Agents and the inbox: what to automate and what to never automate
Not every thread should reach an agent unsupervised, and pretending otherwise is how automation projects earn a bad reputation internally. The useful question is not whether to automate the inbox. It is where to draw the line, and how to make the line hold in configuration rather than in hope.
Here is a practical way to draw it.
Automate the knowable
If the right answer is a lookup, automate it. Order status, delivery windows, plan changes, refund eligibility under a clear policy, appointment confirmations. These are questions where being right is a matter of having the information, and the agent has the information.
This is the bulk of most inboxes, and handing it over is almost pure upside. The replies get faster, they get more consistent, and your team stops spending its day on work nobody was hired to do.
Never automate the load-bearing
Some threads carry weight out of proportion to their length. A cancellation from your largest account. A message that mentions a lawyer, a regulator, or a death. A complaint that has already escalated once. Anything where a confident wrong answer does lasting damage.
These should never go out without a person, no matter how good the draft looks. The rule is not about the agent's accuracy. It is about the cost of the rare miss, and on these threads the cost of a miss is too high to average away.
Enforce it in configuration, not in trust
A line you hold by asking people to be careful is not a line. Scope each agent to specific queues and specific actions, set the approval thresholds explicitly, and route the sensitive categories to a human by rule. The agent should be structurally unable to send the things you decided it should not send.
This is the difference between a policy and a wish. A wish depends on everyone remembering. A policy is in the configuration, and it holds at three in the morning when no one is remembering anything.
Revisit the line as trust grows
The line is not fixed forever. As the log builds evidence, categories move from held to automatic, one at a time, on the strength of what you have actually seen. The point is that they move deliberately, with a reason, not because someone got impatient.
Draw the line in configuration, watch the log, and move it on evidence. That is the whole discipline, and it is what separates an automation you trust from one you quietly turned back off.
Where this tends to go wrong
The failure mode is almost never the agent inventing something wild. It is the small, plausible miss: a reply that is correct in general but wrong for this one account. That is exactly what the approval step and the log are for, and it is why we tell teams to read the log before they widen scope.