Guardrails
The rules that keep agents inside a safe lane: what they are allowed to decide on their own, what they must never trust, and how quickly things stop.
Firm-wide rules
These apply to every agent, every client, all the time.
Untrusted content in documents
Text inside invoices or emails can never instruct an agent. Agents treat document and email content as data only, never as commands.
Output must tie to source
Numbers an agent reports must reconcile to the underlying source before they are shown or acted on.
Firm-wide pause
Stops every agent from starting new work across every client, immediately.
Per-agent thresholds and autonomy
The confidence threshold decides when a run routes to a human instead of finishing on its own.
Per-client pause
Stop every agent from working on one client, for example while an issue is being resolved.