What's New: Prepare Your Copilot Extension - Practical Considerations (2025 release wave 1)

Overview

The session explains that partners can upload a Copilot extension to a Business Central sandbox and run it in the Microsoft cloud on Microsoft AI resources, without managing LLM models, deployment or availability. Partners now use the same framework and infrastructure as Microsoft. GPT-4o and GPT-4o mini are available through an abstraction layer. Models such as Deep Seek R1 and other reasoning models are not yet available in Business Central, because capacity is needed in all Azure regions first.

The rest covers practical considerations. Model versions change behavior and are not necessarily backward compatible, so test automation should start from day zero. Fabrications are called the number one priority. The video shows validating LLM responses with a marketing text example (accuracy raised from 6%) and a documentation chat grounded in Microsoft Learn. It also covers tools defined in AL for grounding, with a recommended limit of around five.

Key points

• Partners can upload a Copilot extension to a latest sandbox environment and run it in the Microsoft cloud on Microsoft AI resources, with no LLM deployment or availability to manage.
• GPT-4o and GPT-4o mini are available through an Azure OpenAI abstraction layer, so partners do not manage model versions themselves.
• Deep Seek R1 and other reasoning models are not yet in Business Central. They come only after capacity exists in all Azure regions where customers are, and no timeline was given.
• Model versions can change behavior overnight and are not necessarily backward compatible. Write test automation from day zero to catch prompt regressions, which matters more than prompt styling.
• Fabrications come from models predicting likely text. Validate LLM responses, for example with GPT-4o mini as a cheaper validator, and do not show them to users unless they are verified as grounded.
• Validation added to the marketing text feature raised accuracy from 6% to very high numbers. The documentation chat checks grounding against Microsoft Learn articles, which may add about 200 ms of latency.
• Define tools in AL to pull real Business Central data into responses. Keep to around five tools, since too many reduce reliability.

Features

AL objects mentioned

As heard in the captions. A name that matches one object page by exact type and name links to it; the others stay as named.

• other “Sales Order Page” at 12:42
• other “Production Order” at 11:20
• other “Credit Memo” at 11:20
• other “Sales Invoice” at 12:01
• other “Purchase Order” at 12:01
• other “code unit” at 22:41

Quotes

• 0:05 “in this release we allow you to bring your capalot extension and finally run it in the Microsoft cloud based on our Microsoft II”
• 1:26 “finally you and us have a same framework and infrastructure how do we build our compal experiences”
• 2:18 “as a part of our def tools today we give you access to GPT for or and GPT for all mini”
• 3:39 “before we as a business Central can give you access to those models we need to make sure that all the world where we”
• 8:48 “new versions can be faster or sometimes cheaper and so on so forth so and because every model is not better is different sometimes”
• 10:01 “it’s much much more important for practical considerations to make sure your prompts will continue to work as we’re going to update models”
• 12:01 “large language models just copy those uh type of response if you will because it’s trying to be helpful but it definitely lead to”
• 14:03 “large language models always try to predict next words or talking and it always try to be helpful and finish the sentence and instead”
• 14:43 “always validate l l m response”
• 15:44 “whenever we never surface any llm response from a system to the user unless we verify they’re grounded and they true”
• 16:04 “you might use 40 mini as a cheaper faster model to validate its responses and this is like a very simple pattern”
• 17:45 “having that small technique just to validate llm for this particular feature allow us to increase accuracy meaning it works from 6”
• 22:01 “exactly same approach we can use in business Central to make sure we can answer the question about business data in a trustful and”
• 25:23 “tools are awesome it’s a best tool for a job to retrieve data from a system to enrich your responses to make sure it’s”
• 25:44 “if you look at Microsoft features we use around five fiveish and we see pretty pretty good U results so we would suggest you”
• 26:58 “if you want to build like a gut capil which can help you with everything uh you probably don’t have enough technology for now”

Disclaimers in the video

• 0:25 other: in this release we finally give you ability to run your extension in our Microsoft cloud
• 3:39 coming-later: Deep Seek R1 reasoning Model and other reasoning models we need capacity before we can give you access to those models
• 19:48 other: dealing with Fabrications should be a number one priority as you consider to bring copilot to your users just to avoid lack of trust frustration bad reputation brand damage
• 27:12 other: over the last releases and this release we have a number of great great videos how to extend copilot in business Central and how to build your own copilot