations or certified architectures. During discovery, the creator verifies the connections your design needs rather than assuming the services work together. The examples below illustrate possible uses. They are not customer results or deployment templates.

Application families

The Microsoft products shown below are examples, not required services or approved architectures. Discovery must verify product fit and any connections your design needs. Some examples cover different parts of a solution and can work together. They are not interchangeable. Check the selected features' availability and release status. Product icons use Microsoft artwork where available. Neutral document and code symbols are illustrations, not product logos.

Knowledge assistants

Help a user find and understand information they are allowed to access. For example, an operations assistant could explain a procedure and cite the current source, then investigate a follow-up question using other permitted evidence.

Name who owns each source and how it stays current. The design must explain how user permissions limit retrieval and what happens when evidence is absent or contradictory. A fixed lookup may need retrieval without an agent loop.

Example options
Azure AI SearchSearch and retrieval for an application-owned assistant.
Microsoft Copilot StudioA platform-managed agent grounded in configured knowledge sources.
Answer with authorized evidence The assistant searches sources the user may access. It cites evidence in its answer and reports missing or conflicting evidence. Ask a question Search permitted sourcesApply the user's permissions Answer with citations
If evidence is missing or conflicting, say so. Retrieval alone does not require an agent loop.

Agents that use business tools

Investigate a task and prepare or carry out operations within agreed limits. A maintenance assistant could gather request details and propose a work order for a manager to approve.

Tool access is a separate decision from knowledge access. Specify which actions are read-only, which require approval, and how the system handles retries without duplicating a business operation. The system must check authorization when it performs each operation.

Example options
Azure Logic AppsConnector-based business workflows.
Foundry Agent ServiceAn agent that selects configured tools. Enforce business authorization separately.
Propose an action before approval The assistant investigates a maintenance request and proposes a work order. A manager approves or rejects it before any operation. The system must authorize the operation before acting. Receive a request Investigate with allowed tools Propose a work order Manager reviews the proposalApprove or reject
The system must authorize each operation and prevent retries from repeating it. Proposing an action does not authorize it.

Document and media processing

Turn incoming material into structured results. For example, classify a document bundle, extract fields, and send uncertain results to a reviewer.

Agree the output schema and the review threshold. Define how a reviewer corrects an error and how a corrected result reaches the downstream system. This may be a fixed processing pipeline; a conversational interface is optional.

Example options
Azure Document IntelligencePrebuilt or custom extraction models for known document types.
Azure Content UnderstandingAnalyzers for unstructured documents or mixed-media inputs.
Turn a document into a reviewed record The pipeline classifies a document and extracts its fields. A reviewer checks uncertain fields before the pipeline publishes the record. Receive a document Classify it and extract fields Review uncertain fieldsApply the agreed threshold Publish the structured record
The output must match the agreed schema. Send reviewer corrections to the downstream system.

Voice and audiovisual experiences

Use speech or an audiovisual presentation when the user's task requires it. An illustrative voice assistant could guide a worker through a permitted procedure, with confirmation before a consequential action.

Decide how users interrupt, correct, or leave the interaction. Identify privacy and accessibility requirements, and verify the selected channel's capabilities. An avatar or generated video alone is no reason to add an agent.

Example options
Azure SpeechCompose speech recognition and synthesis with your application logic.
Voice Live APIA speech-to-speech API within Azure Speech, rather than a separate product.
A voice response is separate from an action The assistant interprets speech and responds by voice. Consequential actions require a separate confirmation step. Listen to the user Interpret the request Respond by voice Confirm before taking action
The dashed path requires confirmation before consequential actions. Let users interrupt the assistant or correct what it understood.

Background workflows

Respond to an event or process a job without a new chat interface. An incoming service request could trigger evidence gathering and then wait for an owner to approve the proposed next step.

Identify who owns job state and how work resumes after a failure or a long wait. Keep fixed branches explicit. If the system investigates as it runs, define what it may investigate and when it must stop.

Example options
Azure Logic AppsConnector-based workflows with explicit steps.
Azure Functions with Durable FunctionsCode-defined workflows with persisted state and recovery.
Keep job state across an approval wait An event starts a job that gathers evidence. The job saves its state while it waits for approval. It resumes only after approval. An event starts a job Gather evidence Wait for the owner's approvalSave the job's state Resume after approval
The job can wait without an open chat. Stop the proposed action if the owner rejects it. Define how the job recovers from failures.

Coordinated specialist agents

Split responsibilities when one agent cannot meet the requirement reliably. For example, an investigator and a separately authorized action agent could have different tools and access boundaries.

Explain why separate agents are needed and define their handoff contracts, including how agents handle conflicting results and shared state. Define what happens when one agent fails. Coordination adds overhead. Microsoft's orchestration guidance recommends the lowest complexity that meets the requirement.

Example options
Foundry Agent ServiceA managed agent runtime. Verify support for the selected coordination pattern.
Microsoft Agent FrameworkCode-defined orchestration. Choose hosting separately, including Foundry when appropriate.
Coordinate specialists with separate permissions A coordinator assigns work to specialists with separate permissions. It checks their results before submitting a proposal for approval. Coordinate the task InvestigatorRead evidence Action agentPropose work Check and submit the proposal
This example gives each specialist separate tools and permissions. The coordinator must resolve conflicting results. Consequential actions need approval.

Capabilities you can combine

These families can overlap. A document workflow may feed a knowledge assistant; an investigation may finish with an approval-controlled operation.

Persistent memory and personalization are additional choices. Define what the system may remember and who owns that context. Set rules for expiry and deletion. Multiple knowledge sources need their own permission and provenance boundaries.

Add these capabilities when the journey needs them. You can start without multiple agents or persistent memory.

Where this workflow stops

The creator records the architecture and requirements, with an implementation plan for a separate delivery session. You confirm discovery before it prepares the complete package, then approve that package before handoff.

Use specialist workflows for standalone SDK questions, incident troubleshooting, or live agent operations. Architecture approval grants no permission to implement or deploy.

Describe your first journey

Tell the creator who will use the application and what their first task is. You can leave the platform open or state a constraint:

/ai-agent-creator Design a maintenance assistant for property managers. Help them investigate
resident requests and prepare a work order for approval. Reuse our existing
identity and service system. Compare an agent with a simpler workflow before
recommending the architecture.

Set up your application workspace, then bring your own idea. If you are extending a system, read Add AI to an existing application.