How accurate is Ella and how does Ella handle errors?

Ella gives every field it extracts (like customer name, SKU, quantity, ship-to, and price) a confidence score, colour-coded high, medium, or low. Low-confidence fields are flagged for the rep to review before submission. High-confidence orders can be set to auto-submit straight to the ERP.

» The three accuracy checks

1. Confidence Scores On Every Field: Every extracted value is scored before it reaches the rep. The rep sees uncertainty before the order goes through, not after. High-confidence lines are ready to submit. Medium and low-confidence lines are highlighted for review.

2. Source Traceability: Clicking any field on the draft order shows exactly where Ella pulled that value from on the source PO; the relevant section of the document is highlighted. There is no guessing about where a value came from. If something looks wrong, the rep can check the source in seconds.

3. Pre-Submission Validation: Before pushing to the ERP, Ella runs a final check on key fields (like customer, items, and prices) against the PO one more time. Anything that doesn't match is flagged for human review. No order bypasses this check.

» Where Does This Data Come From?

Every field on the draft order carries a lineage flag showing how that value was determined:

• AI-Extracted: Pulled directly from the source PO
• Defaulted: Filled from the buyer's saved profile (address, pricing, preferences)
• Manually Edited: Changed by the rep on the draft

This makes every order fully auditable. For compliance reviews or dispute resolution, the lineage flag shows exactly what Ella did and what a human changed — with no ambiguity.

» Things to keep in mind

• Auto-submit for high-confidence orders is configurable per tenant. If your team prefers a human review on every order regardless of confidence, that can be set during onboarding.
• Confidence thresholds (what counts as high, medium, or low) are also configurable. Your implementation contact can adjust them based on your tolerance for automated processing.