Why your chatbot really needs a spreadsheet

I recently wrote about using ChatGPT as an impressive analytic partner for spreadsheet data. You can do something similar with Anthropic Claude and CSV files.

However, in the last paragraph of that post, I hinted at a problem. A client with whom I had been experimenting, a small construction company, had some very specific needs. In construction, numerous calculations need to be accurately carried out - calculating the volume of concrete to order, the size of a retaining wall, or even the hourly rate to charge for a specific project.

Here's one very similar to my client's spreadsheet, in this case from BuildBook, a construction software vendor.

These are not analyses or predictions; they demand precise, reliable results.

Why can't an LLM do this?

Although LLMs are getting better at simple arithmetic, ChatGPT and similar AI systems don't actually perform complex calculations directly. Instead, they act more like coordinators:

• They use libraries of code for analytics, that have been thoroughly tested, mostly by the Javascript community

• They write instructions (code) to pass data to these libraries

• They interpret and present the results back in plain language or as graphics.

But they are not doing the calculations. This does not prevent them from giving confident answers, as you should probably realize by now.

This becomes particularly important with spreadsheets, which are the backbone of business calculations. Most businesses, like my construction client, have a ton of spreadsheets with specialized calculations. I can't expect them to rebuild all these calculations as code libraries.

I am glad to say that now they don't have to.

Hooking up to the GRID

The solution I like is from GRID - an Icelandic company I know very well. Their tagline is Spreadsheets run the world and they are not wrong.

When it comes to integrating these world-conquering spreadsheets with LLMS, GRID's approach is to let each component do what it does best:

• LLMs handle natural language understanding and generation

• Specialized engines (like GRID) handle calculations and spreadsheet logic

• Function calls connect these components seamlessly

This creates a more reliable system because calculations are verifiable, business logic remains intact and results can be trusted. Here's how it works …

Using the GRID API (You can get access via their site), upload your spreadsheet to their server. I loaded what is, in essence, a version of the BuildBook calculator for the sake of a demo.

You then set up a connection to ChatGPT by creating a new custom GTP and copying the custom instructions that GRID generates for you. You will add an API key and a schema that GRID also generates.

In short, you'll follow some straightforward instructions and paste GRID's customized details into ChatGPT. The result is a custom GPT that understands how to connect to your spreadsheet in natural language. Look at this; it's a real conversation …

This scenario is so simple that it may not initially seem like a practical application of a GPT. However, this simple model could be used on a phone with voice recognition as a simple chatbot. Meanwhile, the usability of more complex spreadsheets could be greatly improved.

In every business, you will find more people with spreadsheet skills than coding skills, so GRID makes it much easier to automate some processes requiring accurate calculations.

And finally, here's something Excel cannot do ...

Try it for yourself … www.GRID.is

I’ll have more to say about other spreadsheet integrations with AI in future posts. After all, spreadsheets do run the world … for now.