Every AI question is billed by how much the model has to read. The usual “AI over your data” makes it read your records — so the more you keep, the more each question costs, and past a point it can’t hold it all. North Star sends the model the question, never the records. Same price at ten thousand rows or ten million.
Because one approach reads your records and the other reads your question, the two bills scale differently — and the gap between them widens with every record you add.
Illustrative — assumes ~50 tokens per record and frontier‑model pricing (~$3 per million words in). Log scale. The point isn’t the exact dollar — it’s the shape: theirs climbs with your data and eventually can’t hold it; ours doesn’t move.
The same question over a real job‑shop node of 9,412 records. To answer it correctly you must consider all of them — a global ranking can’t be sampled.
| Ship the data to the AI | Send the question | |
|---|---|---|
| What the model reads | ~470,000 words every record, every question | ~480 words the field names + your question |
| What it writes | the whole answer | ~50 words — the query |
| Cost of one question | ~$1.40 | ~$0.0024 |
| Records you ask over | Ship the data | Send the question | The gap |
|---|---|---|---|
| 10,000 | ~$1.50 | ~$0.0024 | ~600× |
| 100,000 | ~$15 | ~$0.0024 | ~6,000× |
| 1,000,000 | ~$150 | ~$0.0024 | ~60,000× |
| 10,000,000 | won’t fit — must sample | ~$0.0024 | can’t answer “all” |
Past a few hundred thousand records the read‑it‑all approach can no longer hold your data at once — it retreats to reading a sample. A sample can’t count, rank, or prove across all of it. So you’re paying more for an answer that’s also less complete. Ours stays flat — and exact.
The read‑it‑all approach keeps a second, prepared copy of everything for the model to search — built, hosted, and rebuilt every time your data changes. Your box already holds it all; nothing to prepare.
Every question ships records out of your building — metered on the way out, and a liability once they land. The question leaves; the records don’t.
It re‑sends and re‑reads your data on every single question — nothing is ever reused. Your box answers from what it already holds, in place.
Tell us what you keep and what an answer costs you today. We’ll price the same questions both ways on a slice of your data — and you check every figure.
Go deeper · AI over your data · Why North Star · The Commons