Business Central Under the Hood 11: New Data-Driven Copilot Experiences – Summarize with Copilot

Overview

Episode 11 of Business Central Under the Hood explains how Summarize with Copilot works. It runs on single entity pages such as card, document and list pages, and shows the top 2-3 insights in a summary fact box, with a Show more option that opens a chat experience. Behind it is a pipeline: a find insights prompt reads data from the current page, related pages and fact boxes in parallel, then a summarize prompt picks the top insights and applies formatting.

The second half covers engineering practice: testing AI with AI, grounding and harm checks, quality metrics, experimentation, model selection and migration to smaller models. It closes with guidance for partners and developers on custom entities, where summarization works out of the box but cannot yet be extended directly from AL.

Key points

• Summarize with Copilot is available on any single entity page (card, document, ListPlus), including custom partner pages, and appears in a summary fact box at the top of the fact box pane.
• The pipeline has a find insights prompt (current page fields, related pages and fact boxes, run in parallel) followed by a summarize prompt that selects the top-scored insights and formats them in markdown.
• Related pages are actions on the current page that lead to pages named statistics. Aggregated data performed better than sending raw ledger entries.
• Role is the current user context signal used to prioritize insights. Inferring user intent from navigation patterns and frequently used pages is planned, as are more data sources.
• Insights must reference source data IDs via function calling, which supports grounding. Testing also covers harm checks and functional checks with large role and scenario datasets, plus metrics like brevity and precision.
• Summarization prompts moved from 4o and 4o mini to 4.1 mini for speed. The scoring prompt might benefit from reasoning models such as o1, and the team is still experimenting.
• Developers cannot yet add data sources or override summary text from AL (work in progress). They can improve results with metadata, extra fact boxes, actions to statistics pages, and clear captions and tooltips.

Features

Quotes

• 1:43 “So summarizing with copilot highlights what’s most important and urgent for the user on almost every business central page.”
• 2:32 “So summarization with copilot is available on any business central uh single entity pages like card documents or list plus.”
• 3:52 “The way we build summarization, it’s pretty much a three-step process.”
• 7:56 “we filter out only the pages that are named statistics because that’s where the aggregated data exist and that’s very good for summarization”
• 8:51 “So we started with sending pretty much uh all the ledger entries and all and we soon realized uh it’s not very performant cuz”
• 13:34 “in future we do have plans of increasing both the tracks the user context track but also the data context.”
• 14:17 “there’s definitely very good early adoption. We are seeing that. Um uh there’s a little bit of lesser percentage of people clicking show more.”
• 16:14 “what I learned more is that tests are super super important especially because you change one line of prompt and the response could be”
• 18:49 “Uh definitely one of the new things about uh testing with AI features is that you use AI to test AI features.”
• 21:27 “So we came up with some tremendous large test data set containing multiple roles and multiple scenarios with different entities. uh some for example”
• 25:19 “And now uh we have migrated all of them to 41 mini.”
• 25:56 “Uh it depends on the need of the feature, right? Definitely. uh for for summarization specifically um we believe that 41 mini is enough”
• 26:30 “Some of the prompts, for example, the scoring part could be probably better done with uh other models than 41 mini uh because they”
• 27:07 “do try it with other models do and that’s why have a good set of test so you could easily switch between models”
• 27:27 “Another advantage of small language or small language mod is that they’re also cheaper, right? Definitely. Yes. But you need to maintain them yourself.”
• 28:04 “so the as i said summarization works ex out of the box with custom entities”
• 28:17 “Right now it is not directly extensible from AL in the sense that uh they they cannot provide for example more data sources or”
• 29:25 “as a minimum put good tool tips on the on the pages then you’ll get probably a better summary because that”
• 29:41 “think of AI as your co-pilot. There’s a reason we call it co-pilot because it is uh one of them, right?”

Disclaimers in the video

• 12:03 other: In future we do have plans of increasing both the tracks the user context track but also the data context.
• 13:34 preview: we also want to stabilize uh more and more with the current version and and then in future we do have plans of increasing both the tracks the user context track but also the data context
• 13:47 coming-later: We could easily think of uh you know integrating with external data be it um data from LinkedIn or any other external sources
• 23:37 coming-later: we’re working um uh on doing what we call experimentation. I don’t want to tell much about it because that’s that could be a subject an entire an entire episode
• 26:30 preview: Some of the prompts, for example, the scoring part could be probably better done with uh other models than 41 mini uh because they might have better reasoning capabilities and we are still figuring out
• 28:17 coming-later: Right now it is not directly extensible from AL in the sense that uh they they cannot provide for example more data sources or they cannot override the text that they see in the summary in one way or another yet but we are working on it.

Presenters (as heard): Vincent, Monica.