Business Central Copilot Extensions Get Microsoft-Managed AI Resources

TL;DR

Update 28.5, dated 2 September 2026, lists "Use AI resources for your Copilot extensions" at general availability. Public preview opened 6 June 2025, so it spent 15 months in preview.

Marketplace publishers can now call large language models through Microsoft-managed Azure OpenAI resources. The customer no longer needs their own Azure OpenAI subscription.

Microsoft calls this "the default and recommended approach", and points billing questions at the Copilot Studio rate card.

Read that rate card and the tiers are 1, 15 and 100 Copilot Credits per 10 responses. End to end that is a 100x spread for the same number of answers.

The consequence nobody has written down: Copilot Credit capacity is pooled across the tenant, and enforcement disables custom agents at 125% of prepaid capacity. Whose agent stops is not decided by whose agent overspent.

What shipped, precisely?

One line in a cumulative update, with fifteen months behind it.

Update 28.5 for Business Central 2026 release wave 1 lists two feature changes at general availability. One is Verifactu mode for Spain. The other is "Use AI resources for your Copilot extensions", and it is the one that changes how partner software gets built.

Microsoft's description: "publishers in their Marketplace apps use large language models through Microsoft-managed Azure OpenAI resources when they extend Copilot in Business Central. This approach eliminates the need for customers to independently get and manage Azure OpenAI subscriptions."

The dates are worth holding onto. Public preview 6 June 2025. General availability September 2026. Fifteen months is a long preview for a feature whose entire purpose is to remove a setup step, and it suggests the hard part was never the API call.

What does an ISV stop doing?

Roughly everything that is not the actual product.

Managed Business Central AI resources. Bring your own Azure OpenAI

Runs the infrastructure. Microsoft. The publisher

Customer needs an Azure OpenAI subscription. No. Yes

Model choice. What Business Central provides. Any, including fine-tuned models

Scaling, throttling, load balancing, fault tolerance. Microsoft. The publisher

Data residency compliance. Built in. The publisher's problem

Who bills the customer for AI. Microsoft, on one invoice. The publisher

AL authorisation call. SetManagedResourceAuthorization. SetAuthorization

There is a small developer detail with a real cost attached. The managed path previously still made you "provide your own Azure OpenAI subscription details, but the authentication process doesn't use them." Wave 1 adds a SetManagedResources signature that drops the pretence: "There's now a new method signature for SetManagedResources that doesn't require you to provide your subscription details anymore."

If you are shipping a Business Central Copilot extension, that is the line to change this month.

Who pays, and how much?

This is where the release note gets quiet and points elsewhere. For costs it says only: see AI tools at Message scenarios, and links to the Copilot Studio billing page.

So read that page. The published rates:

Agent feature. Copilot Credits

Classic answer. 1

Generative answer. 2

Agent action. 5

Tenant graph grounding. 10

Text and generative AI tools, basic, per 10 responses. 1

Text and generative AI tools, standard, per 10 responses. 15

Text and generative AI tools, premium, per 10 responses. 100

Content processing tools, per page. 8

Two things fall straight out of that table.

The tier spread is 100x. The same ten answers cost 1 credit on basic and 100 on premium. Per thousand tokens the ratio is identical: 0.1, 1.5 and 10 credits. Nothing about the user's question changes. The model behind the prompt does.

A reasoning model is billed twice. Microsoft is explicit: total cost equals the feature rate plus the premium tools rate per 1,000 tokens. So a generative answer over a 3,000-token exchange, using a reasoning model, works out as:

generative answer 2 credits

premium reasoning tokens 3 x 10 = 30 credits

32 credits

The same answer without reasoning is 2 credits. Turning on deep reasoning multiplies that single interaction by 16. At a 5,000-token exchange it is 2 + 50 = 52 credits, a factor of 26.

The part that is not in either document

Put the two documents side by side and something appears that neither states on its own.

Copilot Credit capacity is pooled. Microsoft's own words: "When you purchase a Copilot Studio license, you get a specific number of billed Copilot Credits. You pool this capacity across the entire tenant." Environments with no explicit allocation draw from that shared pool.

Enforcement is also tenant-wide. At 125% of prepaid capacity, "Custom agents are disabled." Users then see one of two messages: "There is a billing issue." or "This agent is currently unavailable. It has reached its usage limit."

Now run the numbers on Microsoft's own example capacity of 25,000 credits. Enforcement fires at 125%, which is 31,250 credits.

Extension on the premium reasoning path

200 answers a day x 32 credits = 6,400 credits a day

31,250 / 6,400 = 4.9 days to enforcement

The same extension on the basic path

200 answers a day x 2 credits = 400 credits a day

31,250 / 400 = 78 days, so it never trips in a month

One extension on the premium path can walk a 25,000-credit tenant into enforcement in under 5 days. The identical extension on the basic path cannot trip it at all inside a billing month. Same feature, same usage, a 16x difference in how long the tenant survives.

And when it trips, it disables custom agents, not the offending one. A finance customer running three partner extensions can have the quiet two stopped by the chatty one.

One caveat, stated plainly. Microsoft's release note routes Business Central publishers to the Copilot Studio rate card for costs, and that card describes the pooling and the 125% enforcement. What no single Microsoft sentence says is that Business Central extension consumption draws from the same tenant pool as Copilot Studio agents. The documents are linked, not welded. Ask your partner to confirm it in writing before you budget on either answer. That question is the most valuable thing in this article, and we could not close it from public documentation.

What should a partner or customer do about it?

Allocate capacity per environment rather than leaving it in the pool. Microsoft documents the control: Power Platform admin center, Licensing, Copilot Studio, Manage Copilot credits. An environment with its own allocation is not enforced while it still has headroom, even when the tenant is in overage.

Cap the agents individually. Licensing, Copilot Studio, Manage Agents sets a monthly consumption limit per agent. That converts a tenant-wide outage into one agent slowing down.

Consider pay-as-you-go for the noisy environment. In Microsoft's worked example the pay-as-you-go environment is unaffected when the tenant is in overage, because overage bills to Azure rather than blocking.

Ask every ISV which tier its prompts run on. Basic, standard or premium is a 100x question and it is almost never on a datasheet. It should be a line in the app's documentation, and if it is not, it should be a line in your contract.

Watch the reasoning switch. The 16x is not a pricing error. It is the documented cost of a feature that reads as a quality setting and behaves as a spend setting.

How does this sit against the rest of the platform?

It is the third change this quarter pointing the same way, which is what makes it worth an article rather than a release note.

Microsoft is retiring Release Plans in favour of an always-on roadmap, which moves disclosure from a document to a feed. It has been wiring agent governance into MCP actions, which moves control from configuration to policy. Now it takes the AI infrastructure off the publisher and puts the consumption on the customer's meter.

Each step removes a thing a partner had to run. Each step also removes a thing a partner could see. The trade is real in both directions, and the second half is the one that does not appear in the announcement.

FAQ

Does the customer still need an Azure OpenAI subscription?

Not on the managed path. Microsoft states the approach "eliminates the need for customers to independently get and manage Azure OpenAI subscriptions."

Can a publisher still bring their own model?

Yes. Microsoft supports it for specific or fine-tuned models the managed resources do not offer. In that case "the publisher is responsible for deploying, maintaining, scaling, and securing the service" and for billing the customer.

Does this replace Marketplace monetisation?

No. Microsoft says it "isn't designed to replace the existing Marketplace monetization pathway" and that ISVs continue to monetise their intellectual property through the Marketplace.

What happens to an agent when capacity runs out?

At 125% of prepaid capacity, custom agents are disabled. An in-progress conversation finishes, then the agent stops answering. Agent flows are handled separately: new runs are blocked while the agent keeps working for everything else.

Is this on-premises?

No. It concerns Marketplace apps extending Copilot in Business Central online.

The last mile

Managed AI resources are a genuine simplification, and the fifteen-month preview shows how much of the difficulty sat in billing rather than in inference.

But simplification moves a problem, it does not delete one. The publisher stops running an Azure subscription and the customer starts carrying a meter they did not previously read, with a shut-off at 125% that does not distinguish between the agent that overspent and the agent that did not. Getting value out of that setup is less about the model than about knowing which decisions justify a premium-tier prompt and which ones were always a two-credit lookup. Working that out inside the system a business already runs on, with the arithmetic visible rather than assumed, is the layer Cognilium builds in.