Reviewing AI Extractions
How to review, trust, and correct AI-extracted data in Muin — confidence scores, what they mean, and the habits that keep extracted data audit-ready.
Muin extracts structured data from your documents in several places — invoice fields, organization facts for the Data Vault, document metadata. Extraction is fast; your review is what makes it trustworthy. This page covers how to read what the AI gives you and the habits that keep extracted data audit-ready.
Confidence Scores, Plainly
Every extracted value carries a confidence score — the model’s own estimate of how sure it is.
| Confidence | How to treat it |
|---|---|
| High (90%+) | Usually right; spot-check totals and identifiers |
| Medium | Worth a look against the source document |
| Low | Treat as a draft — verify before anything depends on it |
Confidence is a prioritization tool, not a guarantee: a high-confidence wrong value is rare but possible, which is why amounts and IDs that drive payments deserve a glance regardless.
Where Review Happens
- Invoices — the invoice detail page marks AI-extracted values and shows their confidence. Edit any field to correct it; your value wins. See AI Invoice Extraction.
- Data Vault — extracted organization facts sit as Pending Review until someone verifies them at Settings → Data Vault. Verification records who confirmed the value.
Review Habits That Pay
- Check the few fields that matter most — totals, dates, account and tax identifiers. A wrong description is cosmetic; a wrong total is a financial error.
- Compare against the source, not memory — the original document is one click away; use it.
- Correct in place — fixing the extracted value (rather than working around it) means reports, approvals, and exports all inherit the fix.
- Make low-confidence review a routine — a weekly pass over pending/low-confidence items keeps the backlog at zero and the data trustworthy.
Why Corrections Matter
Corrected values take precedence over machine readings everywhere downstream — approval rules, analytics, filings. Reviewing isn’t bureaucracy; it’s the step that converts fast AI output into data your auditor (and your board) can rely on.