Hi Business Central tech enthusiasts and consultants, as of the end of May 2026, there are important changes in the Microsoft Dynamics 365 Business Central ecosystem, especially regarding the AI model structure.
It’s no secret that Microsoft has been reshaping key AI components across its app ecosystem, aiming to better fit how consultants, devs, and uss actually work — without limiting what they can do — and, of course, getting the platform ready for everything that’s coming next.
Now, getting into it: Microsoft announced that between May and June 2026, Dynamics 365 Business Central agents will start moving to GPT-5.3-chat as the default model starting with version 28. This isn’t just a simple switch, though-it also brings model management capabilities into the agent design experience, plus new SDK methods to control model selection from AL. So it’s definitely more than just a “new engine” under the hood.
In practical terms, Microsoft isn’t just improving the baseline quality of agents; they’re also introducing a new level of governance, traceability, and technical control. That’s going to matter a lot for anyone building automations and agent-driven experiences in Business Central. Microsoft’s documentation even notes that agent behavior and accuracy might change after the update, so they recommend reevaluating your agents once the upgrade is deployed to your environment.
Context: from reactive Copilot to operational agents inside the ERP
To really get the scope of this, it helps to separate Copilot from agent capabilities in Business Central. Microsoft describes Copilot as a collaborative assistant that helps users get things done, while agents are more like “autonomous workers” that can run processes with minimal human involvement — while still keeping things transparent and allowing for human oversight when needed. In other words, Copilot responds; an agent can observe, decide, and act within defined limits.
This distinction matters because model selection is no longer just a behind-the-scenes AI detail — it directly impacts how agents understand instructions, prioritize actions, and solve tasks inside the ERP. Microsoft has been ramping up its investment in agentic capabilities in Business Central as part of wave 1 of the 2026 release, where we’re also seeing improvements such as agent design tools, AL support for agents, and new operational surfaces.
What exactly did Microsoft announce?
The update was published in the News.md file of the Microsoft/AL repo, and the achoed in the official Business Central docs confirm five key points. First, GPT-5.3-chat becomes the default model for agents. Second, admins will be able to see which models each agent is using in the UI and choose them directly from the agent settings page. Third, AL-based agents will have new methods for selecting models programmatically. Fourth, the rollout of the model and the UI/methods don’t necessarily happen at the same time. And fifth, environments in the UK, India, and Australia will get this update later than other regions.
From a technical standpoint, this means an organization might receive the new default model first and only later be able to manage it explicitly. Microsoft notes that during this transition, Copilot and agents will continue to run as usual — but that doesn’t remove the need to validate results, especially in scenarios where the agent is involved in the sensitive business operations.
Why does this change matter more than it seems?
The “AI models for agents (preview)” docs make it pretty clear: the model directly affects an agent’s behavior, performance, and outputs. On top of that, Business Central can expose multiple models at the same time, and each agent can either use a specific model or rely on automatic selection based on the environment’s default. That basically changes the conversation — this is no longer some “blackbox AI” inside the product, but a runtime layer with its own lifecycle, resolution rules, and architectural decisions.
Microsoft also adds an important detail: the default model can change over time, and agents running in automode might change behavior when that happens. Plus, models follow a defined lifecycle — Preview → Available → Obsolete → Retired — which introduces something developers already know well: versioning, compatibility, and deprecation patterns, now applied to AI.
The new model control layer: how Business Central decides which to use?
One of the most relevant pieces for architecture and devs is how model resolution works. According to Microsoft, when an agent runs a task, Business Central picks the model in this order:
- The model defined at the task level (if any)
- If not, the model is defined at the agent level.
- If neither exists, the environment’s default model
That hierarchy turns model selection into a multi-level architectural decision — not just a global toggle.
This opens up some interesting scenarios. For example, a company could use a general model for most agent flows but “pin” critical tasks — like document classification, accounting proposals, or actions on sales docs — to a specific model while validating accuracy, cost, or stability. From an engineering perspective, this enables features such as model pinning, task-level A/B testing, and safer, gradual rollouts.
Impact for admins: more visibility and governance
For Business Central admins, this isn’t just “hey, there’s a new model now”. The Compilot & agent capabilities page already lets you control which features are available, enable or disable specific capabilities, assign access via permissions, and apply governance rules such as cross-geo data movement. Microsoft also notes that, since version 25.0, these capabilities are enabled by default, including cross-geo data movement when applicable — though admins still retain control.
With this update, that same panel becomes even more important, since it’ll also be where admins can see which model an agent is using and switch it when the feature rolls out in their environment. On top of that, for custom agents in preview, Microsoft documents specific permissions, such as AGENT — ADMIN and AGENT — DIAGNOSTICS, which are key for designing, evaluating, and analyzing agent behavior.
Impact for AL developers: the model becomes part of agent engineering
From a dev perspective, this ties into a bigger move by Microsoft: turning Business Central into a platform where agents are designed, tested, and coded using native tools. The Coding agents in the AL (preview) docs explain that programmatic agents are built around interfaces such as IAgentFactory, IAgentMetadata, and IAgentTaskExecution, and that these capabilities are available for sandbox testing starting in 27.4, with production support from 28.1.
The other half of the picture is the IA stack itself. Microsoft documents the use of the System.AI module and the Azure OpenAI codeunit to register capabilities, manage authorization, configure chat-completion parameters, apply metaprompts, enforce guardrails, and collect usage telemetry. In other words, agents and Copilot features are not some “magic add-on” — they’re a structured dev surface with AL objects, explicit config, and clear responsibility boundaries.
Under that lens, the new SDK methods for model selection are a big deal. The model is no longer just backend infrastructure controlled by Microsoft — it becomes an operational variable you actively design around. Product teams and partners can now choose which model to use based on the agent type, task type, or the solution’s lifecycle stage.
What risks come with changing the default model?
Any default model change in an agent-based system brings the risk of functional drift. An agent might still “work” technically but change how it classifies inputs, summarizes context, suggests actions, or interprets ambiguous instructions. Microsoft is pretty explicit about this, recommending that anyone already designing or coding agents reevaluate them after the update, since bot behavior and accuracy can shift.
In Business Central, this is especially sensitive because agents operate directly inside the ERP — handling emails, orders, invoices, reconciliations, proposals, document suggestions, or workflows with human interaction. When the model changes, the practical boundary between “useful automation” and “risky automation” can shift, too. That’s why Microsoft strongly pushes thorough testing, especially in a sandbox, before going live.
Technical recommendations to get ready
Here are some recommendations for you. I am already applying this in BETABOX for its clients, so here we go.
First, establish a baseline before your environment gets GPT-5.3.chat. If you already have agents (prod or prototype), collect real examples of inputs, expected outputs, response times, edge cases, and the amount of human intervention needed. Microsoft provides guidance for evaluation and testing within the agent tooling and stresses full sandbox validation for preview features.
Second, segment by criticality. Since model resolution occurs at the task, agent, and environment levels, a good strategy is to leave low-risk flows in auto mode while explicitly pinning models for high-stakes tasks with stricter quality or compliance requirements.
Third, revisit prompts and instructions. Microsoft highlights that behavior depends on factors such as metaprompts, system messages, and token limits. A model change might reduce the need for heavy prompt scaffolding — or require stricter instructions. Fine-tuning won’t be optional anymore for serious implementations.
Fourth, audit permissions and observability. Agents run within Business Central’s security model and need explicit permissions for data and actions. Plus, Microsoft provides diagnostics and telemetry, which means you should treat each agent like a real software component — with logs, monitoring, and operational review — not a black-box automation.
Conclusion
The move to GPT-5.3-chat for Business Central agents isn’t just about better quality or performance — it marks a shift where the AI model itself becomes a manageable layer inside the ERP architecture. Microsoft is combining the default model switch with admin visibility, design-time model selection, and programmatic control via AL, pushing the platform toward a more mature, engineering-driven approach to agentic systems.
For customers, partners, and devs. The takeaway is clear: it’s no longer enough to just “turn on Copilot.” Going forward, you’ll need to design, evaluate, govern, and version the model layer behind your agents. And in an ERP like Business Central, that’s the difference between flashy automation and something the business can actually trust.