Coaching versus blocking: how should AI controls respond?

What are the three stances?

• Block: the action does not happen. Clear and safe for the few actions that should never run, such as an agent reading credential files outside an approved flow.
• Warn or ask: the person is told why the action looks risky and decides, or confirms that it should go ahead. Bay calls this Ask, Harmonic Security calls it warn, Onyx Security lists both ask and steer.
• Watch: nothing is interrupted, but everything is recorded. Bay documents a Simulation Mode for new rules; Harmonic Security lists log as an action.

Why does blocking alone tend to backfire?

When people meet a block on a tool that helps them, many of them find another tool, another browser or another device. The use does not stop; it moves somewhere you cannot see. That is the same shadow AI the control was meant to reduce. A warning that explains the reason, and points to an approved alternative, keeps the work in view.

Blocking also has a cost in developer time. A coding agent that is stopped without explanation halfway through a task leaves the developer guessing. Asking first, with the exact command shown, lets them decide in a second.

When is blocking the right start?

When the data or systems in reach are sensitive enough that a single wrong action matters more than friction. Production credentials, regulated data and destructive commands are the usual cases. Even then, the list of blocked actions should be short and specific, and everything else should ask or warn.

How do you move from watching to enforcing?

1. Run new rules in a watch-only or simulation mode for a week or two with a pilot team.
2. Review what would have been blocked or asked, with the people affected.
3. Turn the rules that fired rarely and rightly into asks, and the few that must never run into blocks.
4. If people approve an ask almost every time, consider moving it to allowed.
5. Publish an exception path: who to contact when a rule stops legitimate work.

Which products describe which responses?

The finder checks three responses separately: block, warn or ask, and watch-only or simulation. Fully matching warn or ask on the pages we read: Akamai Workforce Protector, Bay, Ent, Harmonic Security, Neo and Onyx Security. Fully matching a watch-only or simulation mode: Bay, Harmonic Security and Neo. Many more describe blocking. The line-up page shows all three for any set of products.

Whichever products you shortlist, ask to see the person's side of each response during the pilot: the wording of a warning, the information shown with an ask, and what happens when someone declines. Those few seconds decide whether people keep using approved tools or start looking for ways around them.

Sources