Three years ago, I was a financial analyst at a Series B startup. I spent more time fighting Excel than I did analyzing the data inside it. The breaking point came on a Tuesday in March, when I lost an entire afternoon to a missing dollar sign in a SUMIFS formula that had been pasted incorrectly through three iterations of a model.

That night, I wrote the first prototype of what would eventually become Formula Byte. We've rebuilt it from the ground up three times since. This is what we learned each time, and why I think the current version is finally the right one.

01 · OriginThe breaking point

Most analysts I know have a story like that Tuesday. The specifics vary — for some it's VLOOKUP returning #N/A on a range that should match perfectly. For others it's a pivot table that won't refresh because of a rogue blank row. The pattern is the same: a small mechanical issue eats hours, and the actual thinking part of the job — the analysis — gets squeezed into whatever's left.

Spreadsheets are powerful. They're also unforgiving. The interface assumes you know exactly what you want to build before you build it. That's not how real work happens.

02 · LessonsA wrong abstraction

The first prototype was an Excel add-in. You'd select a range, type a question in natural language, and it would suggest a formula.

// User typed:
"Sum the rows in column C where column A says 'Rent'"       

↓

// Formula Byte returned:
=SUMIFS(C:C, A:A, "Rent")

It worked. We had three hundred users in two months. Then we noticed something weird: nobody was opening it twice.

The problem was the abstraction. Translating English to a formula is a translation problem — and translation is only useful if you don't already know the answer. The people who needed Formula Byte most weren't analysts; they were the operations manager at a small business who'd never written a SUMIFS in their life and never would. To them, “translation” meant nothing. They wanted the whole spreadsheet.

What we learned

  • Most spreadsheet users don't think in terms of formulas. They think in terms of outcomes.
  • Translation tools are useful for experts; structure tools are useful for everyone else.
  • A prompt that returns a single formula doesn't change anyone's workflow. A prompt that returns a system does.

03 · IterationTemplates, almost

The second attempt was templates. You'd pick from a library — P&L, invoice, inventory tracker — and we'd customize it to your business based on a quick description.

This worked better. Users came back. Some of them upgraded.

But the templates were still mostly static. We were essentially a faster way to download a generic Excel file. The AI part was buried in the customization step, and most users didn't realize that's what was happening.

The lesson was almost the inverse of the first attempt: structure wasn't enough on its own. You needed generation. Every business runs a slightly different bookkeeping process. A template that fits everyone fits no one.

The best spreadsheet is the one you don't have to think about. The second-best is the one already shaped like the problem you're solving.A user we lost in week six, who came back in month four

04 · BreakthroughWhat actually worked

The Formula Byte that exists today does one thing: it generates the whole spreadsheet from a description.

Not a formula. Not a generic template. A complete, customized Excel system — worksheets, formulas, dashboards, and reporting — built to fit the specific workflow you describe.

The architecture looks something like this:

User prompt→Intent parser→Structure planner→Sheet generator→Formula composer→Excel output

The interesting work happens in the middle two steps. The structure planner figures out how many sheets you need and what each one tracks. The sheet generator lays out columns, validation rules, and reference relationships before any formulas are written. By the time the formula composer runs, the spreadsheet already knows what it is.

That was the unlock. The output isn't a template you adapt. It's a system that's already adapted.

Key Takeaway

The right abstraction is the one that matches how users already think about the problem. Translation tools are for experts. Templates are for shoppers. Generation is for everyone who just wants the spreadsheet to exist.

05 · MistakesWhat we got wrong (and what we're still getting wrong)

A few things that look obvious now, that took us 18 months to figure out:

  1. The output is the product. Not the prompt, not the AI, not the interface. The Excel file that lands on your machine is the product. Every other design choice should be in service of that.
  2. Dashboards are not extras. Every spreadsheet should generate with a dashboard on by default. We learned this from a hardware-store owner who'd been using us for three months and didn't realize his template had a P&L summary on sheet 9.
  3. The naming convention matters more than we thought. Q4_REV_v3_FINAL_actual_final.xlsx is not just a meme. We now generate filenames as part of the output.

We're still wrong about plenty. Comments and reviews surface things we missed every week. That's fine — that's how it's supposed to work.

06 · ForwardWhat comes next

Formula Byte today is good at generation. It's not yet great at evolution — what happens when your business changes and the spreadsheet needs to follow. That's where we're spending our next year.

If you've read this far, you probably care about spreadsheets more than is healthy. Welcome. We're hiring, and we read every email that comes to hi@formulabyte.com.