--- id: video/jdevJbGgB-k type: video title: "Business Central Under the Hood episode 1: Developing AI in Business Central" summary: "Developing AI features in Business Central: how Microsoft's team designs LLM features, covering prompt techniques (few-shot, Chain of Thought, function calling, meta-prompting), the Copilot chat service architecture, hallucination and jailbreak prevention, and the AI toolkit for partners." tier: official language: en tags: - ai development - llm integration - prompt engineering - few-shot learning - chain of thought - marketing text generation - bank reconciliation - immersive features - chat copilot - testing and validation - function calling - hallucinations system: copilot review: state: reviewed by: opus at: "2026-10-07T23:12:08.911Z" flags: [] generated: at: "2026-10-07T23:12:08.965Z" pipeline: 0.2.0 prompts: extract-video: 1 summarize-video: 2 input_hash: bf0f81e0510ff82d17dd08f7ff22ae72b23917ea73e6c66abcf4cbb59fc7b015 evidence: - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=213s title: "Marketing text generation: generally available" date: "2024-02-13T15:07:50.000Z" commit: null t: 213 quote: we have released at least marketing text where we can kind of generate the marketing text based on kind of keywords - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=233s title: "Chat Copilot for Business Central: announced" date: "2024-02-13T15:07:50.000Z" commit: null t: 233 quote: one of the big features we're working on is to kind of have a chat inside business Central - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=10s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 10 quote: we are going to tell you how we develop these features how we design them how we architect them - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=213s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 213 quote: we have released at least marketing text where we can kind of generate the marketing text based on kind of keywords - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=233s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 233 quote: we are starting slowly we're also learning a lot by building these features turns out this is this is probably harder than we thought - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=287s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 287 quote: we have already shipped uh bank reconciliation right try to match use llms to match bank records behind the product - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=490s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 490 quote: prompting is not an exact science right because it's no one let's put it that way no one can really predict exactly what the - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=652s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 652 quote: because this may not be really predictable right you actually need another prompt to to validate in many cases at least right - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=734s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 734 quote: one of them and kind of a very successful technique is what kind of called few shot learning right and this basically that in - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=815s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 815 quote: CH for is really just like what humans does right you kind of break break down the problem you you solve small sub parts - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=855s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 855 quote: let's think about it step by step it's really breaking down the you write that in the prompt you simply just put that in - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=976s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 976 quote: a chat is like it's it's a multi-turn of or turn of questions and answers right and the user may clarify or change subject - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1158s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1158 quote: it tells you how to execute it doesn't actually execute the code right that's up to you but it guides you - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1178s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1178 quote: Hallucination is basically when you don't give it up to-date information right and the way this whole thing is trained it really wants to - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1218s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1218 quote: you always feed the data and you can instruct in the prompt that you really only use this data - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1259s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1259 quote: jailbreak the prompt in getting it to do something else then you know the developer or whoever made this prompt was intended it to - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1420s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1420 quote: maybe you should not start with the absolutely mind-blowing stuff unless that mind-blowing stuff is really simple I think it's good to develop a - kind: video url: https://www.youtube.com/watch?v=jdevJbGgB-k&t=1502s title: "Business Central Under the Hood episode 1: Developing AI in Business Central" date: "2024-02-13T15:07:50.000Z" commit: null t: 1502 quote: on Azure there's also the Azure um open a studio yeah right which is also a it's very useful tool to kind of construct links: learn: [] objects: [] features: [] topics: [] localizations: [] videos: [] posts: [] guidelines: [] video_id: jdevJbGgB-k channel: yt-microsoft source_name: Microsoft Dynamics 365 Business Central (YouTube) url: https://www.youtube.com/watch?v=jdevJbGgB-k published_at: "2024-02-13T15:07:50.000Z" duration_s: 1565 captions: full audience: - developer - functional consultant - partner chapters: - t: 0 title: Introduction to the Under the Hood series - t: 91 title: "From web technology to AI: Michael's journey" - t: 193 title: "AI features in Business Central: marketing text and chat" - t: 287 title: Immersive features and developer toolkit - t: 368 title: Prompting and LLM architecture basics - t: 490 title: Testing and prompt engineering challenges - t: 591 title: Testing fuzzy inputs and outputs with LLMs - t: 714 title: Few-shot learning technique - t: 815 title: Chain of Thought reasoning and prompt engineering - t: 976 title: Copilot service architecture and chat orchestration - t: 1057 title: Function calling and hallucination prevention - t: 1259 title: Jailbreaking prevention, meta-prompting, and AI toolkit - t: 1502 title: Azure OpenAI Studio and prompt development tools features: - name: Marketing text generation status: ga t: 213 verified: true status_source: video - name: Chat Copilot for Business Central status: announced t: 233 verified: true status_source: video - name: Bank reconciliation with LLM status: unclear t: 287 verified: false status_source: video - name: AI developer toolkit status: unclear t: 327 verified: false status_source: video - name: Few-shot learning prompting technique status: unclear t: 714 verified: false status_source: video - name: Chain of Thought prompting status: unclear t: 815 verified: false status_source: video - name: Copilot service architecture status: unclear t: 915 verified: false status_source: video - name: Chat orchestration status: unclear t: 956 verified: false status_source: video - name: Function calling status: unclear t: 1017 verified: false status_source: video - name: Hallucination prevention through data feeding status: unclear t: 1158 verified: false status_source: video - name: Content moderation and jailbreaking prevention status: unclear t: 1299 verified: false status_source: video - name: Meta-prompting technique status: unclear t: 1340 verified: false status_source: video - name: AI toolkit for partners and developers status: unclear t: 1380 verified: false status_source: video - name: Azure OpenAI Studio status: unclear t: 1482 verified: false status_source: video objects_mentioned: - other bank reconciliation - other Chat pane - other Orchestrator - other Plugin service quotes: - t: 10 text: we are going to tell you how we develop these features how we design them how we architect them check: exact - t: 213 text: we have released at least marketing text where we can kind of generate the marketing text based on kind of keywords check: exact - t: 233 text: we are starting slowly we're also learning a lot by building these features turns out this is this is probably harder than we thought check: exact - t: 287 text: we have already shipped uh bank reconciliation right try to match use llms to match bank records behind the product check: exact - t: 490 text: prompting is not an exact science right because it's no one let's put it that way no one can really predict exactly what the check: exact - t: 652 text: because this may not be really predictable right you actually need another prompt to to validate in many cases at least right check: exact - t: 734 text: one of them and kind of a very successful technique is what kind of called few shot learning right and this basically that in check: exact - t: 815 text: CH for is really just like what humans does right you kind of break break down the problem you you solve small sub parts check: snapped - t: 855 text: let's think about it step by step it's really breaking down the you write that in the prompt you simply just put that in check: exact - t: 976 text: a chat is like it's it's a multi-turn of or turn of questions and answers right and the user may clarify or change subject check: exact - t: 1158 text: it tells you how to execute it doesn't actually execute the code right that's up to you but it guides you check: exact - t: 1178 text: Hallucination is basically when you don't give it up to-date information right and the way this whole thing is trained it really wants to check: exact - t: 1218 text: you always feed the data and you can instruct in the prompt that you really only use this data check: exact - t: 1259 text: jailbreak the prompt in getting it to do something else then you know the developer or whoever made this prompt was intended it to check: exact - t: 1420 text: maybe you should not start with the absolutely mind-blowing stuff unless that mind-blowing stuff is really simple I think it's good to develop a check: exact - t: 1502 text: on Azure there's also the Azure um open a studio yeah right which is also a it's very useful tool to kind of construct check: exact --- # Business Central Under the Hood episode 1: Developing AI in Business Central > Developing AI features in Business Central: how Microsoft's team designs LLM features, covering prompt techniques (few-shot, Chain of Thought, function calling, meta-prompting), the Copilot chat service architecture, hallucination and jailbreak prevention, and the AI toolkit for partners. [Watch on YouTube](https://www.youtube.com/watch?v=jdevJbGgB-k) · Microsoft Dynamics 365 Business Central (YouTube) · 2024-02-13 · 26:05 · tier official · reviewed (checked by Opus) ## Overview Episode 1 of the Under the Hood series, published 2024-02-13, explains how the Business Central team builds AI features. It mentions marketing text generation and bank reconciliation matching as shipped features, and describes a chat Copilot that focuses on guidance and navigation and is still improving in quality. The talk then covers practical prompt engineering: few-shot learning, Chain of Thought, function calling, feeding the LLM current data to limit hallucinations, and meta-prompting against jailbreaks. It also describes the Copilot service architecture with a chat pane, orchestrator and plugin service, the difficulty of testing fuzzy inputs and outputs, and Azure OpenAI Studio as a tool for building prompts. ## Key points - Marketing text generation (generally available) creates item marketing text from keywords; bank reconciliation already uses LLMs to match bank records. - Few-shot learning shows the LLM examples of input and expected output, and makes it easy to specify formats such as JSON or XML. - Chain of Thought tells the LLM to reason step by step; it outputs intermediate steps and takes longer to respond. - Function calling gives the LLM function definitions and parameters; the LLM does not execute code, it only says what to call and with which parameters. - To limit hallucinations, always feed the LLM current data and instruct it in the prompt to answer only from the provided data. - Chat architecture: a chat pane in the client talks to a cloud Copilot service with an orchestrator for multi-turn history and context and a plugin service for skills. - Meta-prompting and Azure OpenAI content filters reduce jailbreaking, but new jailbreak techniques keep appearing; the user must stay in the loop. ## Chapters - [0:00](https://www.youtube.com/watch?v=jdevJbGgB-k&t=0s) Introduction to the Under the Hood series - [1:31](https://www.youtube.com/watch?v=jdevJbGgB-k&t=91s) From web technology to AI: Michael's journey - [3:13](https://www.youtube.com/watch?v=jdevJbGgB-k&t=193s) AI features in Business Central: marketing text and chat - [4:47](https://www.youtube.com/watch?v=jdevJbGgB-k&t=287s) Immersive features and developer toolkit - [6:08](https://www.youtube.com/watch?v=jdevJbGgB-k&t=368s) Prompting and LLM architecture basics - [8:10](https://www.youtube.com/watch?v=jdevJbGgB-k&t=490s) Testing and prompt engineering challenges - [9:51](https://www.youtube.com/watch?v=jdevJbGgB-k&t=591s) Testing fuzzy inputs and outputs with LLMs - [11:54](https://www.youtube.com/watch?v=jdevJbGgB-k&t=714s) Few-shot learning technique - [13:35](https://www.youtube.com/watch?v=jdevJbGgB-k&t=815s) Chain of Thought reasoning and prompt engineering - [16:16](https://www.youtube.com/watch?v=jdevJbGgB-k&t=976s) Copilot service architecture and chat orchestration - [17:37](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1057s) Function calling and hallucination prevention - [20:59](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1259s) Jailbreaking prevention, meta-prompting, and AI toolkit - [25:02](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1502s) Azure OpenAI Studio and prompt development tools ## Features | Feature | Status | At | Evidence | |---|---|---|---| | Marketing text generation | generally available | [3:33](https://www.youtube.com/watch?v=jdevJbGgB-k&t=213s) | "we have released at least marketing text where we can kind of generate the marketing text based on kind of keywords" ([3:33](https://www.youtube.com/watch?v=jdevJbGgB-k&t=213s)) | | Chat Copilot for Business Central | announced | [3:53](https://www.youtube.com/watch?v=jdevJbGgB-k&t=233s) | "one of the big features we're working on is to kind of have a chat inside business Central" ([3:53](https://www.youtube.com/watch?v=jdevJbGgB-k&t=233s)) | | Bank reconciliation with LLM | status not stated | [4:47](https://www.youtube.com/watch?v=jdevJbGgB-k&t=287s) | | | AI developer toolkit | status not stated | [5:27](https://www.youtube.com/watch?v=jdevJbGgB-k&t=327s) | | | Few-shot learning prompting technique | status not stated, demoed | [11:54](https://www.youtube.com/watch?v=jdevJbGgB-k&t=714s) | | | Chain of Thought prompting | status not stated, demoed | [13:35](https://www.youtube.com/watch?v=jdevJbGgB-k&t=815s) | | | Copilot service architecture | status not stated, demoed | [15:15](https://www.youtube.com/watch?v=jdevJbGgB-k&t=915s) | | | Chat orchestration | status not stated | [15:56](https://www.youtube.com/watch?v=jdevJbGgB-k&t=956s) | | | Function calling | status not stated, demoed | [16:57](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1017s) | | | Hallucination prevention through data feeding | status not stated | [19:18](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1158s) | | | Content moderation and jailbreaking prevention | status not stated | [21:39](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1299s) | | | Meta-prompting technique | status not stated | [22:20](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1340s) | | | AI toolkit for partners and developers | status not stated | [23:00](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1380s) | | | Azure OpenAI Studio | status not stated | [24:42](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1482s) | | ## 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 "bank reconciliation" at [4:47](https://www.youtube.com/watch?v=jdevJbGgB-k&t=287s) - other "Chat pane" at [15:56](https://www.youtube.com/watch?v=jdevJbGgB-k&t=956s) - other "Orchestrator" at [16:16](https://www.youtube.com/watch?v=jdevJbGgB-k&t=976s) - other "Plugin service" at [16:37](https://www.youtube.com/watch?v=jdevJbGgB-k&t=997s) ## Quotes - [0:10](https://www.youtube.com/watch?v=jdevJbGgB-k&t=10s) "we are going to tell you how we develop these features how we design them how we architect them" - [3:33](https://www.youtube.com/watch?v=jdevJbGgB-k&t=213s) "we have released at least marketing text where we can kind of generate the marketing text based on kind of keywords" - [3:53](https://www.youtube.com/watch?v=jdevJbGgB-k&t=233s) "we are starting slowly we're also learning a lot by building these features turns out this is this is probably harder than we thought" - [4:47](https://www.youtube.com/watch?v=jdevJbGgB-k&t=287s) "we have already shipped uh bank reconciliation right try to match use llms to match bank records behind the product" - [8:10](https://www.youtube.com/watch?v=jdevJbGgB-k&t=490s) "prompting is not an exact science right because it's no one let's put it that way no one can really predict exactly what the" - [10:52](https://www.youtube.com/watch?v=jdevJbGgB-k&t=652s) "because this may not be really predictable right you actually need another prompt to to validate in many cases at least right" - [12:14](https://www.youtube.com/watch?v=jdevJbGgB-k&t=734s) "one of them and kind of a very successful technique is what kind of called few shot learning right and this basically that in" - [13:35](https://www.youtube.com/watch?v=jdevJbGgB-k&t=815s) "CH for is really just like what humans does right you kind of break break down the problem you you solve small sub parts" - [14:15](https://www.youtube.com/watch?v=jdevJbGgB-k&t=855s) "let's think about it step by step it's really breaking down the you write that in the prompt you simply just put that in" - [16:16](https://www.youtube.com/watch?v=jdevJbGgB-k&t=976s) "a chat is like it's it's a multi-turn of or turn of questions and answers right and the user may clarify or change subject" - [19:18](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1158s) "it tells you how to execute it doesn't actually execute the code right that's up to you but it guides you" - [19:38](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1178s) "Hallucination is basically when you don't give it up to-date information right and the way this whole thing is trained it really wants to" - [20:18](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1218s) "you always feed the data and you can instruct in the prompt that you really only use this data" - [20:59](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1259s) "jailbreak the prompt in getting it to do something else then you know the developer or whoever made this prompt was intended it to" - [23:40](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1420s) "maybe you should not start with the absolutely mind-blowing stuff unless that mind-blowing stuff is really simple I think it's good to develop a" - [25:02](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1502s) "on Azure there's also the Azure um open a studio yeah right which is also a it's very useful tool to kind of construct" ## Disclaimers in the video - [2:53](https://www.youtube.com/watch?v=jdevJbGgB-k&t=173s) not-in-this-release: we don't talk about it yet we don't really talk about it yet right but but I I hope we will be able to productize this somewhat soon - [23:20](https://www.youtube.com/watch?v=jdevJbGgB-k&t=1400s) other: we are releasing this tool kit uh for partners and developers - describes toolkit being released for external use Presenters (as heard): Vincent Nicholas, Michael.