On messaging apps, people write in bursts. “Good afternoon.” “How are you?” “Do you deliver to Austin?” Three messages in eight seconds that, for the person writing, are a single sentence. An agent that answers each one as it arrives sends “Good afternoon! How can I help?” while the real question is still being typed, and then a second reply that ignores the first.
Answering in batches means waiting a short window after each incoming message and, if another arrives within it, merging everything into a single turn. The agent replies once, to the whole, rather than once per line. A three-second window catches most bursts without the person writing noticing any delay.
What goes wrong without it
- Crossed replies. The agent answers the “hi” while the question is arriving, and the two replies land out of step with what the person wanted.
- Multiplied cost. Every message becomes a language-model turn, with the whole history processed again.
- Half the context. The reply to the first message is generated without the information in the second, and often asks exactly what the person was about to say.
How long the window should be
Too short, and bursts slip through. Too long, and someone who sent a complete question waits for no reason. Three seconds is a good starting point for messaging; audiences who type more slowly need a little more, and a ceiling keeps the wait from growing without bound. In voice the idea does not apply: there, speech detection decides when a turn ends.
Merging messages is not making the customer wait. It is waiting for them to finish talking, which any human agent does without thinking.
Nothing gets lost while waiting
Waiting to reply cannot mean waiting to record. Every message has to be stored the moment it arrives, so the inbox shows the real conversation and a human agent who opens it at that instant sees everything. Only generating the reply is deferred. And when several messages arrive at once, storage has to be atomic: under load, one concurrent write must not wipe out another.
How it works at Take Sales
On Take Sales messaging channels, such as WhatsApp, Instagram and Telegram, the agent has answered in batches since September 2026, with a default window of three seconds and a ceiling of fifteen. The window follows whichever agent actually owns the conversation, including after a transfer between agents. For what happens when the conversation moves to a person, read about handoff.
