An invoice-extraction pipeline shows 97% field accuracy overall, and the team plans to remove human review entirely. Inputs include standard invoices, credit notes, and handwritten receipts. What is the best first step?

Answer: B — Measure accuracy by document type and field, and confirm each segment meets the bar

An aggregate can mask poor performance on specific document types or fields. Validate by segment before reducing review. (A) trusts the aggregate. (C) uses uncalibrated confidence and still skips segment validation. (D) makes the aggregate more precise without revealing weak segments.