You upload the WIP export (50,000 rows) and ask for a customer summary. The answer comes back fast, confident, tidy. One problem: the AI stopped reading at row 30,000, and it didn't make a fuss about it.
So the summary looks complete, the boss gets a number built on 60% of the data, and nobody lied. The machine simply has a meter, and nobody told you about the meter.
The meter is called tokens.
What a token is
A token is the unit of text an AI reads and writes, roughly three-quarters of a word. A 10-page report is about 4,000 to 5,000 tokens. Think of it like the XYZ envelope on a CNC machine: a part that fits gets machined properly. A part that doesn't fit gets cut short, or needs multiple setups.
Every AI model has a context window, the total text one conversation can hold, your uploads and its replies together.
That last bar is why I could drop 334,000 rows of WIP data into one conversation and have the AI cross-reference all of it. In a smaller window, the same job becomes chunks, and chunks can't see each other.
What it costs you
Two currencies. Money: on pay-per-token tools, a fat upload costs more than a short question, because input and output are billed separately. And quality: in a long conversation, the oldest tokens fall out of the window first. That's why a two-hour-old chat starts "forgetting" what you said at the beginning. The early text fell out of the envelope.
- One page of text ≈ 500 tokens
- A 10,000-row spreadsheet ≈ 500,000 to 800,000 tokens
- Hitting the limit? Split by customer or by year, so each chunk is logical
- Every big new analysis gets a fresh conversation.
Ah Hock's rule
Know roughly what your file weighs before you feed the machine. On a $20/month plan, normal office work never hits the meter, but massive exports do, and the AI won't always warn you. So check the row count in its answer against the row count in your file. Ten seconds, and you'll never present a 60% summary as a whole one.
