Screener Pro: policy engine, Purview and the Screener

Your own screening rules for each agent, group and model, your own detectors, owned and dated exceptions, Purview sensitivity labels and Purview Audit, a review queue, and a Screener page for every member. On Enterprise.

Data screening checks text before it goes to an AI model, with one policy for the whole organisation. Screener Pro, on Enterprise, adds your own rules on top, for each agent, group and model, and the tools to run them safely. It is at Agents → Data screening, on the Policy engine and Review queue tabs.

The policy engine

PartWhat it does
RulesWhen the text holds some kinds of data, from a number of items, redact it, ask a person or refuse. Scoped to models in the AI register, agents and directory groups; none picked means everyone. A rule only ever makes a decision stricter. A rule can be report-only, to see what it would do first.
DetectorsYour own patterns, such as matter numbers or policy references, found case-insensitively and redacted like a confidential term. Patterns that could make screening slow are refused. They apply to text Deliverd reads: the API, MCP and the Screener page.
ExceptionsThe one way to let more through: a named reason, an owner and an end date, scoped like a rule. An exception can never cover secrets, and it stops applying the day it ends.
Purview labelsMap Microsoft Purview sensitivity labels to a classification and an action. When an agent sends the label with the text (sensitivityLabelId), a model not approved for that classification gets the action. Load the labels from your tenant, or add them by ID.
Policy as codeThe whole draft as JSON, to keep in your own repository and paste back. Saving checks every part and says what it left out.

Draft, test, publish

  1. Edit and save the draft. Nothing changes until it is published.
  2. Test the saved draft against the last 30 days of checks: how many would have been decided differently, which way, and how many each rule would have matched. Rules, exceptions and scopes are replayed from the counts the audit keeps; detectors and labels need the text, which was never kept, so they are not part of the test.
  3. Publish it with a note. Every version is kept with who published it and what changed; rolling back publishes an older version's rules as a new version.
  4. Set the mode. Shadow runs the published version beside every check and changes nothing; Analytics → Screening shows how often it would have decided more or less strictly. On makes its decision the answer. Off is where every organisation starts.

Microsoft Purview

With your tenant linked on the Data screening tab, Screener Pro can load your sensitivity labels and write each check to Purview Audit, so it appears in Activity explorer beside the rest of your Microsoft 365 AI activity: counts and the decision, never the text. A global administrator grants Deliverd's app the Microsoft Graph permissions SensitivityLabel.Read and ContentActivity.Write. Analytics → Screening adds how often Purview was asked and what it said.

The review queue

Every ask screening put in front of a person, waiting and decided, with who asked and the note the approver left, and every check refused outright. Each ask is an ordinary approval: open it to decide, and your approver rules apply.

The Screener, for everyone

Screener in the sidebar lets any member paste text before they give it to an AI tool, pick the tool from the AI register, and see the answer agents would get: fine to send, send the redacted version (with a copy button), ask a person, or do not send. The text is checked in memory and never stored; the check is recorded under the person's name, like an agent's.

Owners, administrators and agent admins can change the policy engine. Security admins, compliance admins and auditors can read it and the review queue.

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.