How we stopped drowning in "where's my order?" tickets (without a bot making promises it shouldn't)

AI Build Your Business
1

Sharing something that finally worked for our support load, in case it helps someone.

We were getting buried in the same handful of questions — where’s my order, what’s your return window, did my refund go through. All easy, all relentless.

What worked was splitting every ticket into two buckets:

  1. Questions answerable straight from our own data and policies (order status, return window, stock). These are safe to automate — as long as the answers come from real order data and real policy docs, not a model guessing. This was about half our volume.

  2. Anything involving money, a judgment call, or an upset customer (refunds, damaged items, one-off exceptions). We let automation prep these — pull the order, draft a reply — but a human always hits send, and a human approves any refund.

The one rule we stuck to: automate the answers, not the decisions. It cut the repetitive load a lot, and we never woke up to a bot that refunded the wrong customer.

Curious how others here draw that line — what do you let automation handle on its own vs. keep human?

2

I think that’s a sensible approach. The biggest mistakes I’ve seen happen when businesses let automation make decisions instead of handling repetitive tasks.

For me, order tracking, FAQs, inventory checks, and policy-related questions are great candidates for automation. Refunds, exceptions, disputes, and anything involving customer satisfaction should still have human oversight.

“Automate the answers, not the decisions” is a good rule of thumb.

3

Exactly — the disasters almost always come from automating a decision, not from automating a task. One thing I’d add to your list: a confidence check on top of it. Even on the “safe” stuff (tracking, FAQs, policy), if the agent isn’t sure, it should hand off rather than guess — a confident wrong answer about a return window does more damage than a slow one. So the rule we landed on is really two layers: only automate facts (not decisions), and only when the agent
is actually confident. Everything else routes to a person.