Watch, coach, redact, block: choosing the first response

Watch logs the interaction and changes nothing, which gives you a baseline. Coach shows a message, such as a reminder that this tool is not approved for customer data, and lets the person continue or stop. Redact masks or removes the sensitive part of the prompt and sends the rest, so the task still gets done. Block stops the interaction entirely.

Watch fits new tools and the first weeks of a programme. Coach fits tolerated tools and grey areas, where you want people to think before they paste. Redact fits approved tools used with occasional sensitive data, such as a support team summarising tickets. Block fits unapproved tools for sensitive data, secrets and credentials, and agent actions that could cause damage, such as deleting data or sending it out.

Start in watch mode for a few weeks and look at what would have been blocked. Tune classifiers on the false positives. Turn on coaching for the categories people hit most, then redaction for approved tools, and block last, for the narrow cases where nothing else is safe. Announce each step before it happens.

The ActionStrip on each profile shows the responses a vendor states on its pages. On the pages we read, Cato states Monitor, Anonymize and Block; Fortinet states monitoring, coaching and blocking; Palo Alto Networks and Zscaler state coaching and blocking. Use the quiz's first-response question to weight this for your own case.

Source: Cato Networks: AI security · Read 7 Oct 2026

The message matters as much as the rule. A coaching message should say what was detected, why it matters and what the person can do instead, such as using the approved tool. A redaction should be visible, so people understand why an answer looks different. A block should name the policy and the place to ask for an exception. Teams that skip this step see the same pattern: support tickets rise, people switch to personal devices, and the security team loses the visibility the control was meant to give. Write the messages before you turn on the responses, and test them with a few users from each team.