Unlocking Spreadsheet Intelligence with AI

Ben Lorica

Hjalmar Gislason on Spreadsheets, AI, and the Future of Business Modeling.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • AntennaPod • Podcast Addict • Amazon •  RSS.

Hjalmar Gislason is founder and CEO of Grid, a startup focused on bringing spreadsheets into the AI-first era. This episode delves into the evolving landscape of spreadsheets, exploring how AI is transforming these ubiquitous tools. We discuss the shift from traditional spreadsheet software to AI-integrated solutions, examining the technical challenges and opportunities of making spreadsheets accessible to AI through natural language interfaces, web services, and function calls. We also cover the importance of dependency graphs, metadata, and advanced spreadsheet features, along with real-world use cases and future trends in AI-first modeling and richer UI interactions.

Subscribe to the Gradient Flow Newsletter

 

Interview highlights – key sections from the video version:

 

Related content:

If you enjoyed this episode, consider supporting our work by leaving a small tip here and inviting your friends and colleagues to subscribe:

Transcript.

Below is a heavily edited excerpt, in Question & Answer format.

What problem is Grid trying to solve with spreadsheets in the AI era?

Spreadsheets are the foundation of countless workflows in organizations, especially on the business side where people use them to address everyday IT needs without involving IT departments. These spreadsheets contain valuable business logic and data, but they’re currently out of reach for AI and automation because they don’t lend themselves well to natural language interaction or integration with automated systems.

At Grid, we’re making spreadsheets accessible to AI by turning them into web services with RESTful API endpoints that can be integrated into AI applications. While other companies are bringing AI to spreadsheets, we’re bringing spreadsheets to AI.

How does Grid fit into the broader landscape of spreadsheet startups?

There are several categories of new spreadsheet companies. First are those trying to reinvent the spreadsheet itself – companies like Rows, Equals, and Spreadsheet.com aim to replace Excel or Google Sheets as your primary spreadsheet application.

Then there are companies like Sigma Computing that focus on tabular data – they provide a spreadsheet-like UI on top of columnar data, treating spreadsheets more like databases where columns are attributes and rows are entities.

Grid takes a different approach. We believe people will continue to build models in traditional spreadsheet software, but those spreadsheet models need to be accessible via natural language, AI, and automation. We turn existing spreadsheets into web services that can be integrated into AI applications.

What’s the distinction between different types of spreadsheet uses that’s important here?

There are essentially two types of spreadsheets that people work with. First are tabular data spreadsheets, which function almost like databases – columns are attributes and rows are entities or transactions. Most AI applications for spreadsheets focus on this type, treating them essentially as CSV files.

The second type, which is our focus, are free-form models where people encode business logic in formulas and cell relationships. These are effectively programs written in a two-dimensional space that encode business logic. When people build financial models or ROI calculators, they’re essentially programming using spreadsheets.

Traditional AI approaches to spreadsheets have mostly focused on tabular data while ignoring formulas. Our key differentiation is that we handle the formulas and complex spreadsheet models that contain sophisticated business logic.

How exactly does Grid make spreadsheets accessible to AI systems?

We do two important things. First, we turn the spreadsheet into a web service that’s available 24/7 for querying in a scalable fashion.

Second, we analyze the spreadsheet to transform cell references into something that can be accessed using natural language. A cell might just have a number in it, but there are usually labels surrounding it that help humans understand what that number represents. We use heuristics and AI to extract the right metadata and labels for those cells, so they become accessible via natural language rather than cell references like “B4”.

This means an LLM or an application can interact with the spreadsheet using natural language to refer to values like “monthly price per seat” instead of having to know specific cell locations.

Can you provide an example of how Grid works in practice?

One example is a small restaurant chain called Joe and the Juice that we worked with to set up automation for their catering solutions. They frequently receive emails requesting catering for meetings with details like number of attendees, dietary requirements, etc.

Before our solution, employees would manually read these emails, enter the information into a spreadsheet to calculate pricing and required grocery orders. Now, they can paste the text into a custom GPT connected to their spreadsheet through our API. The system automatically extracts the relevant information, fills out the spreadsheet, and generates both a professional PDF quote for the customer and updates to their grocery list.

This happens almost instantly, replacing what was previously a lengthy manual process. The reason they were using spreadsheets is that’s what small businesses do – they run on spreadsheets. We’ve made that process AI-accessible using a simple set of tools.

How does Grid handle complex spreadsheets that reference other spreadsheets?

We handle cross-workbook references quite well, which is especially common in finance where people create chains of interconnected spreadsheets. Our spreadsheet engine supports advanced features like cross-workbook references, named ranges, number formatting, newer functions like LAMBDA and LET, iterative calculations, and Goal Seek.

These are features that advanced spreadsheet users rely on, and most spreadsheet engines can’t fully support them. We’ve built one of the most sophisticated independent spreadsheet engines in the world to handle these complex spreadsheets.

When a user interacts with a spreadsheet through our system that relies on other spreadsheets, we maintain those dependencies and can provide accurate results based on the entire chain of calculations.

What kind of “what-if” scenarios can users explore through natural language?

Users can ask questions like “If interest rates go down by 1 percentage point, how does that affect our cash position, and when will that happen?” The LLM isn’t generating calculations or hallucinating – it’s taking the new parameters (like interest rates), plugging them into the right cells in the spreadsheet model, and then returning the results from the appropriate output cells.

We can also run more advanced analyses like Goal Seek, sensitivity analysis, or even simulation-style approaches where we can run thousands of iterations over a spreadsheet with different inputs and provide a distribution of outcomes.

This is much more reliable than trying to teach LLMs to do math calculations themselves. We’re using proven calculation engines but making them accessible through natural language.

How does Grid provide transparency and explainability to build user trust?

Users can ask the system to explain how it arrived at a particular result. The LLM will walk through how it understood the query and which cells or ranges it referenced in the spreadsheet.

We also provide the ability to return relevant portions of the spreadsheet visually, so users can see exactly which parts were changed and what the results look like. This is particularly valuable during the verification stage when users are still getting comfortable with the system.

While some use cases (like a pricing agent on a website) might not need to expose that there’s a spreadsheet behind the scenes, for internal users who are interrogating spreadsheets they’re familiar with, seeing the relevant parts of the spreadsheet builds trust and understanding.

What’s on Grid’s roadmap for the next 6-12 months?

We’re releasing a commercial version of our solution in February 2024. Currently, we’re in alpha with about 200 POCs that we’re working on. The initial release will focus on our API solution that integrates well with function calling from custom GPTs in ChatGPT and other AI assistants.

Looking further ahead, we’re watching how LLM environments are evolving to support richer user interfaces through artifacts and canvases. We already have UI components that can make spreadsheet models interactive – with charts, sliders that adjust inputs, buttons that change scenarios, etc. We expect to integrate these richer UI components into LLM environments as they mature.

Our vision is to bridge the gap between the natural language interface of LLMs and the visual, interactive nature of spreadsheets that makes them so powerful and understandable.

How will users be able to get started with Grid when it launches commercially?

Starting in February, anyone will be able to use Grid as a self-service solution. You can upload your Excel or Google Sheets files to Grid, which will create an API endpoint that you can use to interact with your spreadsheet through function calls.

For cloud spreadsheets like Google Sheets or Excel files stored in OneDrive or Dropbox, you can connect them live so any changes you make to the spreadsheet are immediately available to the LLM. This means your AI applications always have access to the latest version of your spreadsheet data and logic.

Until the commercial launch, we’re running an alpha program that anyone can sign up for, though we’re providing more hands-on support during this phase.