Support AI is sold on capability. What decides whether a merchant actually connects their WhatsApp number is the opposite list — the things the software will refuse to do even when it could.
Here is ours, and the reasoning behind each one.
Never Invent an Answer
The agent answers from your store data and your written policies. If the answer is not in either, it says it does not know and hands the conversation over.
This is the constraint people underestimate. A language model asked about your return window will produce a confident, well-written, plausible number whether or not it knows yours. The customer believes it, acts on it, and you inherit a commitment you never made. Grounding is not a quality feature; it is the difference between a support tool and a liability.
The useful side effect: every unanswered question gets logged. You add the missing policy once and it is handled from then on.
Never Block a Customer From a Human
"I want to talk to someone" is an instruction, not an objection to be handled. Every deflection-optimised support bot is built to treat it as the latter, because deflection rate is what the vendor reports.
Our agent hands over on request, immediately, and says so plainly. It also hands over unprompted on frustration, on complaints, and on anything sensitive. The handoff carries the order, the thread, and any photos, so the customer does not repeat themselves and the merchant does not start from "which order?"
Never Send Marketing
The agent replies to customers who wrote to you and confirms orders they placed. It does not broadcast promotions, it does not upsell inside an open complaint, and it does not touch your contact list.
Partly because it is obnoxious. Mostly because your WhatsApp number carries a quality rating, and marketing blasts are the fastest way to lose it. The number you would burn is the same number your customers use to reach you about real orders.
Never Decide a Refund Alone
Refunds, chargebacks, disputes, anything with a money or legal consequence — the agent gathers the context and stops. It can pull the order, collect the photos, and record what the customer says happened. The decision is a human one, every time.
There is no setting that changes this. A configuration flag that lets software issue refunds unattended is a configuration flag that eventually gets switched on by someone in a hurry.
Why This Is the Product
An agent that answers everything is a liability. An agent that knows the edge of its own knowledge and stops there is a colleague. The escalation rules are not a safety wrapper bolted onto a chatbot — they are the thing that makes it safe to point at a live customer channel at all.
Ask any support vendor for their version of this list. If they only have the capability list, that tells you what they optimised for.