Screening analytics
What the check before an AI model found, held back and let through, by kind of data, agent and model, over up to a year. Part of Screener Pro, on Enterprise.
Analytics → Screening sums every check your agents ran before sending text to an AI model: how many there were, how many were held, and what was in them. It is built from the counts the audit trail already keeps, so it never shows, and Deliverd never holds, a value that was found.
What the page shows
| Figure | What it counts |
|---|---|
| Checks | Every check run with the SDKs, the CLI, the MCP tool screen_data, the Deliverd Screen proxy or Copilot Studio |
| Held | Checks that were refused, redacted or put to a person, or would have been in report-only mode |
| Secrets stopped | Secrets and credentials in checks that were refused |
| Items redacted | Items replaced with a placeholder before the model saw them |
| Put to a person | Checks that became an approval before the text could go |
| What enforcing would change | Report-only checks that would have been held, by what would have happened |
Below the totals: checks and held checks per day, what was found by kind, the models the text was going to (by the AI register record each check named) and the agents and callers that ran the checks, with the most held first. Choose a window of 7 days to a year; Print saves it as a PDF for a risk committee or your DPO.
Owners and administrators, agent admins, security admins, compliance admins and auditors can open it.
Try it on your own agent.
Free to start, no card. Sign up with Google, Microsoft or GitHub, and the first page after hands you an API key.