An AI Agent That Answers DMs 24/7

By Alvin Ramdin · June 3, 2026 · ~6 min read

Most small operators lose the same way: a message lands in the WhatsApp or Telegram inbox at 9pm, nobody sees it until morning, and by then the person has messaged two competitors. You don't lose those leads to a better pitch. You lose them to a slower reply.

So I built a small agent to close that gap: an OpenClaw agent that watches the DMs, answers the routine stuff in your voice, and pulls you in only when it matters. This is the real build: what it does, the three mistakes that nearly sank it, and the fix for each. No invented numbers, just what actually happened.

"Multiply your output, not your headcount" isn't a slogan here. It's one person now covering a channel that used to need a second one.

What the agent actually does

The job is narrow on purpose. The agent connects to WhatsApp and Telegram, reads incoming messages, and handles three jobs: answer common questions (what you do, hours, rough pricing), drop the booking link when someone shows intent, and quietly tag anything it can't handle so a human picks it up. That's it. It is not trying to be a chatbot personality. It's trying to make sure no message sits unanswered overnight.

Here's the shape of a real exchange it handles without me touching anything:

Setup took an afternoon, not a sprint. The agent is given a handful of facts about the business and a tone to match, and because it runs on files you own, there's no SaaS subscription holding the logic hostage. If you want the full primer on what OpenClaw is and where it fits, I wrote that up in What is OpenClaw.

Mistake #1: I gave it too much rope

The first version answered everything, confidently. Someone asked a detailed question about a service edge case and the agent improvised an answer that wasn't quite right. That's the classic failure: an eager model filling silence with a guess.

The fix: a tight scope and one hard rule: when you're not sure, don't guess; hand it to a human. I gave it a short list of what it's allowed to speak on (services, hours, booking) and told it that anything outside that gets a friendly "let me get Alvin to confirm that for you" plus a flag to me. Narrower scope made it more useful, not less.

Mistake #2: instant replies that sounded like a robot

Replying in under a second, 24 hours a day, reads as automated, and people clock it immediately. A 3am machine-gun response actually lowered trust.

The fix: two small changes. First, a short, natural reply delay so it doesn't feel like a vending machine. Second, the tone got rewritten to sound like a real person who's just quick to respond, not a script. The goal was never to hide that there's automation. It's to make sure the automation doesn't make you look worse than a slow human would.

Mistake #3: hot leads got stuck in autopilot

The worst near-miss: the agent was so good at being helpful that a clearly high-intent lead got a tidy answer and then… nothing escalated. The system was content. I was not.

The fix: intent triggers. Certain signals (words like "price," "book," "today," "urgent") break the agent out of FAQ mode. It drops the booking link immediately and pings me so I can jump in personally if it's worth it. The agent handles volume; you handle the moments that close.

A good DM agent isn't measured by how many messages it answers. It's measured by how few good ones it lets slip.

The part nobody mentions: you can see everything

Because OpenClaw syncs the conversations, you're not flying blind. Every auto-reply is logged, so you can read what it said, catch a wrong answer early, and tighten the scope over the first week. Treat the first few days as a shakedown: read the transcripts, fix the one or two things it fumbles, and then let it run.

Build it, buy it, or wire it up another way?

An agent like this is the right tool when your bottleneck is first response on a messaging channel. If your problem is moving data between apps or triggering workflows on a schedule, a different tool fits better. I broke down where agents beat automation platforms (and where they don't) in OpenClaw vs n8n vs Zapier. And if you'd rather start with something even smaller, 5 Claude skills for solopreneurs has a few one-afternoon wins.

What it costs to run (the proof, not the pitch)

Here's the honest math, since it always comes up. An OpenClaw agent like this starts around a $497 setup with roughly $147/mo to run, or a small API bill if you self-host the logic. A part-time VA covering the same hours runs $12,500 to $52,000 a year. So yes, it's cheaper than a person.

But that's the closer, not the headline. The reason to do it is simpler: the same one operator now covers a channel that used to go dark. That's the multiplier. The cost just makes it an easy yes.