The assistant can now be switched off without a release

Until now, stopping the assistant meant shipping a code change or pulling the model provider’s credential. Both are slow, and the second takes the assistant down for everyone with no record of who did it or why — the worst way to answer a bad prompt publish, a runaway spend spike or an incident at the provider.

The console now carries the switch itself, at two levels.

Platform-wide. A new Assistant switch page states plainly whether the assistant is on or off for everybody, and turns it either way. The platform switch wins: while it is off, every workspace’s assistant is off, whatever its own setting says. Turning it back on restores exactly the workspaces that were not switched off individually — nobody is quietly re-enabled.

One workspace at a time. A workspace’s detail page gains an Assistant card. It says which switch, if either, is holding that workspace’s assistant off, names the platform setting alongside it so nobody switches off a workspace that was already off, and flips the workspace’s own switch.

Both levels take effect on the next request. There is nothing to deploy and no cache to wait out.

Every flip is on the record. A reason is required — a flip without one is refused and changes nothing — and it is written to the activity trail beside the person who made it and the time they made it. Six months later, “who turned the assistant off in March” and “why” are both answerable.

While the assistant is off, the surfaces that use it say so and put their message box out of use, rather than accepting a message that would only be refused: the assistant itself, the requirements qualification conversation, and the setup concierge.

Two things worth knowing

If the setting cannot be read for any reason, the assistant keeps answering. An outage of the off-switch must never be able to do what the off-switch does.

This switch is the conversational assistant, whole. Expert matching and scoring use AI models too, and they are not affected by it — during an incident, turning the assistant off does not stop every model call the platform makes.