Guardrails are the safety checks every AI message passes through, on the way to the assistant and on the way back. They keep the assistant on the work it is meant for, keep credentials out of its replies, and mask the most sensitive personal details before anyone reads them.
What guardrails do
Guardrails screen both directions of a conversation: the message someone sends to the assistant, and the reply that comes back. The same checks cover the assistant and the builders that draft forms, reports, dashboards, and flow graphs.
They guard against a few distinct risks:
- On-task replies. Requests that fall outside what the assistant is built to help with are turned away, so it stays on your data and your tasks rather than being steered into unrelated territory.
- Resistance to manipulation. Attempts to make the assistant ignore its instructions, or to coax it into revealing them, are caught before they take effect.
- No secrets in replies. Access keys and similar credentials are blocked from appearing in a reply.
- Masked sensitive numbers. Payment-card numbers and similar high-risk identifiers are masked. Ordinary details the assistant needs to do its job, such as names and email addresses, are left in place, so it can still answer questions about your entries.
When a guardrail stops something, the assistant returns a short, plain decline rather than an error, and the flagged content is not stored. A card number pasted into a message, for example, never surfaces in a saved conversation title.
The checks screen the traffic whether the assistant runs on the platform default or on a provider your organisation brings, so the protections do not depend on which model answers.
Tiers
Guardrails come at more than one tier of strictness, from a more relaxed level to a stricter one. An administrator chooses which tier applies, alongside the provider and model choices on the AI settings screen.
The tiers differ in one thing: how tightly content filtering is tuned. A stricter tier screens more borderline content and gives the assistant less latitude; a more relaxed tier lets more through. Pick the tier that matches how sensitive your work is and how much room the assistant needs to stay useful.
The core protections hold at every tier. Keeping the assistant on task, resisting manipulation, blocking secrets, and masking payment-card numbers are the floor, and they apply whichever tier you pick. Moving between tiers tunes how cautious the content screening is, not whether these protections exist.
The guardrails settings list the tiers your organisation can choose from. Where your organisation has configured its own set of guardrails, that set is offered in place of the standard tiers; where none has been set up, the settings say so.