AI features

Why a tender was matched or unmatched

Every evaluated tender in Tender Navigator carries the AI's reasoning next to its verdict. The reasoning refers to the filter's prompt as it was at evaluation time, which is stored with the result, so the answer to why a tender was matched or unmatched is always on the result itself.

Last updated

Open the reasoning

  1. Open Results in the dashboard.
  2. Click the tender. The Tender Info panel opens.
  3. Under Evaluation completed, read the verdict and the reasoning. The reasoning is the model's own explanation of how the tender compares with the criteria in your filter's prompt.

The same reasoning is in the digest email for that tender and on a shared tender page.

Check what it was judged against

Further down the panel, Filter info shows the filter as it was when the tender was evaluated: keywords, countries, codes, deadline offset, budget rule and the LLM model. Editing a filter later never rewrites this snapshot, so a verdict from three weeks ago is explained by the prompt of three weeks ago.

Common reasons for a verdict you did not expect

  • The prompt said less than you meant. "IT services" matches a lot; the model can only apply the criteria you wrote. Add the concrete signals and exclusions the reasoning missed.
  • The prompt said more than you meant. A requirement phrased as a hard rule ("only tenders above EUR 1 million") is applied as one. Soften it, or say how to treat uncertainty.
  • The notice is thin. Some portals publish a title and two lines. The model works from the text Tender Navigator ingested; it does not open attachments. See Why some tender fields are missing.
  • The keywords selected the wrong tenders. A keyword such as "cloud" also appears in weather and printing notices. Unmatched is the correct verdict there; tighten the keywords or codes so the filter reads fewer irrelevant notices and spends fewer evaluations on them.
  • Two filters disagree. Each filter has its own prompt, so the same tender can be Matched by one and Unmatched by another.

Change future verdicts

Edit the filter's AI prompt and save. The change applies from the next daily run. Existing results are never re-evaluated. See Writing the AI prompt for how to sharpen it, and Too many irrelevant tenders for the most common fixes.

Ask the assistant

The dashboard AI assistant can help you read a reasoning and rewrite the sentence in the prompt that produced it.