The most common question we get from prospective clients isn't "can your agent handle X" — it's "how do I know it won't handle something it shouldn't." That's the right question. Deciding where the agent's authority ends is a business decision with real consequences, and we treat it that way, not as an afterthought bolted on after the conversational logic is built.
Why this isn't a purely technical choice
Technically, an agent can be given very broad permissions — issuing refunds, modifying orders, applying discretionary credit. Whether it should have those permissions depends on your risk tolerance, your margins, and how much a wrong action would actually cost you, none of which is a technical question. A high-margin brand with generous return policies can afford a more autonomous agent than a thin-margin operation where an incorrectly issued refund meaningfully hurts the bottom line.
We walk through this explicitly during setup: for each category of action, does the agent handle it independently, handle it with a confirmation step, or always hand off to a human? That mapping becomes part of the build spec, reviewed and approved by the client before launch, not a default we quietly ship.
What good handoff actually looks like
A bad handoff drops the customer into a queue with no context, forcing them to repeat everything they already told the agent — which reads as the automation having wasted their time rather than helped them. A good handoff passes the full conversation, the customer's apparent intent, and a note on why the agent is escalating, so the human picking it up can continue the conversation instead of restarting it.
We also design for graceful escalation on tone, not just capability. If a customer is clearly frustrated, the agent is built to recognize that and offer a human handoff proactively, rather than waiting until it hits a hard technical limit. Sometimes what a frustrated customer needs most is simply knowing a person is now involved, even if the agent could have technically resolved the issue.
Where the line usually ends up
In practice, most clients start conservative — letting the agent handle information and low-risk actions, keeping anything involving money or account changes with a human or a confirmation step — and expand the agent's authority over time as trust builds and the transcript history shows it's making good calls. That's a healthier default than starting broad and pulling permissions back after something goes wrong.
Revisiting the boundary over time
The permission map we agree on at launch isn't permanent. We review it against real transcripts at regular intervals, looking specifically at cases where the agent escalated something it might have been able to safely handle, and cases — more rare, but more important to catch — where it came close to acting outside its intended scope. Both directions inform whether a boundary should shift.
Clients tend to expand agent authority gradually as this review process builds confidence, rather than all at once. That pace is set by the client, not by us, because they're the ones who ultimately carry the consequence of a boundary set too loosely.