How Vendarity’s AI reads supplier documents
Vendarity can propose an uploaded supplier document’s key fields with page citations. How it works, how we tested it and what the AI cannot do.
In short
A workspace can turn on AI document reading when it starts, or later in Settings. While it is on, Vendarity reads each new document as it arrives and proposes ten fields from it, each with a page citation or an explicit “not found”. Dates whose day and month order is unclear are left for a person, and the reviewer resolves anything uncertain and decides. On 1 October 2026, the first extraction version matched 472 of 475 required fields on 100 synthetic documents.
A Vendarity workspace can turn on AI document reading when it starts, or later in Settings. While it is on, AI reads each new document as it arrives and proposes its key details, each with the place in the document it came from. The reviewer checks the proposals against the source and makes the decision.
What does the AI propose?
In the first extraction version (v1), the AI proposes ten fields, the same ones a reviewer records without AI: document type, legal entity, certificate or policy number, issuer, standard and version, site or scope, and the issue, valid-from, expiry and review dates. For each field, the AI either proposes a value with its page and the line or snippet it came from, marks the field as not found, or marks it as ambiguous.
How are citations checked?
Where a suggestion quotes the document’s embedded text, Vendarity checks that the cited lines contain the quoted snippet and the proposed value. If they do not, it withholds the suggestion. Dates must be supported by the snippet, and a numeric date whose day and month order is not explicit is left unresolved for the reviewer. A suggestion read from an image, such as a scanned page or a picture within a text page, is labelled as image interpretation. Its citation points the reviewer to the page image to check.
How did we test it?
Before switching it on, we ran a live evaluation on 1 October 2026 against 100 synthetic documents, 20 for each launch country. They covered varied business documents, missing fields and ambiguous dates. Our release bar was 95% exact match on required fields.
| Measure | Result |
|---|---|
| Required fields matched | 472 of 475 (99.4%) |
| All present fields matched | 972 of 975 (99.7%) |
| Ambiguous or absent fields left for a person | 25 of 25 |
| Incorrect non-empty suggestions | None in this test set |
The three misses were expiry dates the AI withheld even though the document stated them. A reviewer would type those dates in; no wrong date was proposed in this set.
On 3 October 2026 we evaluated a later extraction version (v2), which reads each policy on a certificate separately, on the same synthetic set. It matched 469 of 475 required fields (98.74%), and its six misses were also expiry dates it withheld rather than gave. Re-run on 24 October 2026, it matched 475 of 475, left all 25 ambiguous or absent fields for a person and made no incorrect suggestion. v2 is the version in use since the October 2026 release. Its prompt and checks were changed after reading failures on that same set, so its results are not independent estimates.
Neither result measures accuracy on your documents, and any later change to the extraction needs a fresh evaluation. That is why, in this release, a person reviews each document before it can be accepted.
How do you switch it on?
- On your new workspace’s first screen, leave “Let AI read each document and suggest its details” ticked. Later, an Owner or Admin can turn AI on or off in Settings → Organization, after reading what is sent to Google.
- Each new document is read once it passes its security scan, while the plan’s AI pages last. A reviewer can also ask for suggestions on a document that wasn’t read, such as one that arrived before AI was on.
- The suggestions appear beside the document. The reviewer checks them against the cited pages, resolves anything marked uncertain, corrects any wrong value with a reason, and confirms the rest together when accepting.
What can the AI not do?
In this release, the AI cannot:
- Accept a document or approve a supplier.
- Skip review: a person must resolve every required detail and validity date and confirm the displayed details before a document can be accepted; anything left unresolved is recorded as unresolved.
- Use tools: the model is given none, and document text is treated as data, not instructions. These controls reduce the risk of a document steering the AI; they do not guarantee that every manipulation is detected.
- Read more pages than your plan includes: 50 on Free, 150 on Starter, 500 on Core and 2,000 on Growth a month, with no overage charges.
What happens to the documents?
The page previews and embedded text of each document AI reads go to Google’s paid Gemini API, with stored interactions, search and model tools disabled. As checked on 13 October 2026, Google’s paid-service terms say Google does not use prompts or responses to improve its products.
The same terms say Google logs prompts and responses for a limited period to detect policy violations. They also say data may be stored briefly or cached in any country where Google or its agents have facilities. The original suggestions, their citations and your reviewers’ corrections stay in your workspace’s evidence history for its retention period. More detail is on the trust page.
Sources
- How Vendarity works · Vendarity
- Trust, data location and product limits · Vendarity
- Pricing and the AI pages each plan includes · Vendarity
- Gemini API additional terms of service · Google