An AI agent that takes an order in chat is not the same job as one that answers a question. A wrong FAQ answer gets corrected in the next message; a cart closed on the wrong price, the wrong item, or a payment that never actually cleared ships. The fix isn’t turning commerce off for AI — it’s naming the exact checkout moments where the agent should stop and route to a person, before the order confirms.
The moments, not the confidence score
General escalation advice says hand off when the agent is unsure. A cart needs sharper triggers than that, because the agent can be perfectly confident and still be the wrong one to close the sale:
- A price or discount request. “Can you do 300 instead of 350?” is a negotiation, not a lookup. Even a small discount changes the bill your customer pays — that’s a human call every time.
- Anything outside the catalog. A custom bundle, a colour that’s out of stock, a substitute item — none of it is in the product data the agent reads from, so it shouldn’t guess a price or availability.
- A second failed payment on the same order. One failed transfer is normal — a blurry slip, a typo in the amount. Two in a row on the same cart is a pattern worth a person’s eyes, not a third automated retry.
- A cancellation or refund request mid-order. These reverse money that may already be moving. Route them the same way you’d route a payment slip that fails your verification checklist — hold and hand off, don’t argue in chat.
- A request for a tax invoice or company billing details. This touches your accounting, not just the order — a judgement call for a person, every time.
Write the rule where it happens, not after
The general framework for when any AI agent should hand off sorts signals into confidence, intent, and customer state. Commerce-in-chat needs the intent bucket built out specifically for checkout: pair each trigger above with an escalation rule at the point the AI agent would otherwise move to order confirmation, not a review pass afterward. In 1CS, that means routing the conversation — unassigned for the next free agent, to a specific person, or round-robin across the team — the moment the trigger fires, while the customer is still mid-cart and before any bill link goes out.
For teams that already coordinate over LINE, an uncertain cart doesn’t need a separate dashboard check: it can land in the same admin group used for other approvals, with the AI’s draft summary of what the customer wants attached.
What the human needs in the queue
A cart that lands in a human queue empty-handed costs you the goodwill the fast handoff bought you. Attach the same three things every time:
- The cart as the AI understood it — items, quantities, and the price it was about to confirm.
- Payment status — whether a bill link went out yet, and whether a slip has already been received and how it checked out.
- Why it escalated — which of the five triggers fired, in one line.
A human agent who opens the conversation and sees all three can usually finish the order in one reply, instead of re-asking the customer what they wanted.
A rule set to start with today
Pull the last 50 chat orders your team closed and check how many hit one of the five triggers above without ever reaching a person — a discount granted on the spot, a substitute item confirmed without a stock check, a second slip accepted without question. That count is your starting backlog for escalation rules, and it’s usually higher than teams expect once they look for it by trigger instead of by outcome.
None of this slows down the orders that are exactly what they look like: right item, right price, one clean payment. It just makes sure the ones that aren’t never close on their own.
