Lurked here forever, finally have something to share.

About a week ago I build an AI sales agent for Telegram. Got tired of watching warm leads rot in shared inboxes while support copy-pasted the same 5 replies. Sharing what I learned because most content about AI chatbots is written by people who've never run a campaign that actually had to convert.

Background: 8 years in iGaming.

Why Telegram first, not WhatsApp

Everyone asks this. WhatsApp's Business API is a walled garden – template approvals, 24-hour windows, BSP middlemen. Building AI for WhatsApp at scale means dancing with Meta daily. Most WhatsApp CRM tools look identical because they're stuck inside the same cage.

Telegram is the wild west in a good way. Generous Bot API, rich interactions, and the user base in crypto/iGaming/fintech is huge. AI for Telegram lets you experiment without compliance breathing down your neck every time you test a flow.

The casino background matters more than it sounds

Casino marketing teams have been doing "AI-driven personalization" for 15 years. Just without the AI. Behavior triggers, heavy A/B testing on literally everything, cohort analysis by LTV, brutal honesty about what converts (almost never what marketers think it will).

So when I built the AI auto reply layer my first thought wasn't "make it sound human." It was "make it convert." That's where most chatbot projects die.

Stuff I learned

The model doesn't matter that much. Everyone obsesses over which LLM. GPT, Claude, open-source – they all work for sales convos once the prompt and context are right. The hard part is everything around it: CRM integrations, handoff logic, state, analytics.

"Sound human" is the wrong goal. Users on Telegram half-know they're talking to a bot. They don't care if it's useful and fast. What actually moves the needle:

• Reply under 3 seconds (this alone beats better phrasing)
• Remembers what they said two messages ago
• Doesn't make them repeat info
• Knows when to shut up and pass to a human

Hard handoff rules save your numbers. Casino taught me this. Payment issues, complaints, anyone naming a competitor, anyone spending over X – instant escalation, no exceptions. Letting AI try to handle these costs money.

Agent > chatbot. A bot answers questions. An agent does things. Mine pulls order history, checks stock, generates discount codes, books demos, pushes everything into the CRM with the right tags. Most products marketed as AI sales agent are honestly just FAQ bots with better branding.

Read your logs. Every convo logged, every handoff logged. I still go through maybe 50 a week manually. Found 12% of handoffs were unnecessary – AI could've closed but bailed early. Fixed the prompt, conversion up 18% in two weeks.

What I'd do differently

Spent way too much time on conversational quality at first. First version sounded great and converted badly because it couldn't do anything. Rebuilt it backwards: integrations first, conversation on top.

Also underestimated how much is non-AI work. Probably 70% of engineering is queue management, dedup, opt-outs, rate limits, webhook reliability. The boring stuff that keeps high-volume messaging alive. Coming from casino I should've known.

Where this is going

Real frontier isn't smarter AI. It's the plumbing. Your AI chatbot is only as good as the data it can pull and the actions it can take. The winners over the next two years will treat AI as the interface and invest in the orchestration underneath.

WhatsApp will eventually loosen up, or someone builds a WhatsApp CRM clean enough to make the constraints workable. When that happens, whoever has flows already proven on Telegram ports over with a head start. That's the bet I'm making.

Happy to go deeper in comments on any of it – prompt structure, handoff logic, the iGaming-SaaS shift, whatever. Also curious what others are seeing if you're building in this space. Good content on AI for Telegram is basically nonexistent and I'd love to compare notes.