Build a follow-up that has a reason to exist.

Start with the decision that stalled.

Read the conversation before writing. Identify the unresolved point: price, timing, scope, trust, approval from another person or a missing technical answer. If there is no clear obstacle, there may be no good reason to follow up yet.

An AI agent needs the same context a careful salesperson would use. That includes the customer’s original goal, products discussed, questions already answered, files or voice notes they sent, previous objections and every promise the business made. Supply that context explicitly and test what survives between messages. Do not assume the system can match a WhatsApp conversation to an email conversation unless the customer’s identity has been confirmed.

Return with something new.

Useful follow-up reduces uncertainty. Depending on the conversation, that might be a direct answer to the open question, a shorter scope, a corrected price, a relevant example, a requested document or two appointment options. It should be specific to the customer rather than a generic sequence sent because a timer expired.

Keep the message easy to answer. One new piece of information and one clear question usually create a better next step than a long recap with several calls to action.

Use a structure that earns a reply.

A practical follow-up has four short parts:

1. Context: name the specific request or concern from the earlier conversation.
2. New value: provide the answer, option or resource that was missing.
3. Next step: ask one question the customer can answer quickly.
4. Exit: make it easy to say no or pause the conversation.

“Just checking in” contains none of these. A stronger version would say that the requested scope has been reduced to the two essential items, link to the approved summary and ask which start date is more practical. The details will differ, but the purpose is clear.

Set stopping and escalation rules.

Decide the maximum number of follow-ups, minimum spacing and events that end the sequence. Stop when the customer declines, asks not to be contacted, chooses another option or becomes angry. Do not manufacture urgency or keep changing the offer to force a response.

Prices and links also need controls. Configure and test the source, validation step and failure action instead of relying on an instruction to “be accurate”. Any follow-up that changes money, scope, delivery dates or a previous promise should reach a person. Kai currently lets the owner pause or take over one conversation through Telegram. The detailed framework is in the guide to customer escalation rules.

Review quality, not only reply rate.

A response is not automatically a success. Track whether the reply moved the decision forward, whether the information was correct, whether a human had to repair the message, and whether the customer asked to stop. Review failed and uncomfortable conversations as well as completed sales.

For implementation details, see Kai’s current capabilities and operating profile. The Customer Service & Sales role is one of Kai’s roles available now; the jobs page lists current plans and included message volumes. You may also want the guides to teaching brand voice and controlling prices.