When to let AI reply, and when not to

The decision is usually framed as on or off. In practice the useful question is narrower: which messages should be answered without anyone looking, and which should wait for a person.

Start in review mode

For the first week, have the assistant draft every reply and send nothing. You read the drafts, correct the ones that are off, and send the rest with a click. You get the answer written for you, the customer gets a fast reply, and nothing goes out that you have not seen.

That week also tells you something a demo cannot: what your customers actually ask, and how often the assistant gets it right.

Then automate the boring half

Opening hours, location, whether something is in stock, delivery areas, standard pricing — these repeat endlessly and have one correct answer. Automating them is where nearly all the time saving comes from, and the risk is close to zero.

Keep people on the parts that need judgment

• Negotiation and anything involving a discount.
• Complaints, refunds, and anyone who is already unhappy.
• Custom or unusual orders where the answer depends on stock nobody has checked.
• Any conversation where the customer asks for a human. Always honour that immediately.

Make the handoff sharp

The worst outcome is not a wrong answer, it is an ambiguous one: a customer told someone will call, while nobody was actually told. A handoff should flag the contact, stop the assistant replying, and put the whole thread in front of a named person. Half a handoff is worse than none.

whacl runs review mode or auto-send per channel, with escalation triggers you set, so you can automate the repetitive half and keep people where they are worth the most.