It started with a failure that never showed up in a bug report.

Someone attached a program status deck to a project chat and asked the AI about it. The app did what basically every chat app does by default: it pulled the text out of the file and threw everything else away. The words survived. But the budget chart came through as a few stray axis labels, and the milestone table lost its shape. The answer looked confident and was quietly missing the most important thing in the deck. That kind of silent failure is worse than an error, because you have no reason to distrust the answer.

What we learned about how AI reads files

AI models do not eat files. They eat either text or pages.

A model can genuinely read a PDF the way you do, looking at the rendered pages, charts and scans included. But no model on the market accepts a PowerPoint or Excel file directly. Not ours, not anyone's. Every AI app you have ever dropped an Office file into is quietly converting or extracting behind the curtain.

So the real question was never how to send the file. It was which of two very different reading experiences each file deserves, and what each one costs.

The costs are delightfully backwards

Cost follows what the model reads, not what you upload:

What we sent	File size	Relative cost to answer
A full slide deck, as extracted text	24.7 MB	about a third of a cent
A two page resume PDF, read visually	73 KB	several times the giant deck

The deck was mostly pictures, and words are cheap. The little PDF cost more because the model actually looked at the pages, and page reading is the premium experience.

We also tested the obvious compromise: rip the images out of a deck and send those along with the text. Real data killed it in one afternoon. On our test deck it cost the same as full page reading while delivering a fraction of the value, and the biggest images in the file turned out to be decorative slide backgrounds, not the charts that mattered.

The balance we landed on

Text extraction stays free and universal. Works with every model including your own keys, costs pennies, and for text heavy documents it loses almost nothing.

Full visual PDF reading is coming to Pro. It is the genuinely premium experience, and it costs real money per question.

For Office formats, we refused to ship a half measure. For now the honest guidance is export to PDF, which is lossless today. The real fix was already on our roadmap, and every hour spent on a stopgap would have been a bet against it.

The part we're most excited about

That real fix is in final testing now. Instead of straining to read Office files better, we gave the AI a workspace where it opens the actual file programmatically and, more importantly, writes files back. Ask for a tracking spreadsheet built from your meeting and you get a real Excel file, not a wall of text you have to copy into one.

Here's the contrast that stopped us: reading a 25 megabyte deck cost a third of a cent. Creating one small spreadsheet cost almost a dollar. Same AI, three orders of magnitude apart, because generation isn't one read, it's the model reading your meeting, writing code, running it, checking its own work, and reading everything again at every step. That is not a business at a dollar a file. Then the secret sauce went in, and it now runs at about a third of that, which very much is one. We knew both numbers before committing to the build, because we measured them on real meetings first. That habit saved us twice this week.

One thing we'll brag about

Every limit in this system, file sizes, page counts, formats, is a value our servers send to the app, not code baked into it. When the underlying models get better, and they will, your app gets more capable overnight, no app store update required.

And your files stay yours. Attachments live on your device, generated files are saved into your meeting the moment they arrive, and our servers hold bytes for hours, not forever. No dead links in your history, no warehouse of your documents on our side.

More soon. The spreadsheet feature in particular has been fun to watch work.