How Keystone works, and why you can trust the numbers.
A short guide to what Keystone does, how to use it, and — since it's an AI-assisted tool — an honest look at where its accuracy actually comes from. See Why Keystone for the problem it solves.
Using it in three steps
- 1. Start a package. Click + New package on the dashboard, name it, pick the project, and drop in every bidder's quote.
- 2. Clear the confirmation queue. Open Review queue on a package, check each flagged field against its source page, and Accept or Edit it.
- 3. Finalize. Once every field is confirmed, the comparison table is trustworthy end to end — finalize the package to send it to estimator sign-off.
Where the accuracy numbers come from
These two charts are published benchmarks for AI document-extraction technology in general (2025–2026 vendor and academic sources) — hover a bar for its source. They explain two design decisions, not a claim about Keystone's own measured performance.
Extraction accuracy — digital vs. scanned quotes
This is why a scanned quote produces more confirmation-queue items than a clean digital one — it's expected, not a bug.
Automated extraction alone vs. with human confirmation
This is the actual reason Keystone never auto-accepts a flagged field — the jump from “automated” to “trustworthy” happens at the confirmation step, not inside the model.
Frequently asked
How accurate is Keystone's extraction, really?
Honestly: we haven't measured Keystone's own accuracy yet, because this build isn't reading real documents — it's a working prototype of the workflow, with realistic mock data standing in for what extraction would produce. The benchmark charts below are published, cited results for the same class of AI document-extraction technology, not a claim about this specific system. That gap is exactly why confidence scoring and mandatory confirmation are built into the core design rather than bolted on later.
What happens if Keystone gets a number wrong?
Nothing — until a human confirms it. Every extracted field carries a confidence score, and anything below the threshold sits in a confirmation queue next to the exact source page. A flagged field cannot enter the comparison table, and a package cannot be finalized, until every flag is cleared.
Where does the confidence score come from?
It reflects the extraction model's own certainty about a value — not a guarantee that the value is correct. That distinction matters: a confident-looking number can still be wrong, which is why low-confidence fields are flagged for a human to check against the source, not auto-accepted because the score looked high.
Does Keystone ever submit or finalize anything without a person approving it?
No. Finalizing a package only becomes available once every flagged field is confirmed, and it hands off to an estimator sign-off — Keystone doesn't publish or submit on its own.
What if a bidder's quote is scanned or a low-quality photo?
Keystone still attempts extraction, but confidence typically drops — see the digital-vs-scanned comparison below. More fields get flagged for confirmation on a scanned document, by design, rather than the system quietly guessing.