How do I fix a misread vendor name?

When to fix at the document vs at the vendor master record

There are two places to correct a vendor name. Both feed into the same data model, but they have slightly different ergonomics.

• At the document (the flow this article is about) — the right place when you see a wrong vendor name on a specific invoice. One click, one edit, the vendor learning kicks in and future invoices benefit. This is the path 90% of corrections take.
• At the vendor master record (Trading Partners → vendor → Edit) — the right place when the canonical info for a vendor is wrong (their legal name has changed, the VAT ID is wrong on the master record, the bank IBAN needs updating). Edits here propagate to all past + future invoices from that vendor.

When in doubt, edit at the document. The system propagates upwards automatically.

The walk-through

Step 1 — open the document

From the Documents list, the in-app notification when extraction finishes, or the search bar (search by partial vendor name or invoice number). The document detail page shows the original on the left and extracted fields on the right.

Step 2 — click the vendor field

The Vendor field is near the top of the right pane. Clicking it converts the label into a text input. As you start typing, an autocomplete dropdown shows existing vendors in your Trading Partners list — both exact matches and fuzzy matches (so typing "Hetz" matches "Hetzner Online GmbH" even if you have it stored under a slightly different exact name).

Step 3 — type the correct name

Two paths:

• Existing vendor: pick one from the autocomplete dropdown. The document is now linked to that vendor's master record, inheriting the vendor's canonical name, address, VAT ID, default category, etc.
• New vendor: just type the name (no match in autocomplete) and press Enter or click outside. A new vendor record is created in Trading Partners, the document links to it, and you'll be prompted to optionally fill in the rest (address, VAT ID) — though you can skip and add later.

The autocomplete is fuzzy-tolerant. If you've stored "Stripe Inc" and you type "stripe", the dropdown matches. If you type "Stripe" with different capitalisation, the existing record is matched (not duplicated).

Step 4 — save

Tab to the next field, click outside the field, or press Enter. The change is saved immediately. Three things happen simultaneously:

1. The document now shows the corrected vendor name with a small corrected marker on that field.
2. The vendor record is either created (new vendor) or updated (existing vendor) in Trading Partners.
3. The per-vendor learning is recorded — the system has now seen the correction and will use it on future similar documents.

The document is now in a manually-corrected state for that field, distinct from auto-extracted with high confidence. The audit log captures the change with your user attribution and the before/after values.

What happens after your correction — the three propagations

The single click triggers three updates in parallel:

1. This document

The vendor field on the current document is now your value, marked as manually corrected. Future re-scans of this document do not overwrite the manual correction (corrections are sticky; re-scan only updates fields that haven't been touched manually).

2. The vendor master record

The vendor in Trading Partners gets created (if new) or updated (if it already existed). If existing, the canonical name on the master record stays as it was (your correction on one document doesn't rename the master record); to rename the master record itself, edit at Trading Partners → vendor → Edit. If you typed a slightly different variant of an existing vendor, the system links to the canonical record without renaming it.

3. Future invoices from the same source

The "same source" detection uses multiple signals: visual style of the document, email sender if forwarded, page hash, font and address format, layout positions of key fields, vendor logo. It's substantially more robust than text-only matching. The next document where these signals match the corrected document picks up your corrected vendor automatically, with a vendor memory badge to show it was inferred from prior data rather than a fresh extraction.

After 3–4 corrections per vendor template, the system is reliably catching that vendor's invoices. Vendors with stable templates converge faster (1–2 corrections); vendors with quirky one-off invoices need more.

When the AI doesn't generalise

The learning works best when invoices are visually similar across the vendor's invoices. Cases where it doesn't generalise as well:

• Vendor changes their invoice template. New design, new logo, new typeface. The signal mix changes; the system treats new-template invoices as a partially-new vendor. Correct a few of the new template and it picks up the link.
• You only have 1–2 corrections for the vendor. The system is cautious about over-fitting on small samples. The third correction usually closes the loop.
• The misread was in the layout, not the OCR. E.g. the AI picked up the bill-to address instead of the from address, or read a sub-vendor's name on a multi-vendor consolidated invoice. Layout mistakes don't generalise the same way OCR mistakes do.
• The vendor uses very different invoice styles for different products / services (e.g. one for hosting bills, another for support contracts). Each "style" is effectively its own template from the system's perspective.

In any of these cases, correct each occurrence; the system will catch up.

Editing the vendor master record

Beyond per-document corrections, you can edit the vendor itself:

1. Go to Trading Partners in the left nav.
2. Find the vendor and click to open.
3. Edit the canonical name, address, VAT ID, default category, default expense account, bank IBAN. These propagate to all future invoices from this vendor and update the display of past invoices.

The Trading Partners view also surfaces:

• All invoices from this vendor (both manually-linked and auto-detected)
• Recurring-pattern detection (see recurring vendor patterns)
• Total spend in your base currency over different time ranges
• Vendor's IBAN — used for the matching-pipeline IBAN+amount shortcut

Edits at the master record are audited the same way as document-level edits.

Troubleshooting

AI keeps splitting one vendor into two ("Stripe Inc" and "Stripe Payments Europe Ltd"). These are technically two different legal entities (different VAT IDs, different addresses), and the system creates separate records by default. But for your books they might be one. In Trading Partners, open one of the two, click Merge with another partner, pick the other. The merge keeps all invoices linked to the surviving record; the merged-away record's name becomes an alias so future invoices from either name link to the same master.

AI auto-created a vendor that's just garbled OCR text. Open the vendor in Trading Partners, click Delete. Then re-edit the original invoice to point to the correct vendor (or create a new one). Deleting a vendor that has invoices attached re-links those invoices to "Unassigned" — you'd need to re-assign them, which is one click per invoice.

I corrected the vendor, but the VAT number on the document was also wrong — was that picked up? Partially. Editing the vendor field doesn't propagate to other fields automatically; you'd need to also click the VAT-ID field on the document. The Trading Partners record stores VAT, address, bank details — once they're right at the master record, future invoices pick them up.

The autocomplete dropdown shows weird old entries from when I was testing. Clean up Trading Partners. Open each unwanted entry and delete it (re-linking its invoices first, or marking those invoices as Unassigned).

Two team members corrected the same field at the same time. Last-save-wins. Both edits appear in the audit log. We don't surface a "another user is editing this" warning currently; for high-coordination teams, the convention "one bookkeeper" minimises conflicts.

The vendor on the invoice is in a different language script (e.g. Cyrillic, Arabic). Type the correct name in whatever script you want — the system stores Unicode and matches consistently. The autocomplete handles non-Latin scripts; the matching uses character-level fuzziness so transliterations also work.

I want to bulk-rename a vendor across many documents. From Trading Partners → vendor → Edit → Rename. The rename propagates to all past invoices linked to this vendor and to all future ones. If you instead want to bulk-rename via document-level edits (because the documents are linked to different vendor records that should all be one), do the rename via Merge first.

Related

• What the AI reads, and how to correct it — the broader correction flow
• Understanding confidence scores — what the score on the vendor field means
• Recurring vendor patterns — how vendor memory feeds into recurring detection