Prompt Columns: Pre-computed AI insights, ready for every agent
Read the Portuguese translation →
Author: Microsoft Power Platform · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Prompt Columns: Pre-computed AI insights, ready for every agent · Channel: Microsoft Power Platform · 06:19 · Watch on YouTube
Topics: AI agents, Task and workflow automation, Data, databases and embeddings
Summary
Section titled “Summary”Prompt Columns generate AI insights from descriptive data in advance and store them as fields in Dataverse, so agents can answer questions without repeatedly processing large volumes of raw notes. The video demonstrates creating prompt columns, configuring prompts and filters, and generating highlights, risks, and recommended actions when meeting data changes. An agent then uses those stored insights to identify risky accounts and suggest next steps, enabling faster responses and other workflows. Prompt Columns is generally available in Dataverse, with historical-record backfilling and additional input and output options described as coming soon.
Key points
Section titled “Key points”- Prompt Columns address the slowness of having agents reason over large volumes of raw data at run time.
- Generated insights are stored alongside the source data as first-class fields for reuse.
- A prompt column is created by adding a column, selecting the prompt data type, and configuring its prompt in natural language.
- Users can test prompts, choose a model, and filter which records the prompt runs on.
- The demo uses columns for meeting highlights, potential risks or blockers, and prioritized recommended actions.
- Prompts run when the input data changes, such as when a meeting record is added or its notes are updated.
- Agents can use the precomputed insights to answer questions about risky accounts and recommended next steps.
- The insights can also support other uses, such as sending emails or creating tasks with Power Automate.
- Prompt Columns is generally available in Dataverse; backfilling historical records and more input and output options are coming soon.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | The challenge of reasoning over raw data | The presenters describe how large volumes of customer notes make on-the-fly agent analysis slow and introduce Prompt Columns as a way to prepare reusable insights in advance. |
| 01:34 | Creating and configuring prompt columns | The demo shows how to create a prompt column, configure it in natural language, select a model, and filter the records to analyze. |
| 02:38 | Generated insights and when prompts run | The presenters explain the highlights, lowlights, and recommended actions columns, and show that prompts run when the input data changes. |
| 03:41 | Using prompt column insights with an agent | A new meeting record generates AI insights that an agent then uses to identify risky accounts and recommend actions for Contoso. |
| 05:17 | Additional uses and what is coming next | The presenters describe sharing insights and using them in workflows, note that Prompt Columns is generally available in Dataverse, and preview upcoming features. |
Quotes
Section titled “Quotes”The result was unusably slow, and no amount of context window fixed it. — 01:03
prompt columns generate key insights in advance — 01:03
Prompt columns execute only when the input data changes. — 03:08
The result is a faster and more scalable experience. — 04:14
Prompt Columns is a pretty powerful feature, which is now generally available in Dataverse. — 05:48
Concepts and entities
Section titled “Concepts and entities”Prompt Columns, Dataverse, Power Automate, LLM, Microsoft, Insha Arora, Vaishanavi Mirapurkar
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 [ Music ] INSHA ARORA: Hi, I’m Insha, Engineering Manager at Microsoft. VAISHANAVI MIRAPURKAR: And I’m Vaishanavi, Product Manager at Microsoft. INSHA ARORA: With more powerful LLMs, agents are getting better at reasoning over large volumes of descriptive data.
00:32 But doing that on the fly, in a brute force way, is that really the best approach? VAISHANAVI MIRAPURKAR: That’s exactly the challenge one of our customers had. As demand for their products exploded, so did the raw material piling up across their largest accounts. Like customer meeting notes, account reviews, escalation threads, and product feedback. Their sales and account teams wanted insights to answer questions like, which accounts are at risk? What should the account team do next for a specific account? An agent could answer these on the fly,
01:03 but only by doing complex reasoning on large volume of raw notes at run time. The result was unusably slow, and no amount of context window fixed it. INSHA ARORA: This is where prompt columns change the game. Instead of analyzing large volumes of unprocessed data every time an agent needs to reason, prompt columns generate key insights in advance, such as account risk and recommended actions, and store them alongside the data as first class fields. The result is a faster answer, scalable solution,
01:34 and consistent insights the team can reuse at scale. VAISHANAVI MIRAPURKAR: All right, let’s see how we can solve our customer’s problem with prompt columns then. This is what the data set looks like. It’s an extensive collection of meeting notes across customer accounts, dates, and account owners. It’s very easy to create prompt columns. All you do is create a new column, add your column’s name – – and select prompt as a data type.
02:08 You can then configure your prompt in natural language, like so. You can also test your prompt and choose your model. You can apply filters to choose the records you want to run the prompt on. Like in this case, it can be a specific customer name, like Contoso.
02:38 We’ve already created three prompt columns for our customers to analyze the meeting notes. Let’s take a look at them. Meeting highlights capture positive signals, like customer feedback, successful outcomes, business value achieved, etc. Meeting lowlights identify potential risks, concerns, or blockers. And recommended next actions have the most impactful
03:08 follow-ups prioritized by urgency, like low, medium, or high. INSHA ARORA: You might be wondering when the prompt runs. Prompt columns execute only when the input data changes. For example, when a new meeting record is added or meeting notes are updated on existing record, the notes are automatically analyzed to generate Highlight, Lowlight, and recommended next steps. If the relevant input data doesn’t change, the prompt column does not execute.
03:41 Let’s add a new record using the app and see prompt column in action. Here we have a model-driven app built for customer meeting history. I’m going to go ahead and add some meeting notes for Account Health Insights Demo, and “Save”. As you can see, we now have meeting Highlights, Lowlights, and recommended action generated by our prompt columns. The little icon here is indicating that these are all AI-generated fields. These AI-generated insights are now available in our table for further use, such as reasoning by an agent.
04:14 VAISHANAVI MIRAPURKAR: Now when an agent needs to answer a question, it can reason over the insights generated by prompt columns in advance instead of processing all of the raw meeting notes. The result is a faster and more scalable experience. Let’s see this in action. I have a meeting notes agent created here with the customer meeting history table as its knowledge source. Let’s move on to this agent and ask some questions.
04:45 Who are the riskiest accounts? That was pretty fast. Looks like the Account Manager for Contoso, Fabrikam, and Northwind needs to dig into these risks ASAP. Let’s see another one. What should the account team do for Contoso next? So you can see there are some high-priority items, like improve support SLAs, set up enhanced monitoring, and dedicated support channel.
05:17 There are also some medium-priority and low-priority items on the list, along with a key takeaway in the end. INSHA ARORA: Great, this is pretty detailed and fast. An account manager can share this with their team and take actions immediately. All thanks to insights generated and persist as data by Prompt Columns. They can be used in so many other ways, like having a power automate flow to send an email to the account team when there is a critical lowlight or create tasks by recommended next steps.
05:48 VAISHANAVI MIRAPURKAR: That’s awesome. As you can see, Prompt Columns is a pretty powerful feature, which is now generally available in Dataverse. You might be asking, what’s next? We have backfilling historical records, which is coming out real soon. Here’s a little sneak peek of that. I can now select “Apply to existing and new records” option here, which starts backfilling my historical records. We also have more input and output options, like multiple lines of text, choice sets, rich text, files and images coming out soon as well.
06:19 Here is the code to start building your own prompt columns. INSHA ARORA: Thanks for watching, and happy building!