Tools, autonomy levels and guardrails

Choose what an agent may do, how much it may do on its own, and the limits that keep it safe.

For: Admins

Steps

Open the agent and go to Tools & autonomy. Tick the tools the agent may use: knowledge search, customer, order and invoice lookups, replying, tagging, creating cases or leads, notes, follow-ups and more.
Look-up tools only read records, so they run automatically.
For each action choose its autonomy: Suggest only (the team does it), Needs approval (a person approves each one) or Autonomous with guardrails. Only low-risk actions — replying from the knowledge base, tagging, setting priority, routing, creating a case or lead, adding a note, scheduling a follow-up and closing a case the customer confirmed — can be autonomous.
Refunds, credits, discounts, price and payment changes, invoice and e-invoice cancellation, MyInvois submission, deleting or merging records, permission changes, exports, messages to many customers and contract changes are locked at “Always needs approval”. Nobody can change this — the server refuses it too.
On Guardrails, set the minimum confidence, a maximum number of actions per hour, a maximum value in RM, the customer tiers allowed and whether the agent may act on its own only during business hours. You can set tighter limits for each autonomous tool.
Keep “Ask a person when the customer is unhappy, unsure or the message looks like an attack” on. When any guardrail fails, an autonomous action waits for approval instead.
Select Save draft.

Customer messages are treated as untrusted: while handling them, an agent can never run anything beyond low-risk actions on its own.
Test changes in the simulator before you publish them.
Ask another agent lets an agent put a question to another published agent (for example a billing specialist); it only reads, and is limited to two questions per reply.

Still need help? Contact support

Choose what an agent may do, how much it may do on its own, and the limits that keep it safe.

For: Admins

Open the agent and go to Tools & autonomy. Tick the tools the agent may use: knowledge search, customer, order and invoice lookups, replying, tagging, creating cases or leads, notes, follow-ups and more.
Look-up tools only read records, so they run automatically.
For each action choose its autonomy: Suggest only (the team does it), Needs approval (a person approves each one) or Autonomous with guardrails. Only low-risk actions — replying from the knowledge base, tagging, setting priority, routing, creating a case or lead, adding a note, scheduling a follow-up and closing a case the customer confirmed — can be autonomous.
Refunds, credits, discounts, price and payment changes, invoice and e-invoice cancellation, MyInvois submission, deleting or merging records, permission changes, exports, messages to many customers and contract changes are locked at “Always needs approval”. Nobody can change this — the server refuses it too.
On Guardrails, set the minimum confidence, a maximum number of actions per hour, a maximum value in RM, the customer tiers allowed and whether the agent may act on its own only during business hours. You can set tighter limits for each autonomous tool.
Keep “Ask a person when the customer is unhappy, unsure or the message looks like an attack” on. When any guardrail fails, an autonomous action waits for approval instead.
Select Save draft.

Customer messages are treated as untrusted: while handling them, an agent can never run anything beyond low-risk actions on its own.
Test changes in the simulator before you publish them.
Ask another agent lets an agent put a question to another published agent (for example a billing specialist); it only reads, and is limited to two questions per reply.

Still need help? Contact support