Business Central Managed AI Resources: GPT-4.1 Billing Rules, AL Migration and April 2027 GA

The short version​

Microsoft's default and recommended approach is to use the developer toolkit AI resources when extending Copilot in Business Central. Authentication goes through the SetManagedResourceAuthorization method in the AI module's AzureOpenAI codeunit. The idea is that publishers of Marketplace apps no longer ask each customer to buy and run an Azure OpenAI subscription. Microsoft-managed Azure OpenAI resources do that work, so customers don't have to independently get and manage subscriptions.

The roadmap metadata says:

• Preview availability: June 2025
• General availability: April 2027
• Cloud: Worldwide (Standard Multi-Tenant)
• Platform: Web

Treat the April 2027 date as an estimate. It has already moved. Earlier release-plan material listed general availability for Oct 2025, and a later 2026 wave 1 summary listed General Availability: Apr 2026. Anyone planning a product launch around this date should keep that history in mind.

What Microsoft says you gain​

Microsoft lists five benefits:

• Less operational overhead. There is no infrastructure, update, monitoring or troubleshooting work for a custom Azure OpenAI subscription.
• Simpler onboarding. You skip setting up your own subscription.
• Better reliability. Microsoft cites scaling, throttling, load balancing and fault tolerance.
• Built-in compliance and content protection. This includes governance, protection from harmful content, and data-residency handling.
• A smoother customer experience. Customers don't maintain separate AI configurations for each ISV solution.

These are Microsoft's claims, not independently measured guarantees. They are plausible for a shared, centrally operated service, but the documentation offers no performance or uptime figures.

Billing: one invoice, a different meter​

Customers who use AI resources across several apps built with the Business Central developer tools get a unified view. They can track consumption per Copilot feature and receive a consolidated invoice. They don't have to reconcile invoices from separate ISVs.

Billing follows Copilot Studio message billing and the "AI Tools" rates. The rate depends on the underlying model:

I haven't verified current dollar amounts, and they can change. Check Microsoft's billing rates page before quoting prices to customers.

This billing arrangement doesn't replace Marketplace monetization. Microsoft says the AI billing model is an additional option, and ISVs still monetize their own intellectual property through the Marketplace.

Limits to weigh before you commit​

The managed option is narrower than a subscription you run yourself:

• Models: Supported models are GPT-4.1 and GPT-4.1-mini. Microsoft says more are coming. A partner-owned subscription can use any Azure OpenAI model available to it.
• Embeddings: Embeddings aren't yet available on Business Central AI resources. Embeddings and fine-tuning are available with your own subscription, depending on the model.
• Environments: Managed resources work only in SaaS. Your own subscription works in SaaS and for container testing.
• Production only: Microsoft says the managed resources can only be used in customer production environments. It recommends your own subscription for prototyping, development, testing and support.
• Model migration: Partners must migrate to newer models and versions as Microsoft dictates, within the communicated SLA. If they don't, features can stop working in production once a model is deprecated.
• Preview status: Microsoft's page says preview features aren't meant for production use and may have restricted functionality.

Microsoft also asks developers who build on managed resources to enable app- or extension-level telemetry. That lets Microsoft tell them about model changes in time.

The migration obligation is the real trade-off. You hand over infrastructure work and, in return, accept Microsoft's model lifecycle. For a publisher with a stable prompt, that may be fine. For one whose output quality depends on a specific model version or a fine-tune, it is a reason to bring your own subscription.

The AL code: the roadmap example is outdated​

The roadmap text still shows the older call, which passed an account name and API key:

AzureOpenAI.SetManagedResourceAuthorization(Enum::"AOAI Model Type"::"Chat Completions",AzureOpenAIAccountName,AzureOpenAIApiKey,AOAIDeployments.GetGPT4oLatest());

Microsoft Learn now documents the current form:

AzureOpenAI.SetManagedResourceAuthorization(
    Enum::"AOAI Model Type"::"Chat Completions",
    AOAIDeployments.GetGPT41Latest());

The previous signature that required AzureOpenAIAccountName and AzureOpenAIApiKey is obsolete. Partners no longer need to supply their own Azure OpenAI subscription details when using Business Central AI resources. If your extension still uses the old signature, update it.

To use your own subscription, call SetAuthorization with an endpoint, deployment and API key:

AzureOpenAI.SetAuthorization(
    Enum::"AOAI Model Type"::"Chat Completions",
    GetEndpoint(),
    GetDeployment(),
    GetApiKey());

Branching by environment​

Microsoft's documented pattern uses your own subscription outside customer SaaS and managed resources inside it:

if not EnvironmentInformation.IsSaaSInfrastructure() then
    AzureOpenAI.SetAuthorization(Enum::"AOAI Model Type"::"Chat Completions",GetEndpoint(),GetDeployment(),GetApiKey())
else
    AzureOpenAI.SetManagedResourceAuthorization(Enum::"AOAI Model Type"::"Chat Completions", AOAIDeployments.GetGPT41Latest());

This is Microsoft's example, not a universal rule. Some partners or customers will need their own subscription in production, and Microsoft supports that. Microsoft also says an Azure subscription is required to use the AI development toolkit.

Billing type must match the resource you call​

Moving to managed resources is not just a code change. Starting with Business Central 2025 release wave 2, a partner registers each capability with one of three billing types:

• Microsoft Billed
• Custom Billed
• Not Billed

Microsoft says a capability is billed by Microsoft either when Microsoft delivers it or when the partner opts in to Business Central AI resources. At runtime, the declared billing type is checked against the Azure OpenAI resource the code actually calls. The documented combinations for partner-published capabilities are:

If a combination isn't allowed, the user gets an error saying AI resource usage isn't authorized with the chosen billing type. The sandbox exception exists so partners can test with their own subscription while declaring Microsoft Billed for production.

Handling overconsumption​

With managed resources, the AOAI Operation Response codeunit returned by the toolkit lets code distinguish failure modes. Microsoft documents these checks:

• IsSuccess(): the call worked and nothing more is needed.
• Status 402: the Microsoft Entra tenant has run out of AI resources and has no valid billing setup.
• Status 429: the user is sending requests too quickly. Microsoft notes this can point to automations or scheduled tasks the user isn't aware of. It clears after a few seconds.
• Status 503: the Business Central backend is temporarily unavailable. Microsoft calls this rare and typically transient.

In practice, 402 needs an admin or billing message, not a retry. 429 and 503 call for a short back-off. Showing users a clear, distinct message for each will cut support tickets.

What admins should know​

Customer administrators control Copilot and agent features from the Copilot & agent capabilities page. Registered capabilities appear there, and admins can deactivate a capability at any time. Microsoft says registered capabilities are active by default in the sense that customer administrators can switch them off. Admins can also grant or deny access per user, depending on the feature's permissions.

For third-party capabilities, Microsoft's admin guidance says to consult the feature's documentation or publisher for the required permissions. There is also an Allow data movement toggle. It appears only when a capability might process data in a different Azure geography from the Business Central environment. If it is switched off, affected capabilities become unavailable.

Microsoft's own comparison says AI model endpoints are available in limited Azure geographies. Customer environments elsewhere connect to endpoints in the US and are covered by the data governance controls available to administrators. ISVs with customers in regulated regions should read that section before choosing managed resources.

Practical checklist for ISVs​

1. Decide whether GPT-4.1 or GPT-4.1-mini meets your needs. If you need embeddings or fine-tuned models, plan for your own subscription.
2. Replace any call to the obsolete SetManagedResourceAuthorization signature that takes an account name and key.
3. Use your own Azure OpenAI subscription for development, testing and support. Use managed resources in customer production.
4. Set the capability's billing type to match the resource it calls. Test the sandbox and production paths separately.
5. Enable extension telemetry so you get notice of model changes, and budget time for migrations.
6. Handle 402, 429 and 503 explicitly.
7. Treat the preview label seriously until Microsoft announces general availability. The April 2027 target has already slipped from earlier plans.

Analysis​

This is a sensible design for the long tail of Business Central ISVs, where many small teams have no appetite for running Azure OpenAI deployments for every customer. A single metered invoice should also be easier to explain to finance teams.

The costs are less visible: a narrower model list, no embeddings, and a migration duty set on Microsoft's schedule. The billing-type rules add a further compliance step. The decision comes down to control versus convenience. These observations are my own analysis of Microsoft's documentation, not Microsoft statements.