Know when to trust, verify or step in.
Every AI output comes with a confidence score from 0 to 100%. Use it to watch quality, set thresholds, trigger extra validation and send low-confidence results to human review.
A score beside every value.
Each extracted field returns its value, its source on the page and a confidence score. A high score passes on its own. A low score tells you where to look before anything ships.
Northwind Supply Co.
Extracted fields
const run = await client.extract({
file: { url: "https://example.com/invoice-4471.pdf" },
schema: invoiceSchema,
citations: true,
});
// Pass what scores high, send the rest to a person.
const THRESHOLD = 80;
const review = Object.entries(run.output?.citations ?? {})
.filter(([, cited]) => cited.some((c) => c.confidence < THRESHOLD))
.map(([field]) => field);
if (review.length) await sendToReview(run.id, review);Trust, verify or step in.
Set the bands once and let the score decide where each result goes. Thresholds can apply per field or per workflow.
Trust
Results above your threshold move on with no one in the loop.
Verify
Results in the middle band trigger extra validation, such as a rule check or a match against another document.
Step in
Low-confidence results pause for human review with Approvals, the source shown beside the value.
vendor 99% pass total 99% pass due_date 84% verify against the PO ship_to 62% human review
Questions about Score.
How sure the model is about one output, from 0 to 100%. It is a signal for routing, not a guarantee. Pair it with citations so a reviewer can check the source in one look.
See the score on your own file.
Run a cited extraction with your $15 of free monthly usage and read the score on every field.