# What broke when we let AI agents answer real customers on WhatsApp and Telegram — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/84088206f7f9882c

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-11-09 13:36 ingested 2026-11-10 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xrrtp8/what_broke_when_we_let_ai_agents_answer_real/)

## retained text



```
We've been running AI agents that answer small businesses' customers in their real chats (WhatsApp, Telegram, Instagram, email) for most of this year. Not a demo, actual inboxes with prices, bookings and annoyed people. A few things we got wrong, roughly in the order they hurt:

**1. Rules in the prompt fade, rules at the send step don't.** We had a style rule (no long dashes, a dead giveaway of AI text) in the prompt. It held in short chats and washed out in long ones. Moving it into a filter that runs on every outgoing message fixed chat, and then we found email used a different path and skipped the filter. Every channel has to go through the same last step.

**2. "Not sure if it was sent" is its own outcome.** A send that times out often went through. Retrying automatically is how a customer gets the same message twice. Now an unknown result is never retried; a person checks first.

**3. The platform's limits cost more than the model.** An agent writing first messages to new people from a personal account hit the messenger's flood limit, while every single send looked fine in the logs. Replies in chats the customer started and first messages to strangers now have separate budgets and spacing.

**4. Autonomy per type of answer, not per agent.** Answers that come straight from what the owner wrote (hours, prices, how to book) went out on their own first. Anything needing judgment (discounts, exceptions, "can you fit me in today") stayed a draft for a person much longer.

**5. The off switch has to be per conversation.** Owners don't want to turn the agent off for everyone because one chat got delicate. Pausing it in one thread, while the rest keep running, is what made them trust it at all.

**6. Let it say nothing.** An agent that can hand a message to a person instead of answering caused fewer problems than one that always produces something. Counting those handovers per topic told us exactly what was missing from the owner's material.

Curious what others running customer-facing agents found: which of your failures were model problems and which were plumbing? For us it was almost all plumbing.
```



document fetched 2026-11-10 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xrrtp8/what_broke_when_we_let_ai_agents_answer_real
