Edit

Add AI capabilities to a line-of-business WinUI app

AI can add summarization, extraction, classification, search, and assistance to a business workflow. Choose an architecture based on data policy, model quality, supported hardware, connectivity, latency, and operating cost.

A WinUI 3 support app with a ticket list, customer request details, and an option to analyze the ticket with local AI.

Prerequisites

  • Complete the WinUI setup instructions and create a WinUI 3 app.
  • Review the Windows AI APIs overview and the requirements for the API version you plan to use.
  • For cloud AI, configure an approved Azure resource, identity, network path, and data-governance policy.

WinUI 3 is delivered as part of the Windows App SDK; it isn't a separate UI framework that you add to a WinUI 3 project afterward.

Choose an approach

Requirement Approach
Supported built-in on-device language capability Windows AI language-model APIs
A custom model that must run locally ONNX Runtime with an applicable execution provider
A larger general-purpose model, retrieval, or centralized governance Azure AI service
OCR, speech, or other Windows capability The applicable Windows platform API

Don't select an on-device model from the processor name alone. Verify the current API's OS, Windows App SDK, model, and hardware requirements. Supported APIs can use different NPU or GPU configurations.

Use the Windows AI language model

The Windows AI language-model APIs expose readiness states so that an app can detect whether a model is available and prepare it before generation.

using Microsoft.Windows.AI;
using Microsoft.Windows.AI.Text;

if (LanguageModel.GetReadyState() == AIFeatureReadyState.NotReady)
{
    await LanguageModel.EnsureReadyAsync();
}

if (LanguageModel.GetReadyState() == AIFeatureReadyState.Ready)
{
    using LanguageModel model = await LanguageModel.CreateAsync();
    LanguageModelResponseResult result =
        await model.GenerateResponseAsync(prompt);

    if (result.Status == LanguageModelResponseStatus.Complete)
    {
        string response = result.Text;
    }
}

Handle every non-ready and incomplete state that the version you target can return. Show a clear status, disable unavailable commands, and provide a non-AI or approved cloud fallback where the workflow requires one.

Important

Microsoft has announced that Aion Instruct will replace Phi Silica. The Phi Silica transition guidance provides the rollout timeline and states that Aion Instruct doesn't require a Limited Access Feature token. Until Aion is available through the supported API and servicing channel that your app targets, treat it as a preview and continue to check model readiness at run time.

Treat Phi Silica as a transitional API, not as a model that is present on every supported Windows device. Check readiness at run time and provide a graceful fallback.

Use a cloud AI service

For a cloud model:

  • Authenticate with an identity flow approved for your organization.
  • Keep service credentials out of the client app.
  • Apply timeouts, cancellation, bounded retries, and rate-limit handling.
  • Tell users when data leaves the device.
  • Apply the organization's retention, regional-processing, and content-safety requirements.
  • Monitor quality and cost using production-like request sizes.

Prefer a service owned by your organization when the desktop client would otherwise need a privileged secret.

Run a custom model with ONNX Runtime

ONNX Runtime can run a packaged model on the device:

using var session = new InferenceSession("model.onnx");
using IDisposableReadOnlyCollection results =
    session.Run(inputs);

Choose and test an execution provider for the target hardware. Include model files in your servicing, licensing, integrity, and update plans. Dispose sessions and result objects.

Design the user experience

  1. Keep the UI responsive and support cancellation for longer operations.
  2. Show progress and identify AI-generated output.
  3. Let users review consequential suggestions before applying them.
  4. Preserve source data and make corrections possible.
  5. Record enough context for diagnostics without logging sensitive prompts or output.
  6. Evaluate the feature with representative business data and failure cases.

Related content

  • Last updated on