Documentation

Last updated: 12 October 2026 · Evidence Check AI for monday.com

Installing

  1. Install Evidence Check AI from the monday marketplace.
  2. Add the board view Evidence Check to any board that carries requirements with attachments.
  3. The first time you press Assess now, the view asks you to Connect to monday. Approve the read permissions there. That grants us a token to read board content and to download the attachment being judged — nothing is written back to your board. You are sent back to monday afterwards.

How an assessment runs

  1. An item with an attachment and a requirement is triggered — by an automation, or by picking the item and its two columns under Assess now in the board view. The automation needs a monday Standard plan or higher; the button works on every plan.
  2. The requirement is broken down into checkable conditions. If it cannot be broken down — "provide suitable evidence that things are in order" — no assessment is made and the model is never called: you are told to rewrite the requirement instead. A verdict nobody can retrace is worse than no verdict.
  3. The attachment is fetched, downsized if it is larger than needed, and sent to the model together with those conditions.
  4. The model only observes. It describes what the evidence shows and returns the requested values; it is not allowed to return a verdict, and an answer containing one is rejected.
  5. A deterministic rule engine reaches the verdict by comparing each condition against what was observed. The same observation always produces the same verdict, and the reasoning you read is that comparison — not a story written next to a conclusion.
  6. Nothing is charged twice: identical bytes against an unchanged requirement reuse the earlier verdict.

The audit panel

The board view shows the full chain per item: model id, prompt version, the moment of assessment, confidence, reasoning, and every override underneath it. This is the part a native AI column does not have — that one writes a value and overwrites it on reassessment.

Overruling a verdict

Account members can overrule a verdict, with a mandatory reason. Guests and viewers cannot: an override is an audit action, and it is recorded with the monday user ID of the person and the time.

Exporting

The board view exports the log as CSV, per board and over a period you choose. That file is the artefact you hand to an auditor. Cells are neutralised against spreadsheet formula injection before they are written, because part of the text in there came from your board and from the model.

Retention

A verdict does not stay forever. The default retention period is 24 months, counted from the most recent assessment of the same file and requirement. If Assess now re-confirms an earlier verdict, that earlier verdict and its override history are kept for as long as the re-confirmation, and then all of them are deleted automatically. We can set it to any value between 3 and 84 months, either on request or through the app's API. There is no settings screen for it yet, so today it is a deliberate step and not something that changes by accident. The same period applies to the usage record we keep for billing, which holds the board, item and column the call was made for — it is deleted on the same clock, not kept longer. See the privacy page for what happens when you uninstall.