LLM Agents + Spreadsheets = Superpower

While I believe custom, task driven, AI solutions are what will drive holistic business value in the future (because a general use LLM is very expense to operate in comparison), I am convinced that AI agents in a spreadsheet are one of the most powerful generative AI tools available today.

I have a lot of high hopes for Microsoft’s Copilot in Office 365, but currently it is limited. In Excel I get “I'm sorry, but I don't have the capability to…” far too frequently. Therefore, I built a POC for my most common desired use case, LLM prompting as a custom Excel function.

Imagine taking an Excel sheet and sending a series of ChatGPT-like agents out to discover additional information, perform analysis, and then produce information based on those inputs and prompts, all within an Excel sheet and with complete control of the context window. We are now doing this at Advisor Labs and each day we find new ways to leverage it further.

With the POC, you engage the LLM as a function. The function takes input text, input fields, and an LLM prompt that uses those inputs. You can have one input or a series of inputs. The product is the LLM’s response, which is placed in the field of the function. Once the function is working as desired, use Excel’s Autofill feature to drag it across all the rows you want the same prompt applied to in the sheet. Here is simple everyday example, but there are plenty of role specific business use cases:

=AdvisorAI(“The name of a person that gave me a wedding gift”, A2, “The name of the gift”, B2, “Write a short thank you note for the wedding gift. Include their name and reference the gift by name.”) ….

There are many reasons why I think this will be useful as either Copilot catches up or the plugin community capitalizes. The two most noteworthy reasons are:

1. It saves a lot of time to prompt directly in Excel - While there are ways to prompt in ChatGPT and request an output that can translate to Excel, it is time consuming, especially when there is a complicated series of prompts. Having AI Agents as a function in Excel integrates into everyone’s existing workflow. If someone can utilize a SUM function, they can use this.
2. Each field in the sheet can be its own LLM context window - While ChatGPT can produce tables and use lists as an input, anyone that tries it will find that ChatGPT struggles when the input gets big or when you want to perform a series of nested prompts. By having AI Agents that can work in parallel with each other and can operate in as small as a single cells as a context window, you do not run into large context window confusion.

Before building the POC, I did find that there are a few plugins out there already that are doing something similar. I couldn’t find any that formally stated what they are or are not doing with customer data, so I decided to make my own (if anyone doesn’t understand why this is important, read my other post on why businesses need a private LLM and shouldn’t use private or confidential data in ChatGPT).

If anyone wants to put something similar together, I recommend using Azure AI Studio for prototyping. It is straightforward. If you haven’t used it before, below is a tutorial that will give you a lot of hands-on confidence. Then you’ll just need to pick how you want to integrate with Excel/Google Sheets:

Azure AI Studio Tutorial-

https://learn.microsoft.com/en-us/azure/ai-studio/tutorials/deploy-copilot-sdk

Excel Custom Functions-

https://learn.microsoft.com/en-us/office/dev/add-ins/quickstarts/excel-custom-functions-quickstart?tabs=excel-windows

If you run into any problems or want to talk through the potential, feel free to reach out to me on LinkedIn.