Why I Gave My AI a Telegram Leash
We built an AI investor concierge that runs follow-ups automatically — but every message goes through a Telegram approval first. Here's the design philosophy, and how trust levels let us hand the wheel over gradually without losing control.
Most people building AI automation fall into one of two traps.
Trap 1: Too much control. Every message requires manual approval. The AI becomes a glorified draft generator. You're still the bottleneck.
Trap 2: Too little control. Full auto-send. Works great until the AI says something wrong to a $500K investor and you find out about it three days later.
We built an investor follow-up system for a capital raise client and had to solve this exact tension. Here's what we landed on.
The setup
The client is running a real estate syndication — a $3M raise. Over 150 investors in a pipeline: some attended the webinar, some didn't, some asked questions, some ghosted after the first email.
They needed consistent follow-up that was personal and timely. That's what AI is good at.
But they also needed to trust what was going out under their name. You can't have Claude hallucinating financial projections to accredited investors. That's not just embarrassing — it's potentially a securities issue.
The Telegram leash
Every draft the AI generates goes to a Telegram bot first.
The sponsor gets a message like this:
New draft ready — James Whitfield (attended, 2 Q&A replies)
Subject: Quick follow-up after Tuesday's webinar
"Hey James, really appreciated your questions on Tuesday. You asked about the preferred return structure — wanted to follow up with the full breakdown..."
[Send ✅] [Edit ✏️] [Skip ❌]
One tap to send. One tap to skip. Or they hit Edit, type a quick note, and the AI rewrites it.
No message goes out without the sponsor seeing it first.
Why this doesn't become a bottleneck
The obvious objection: "That's just moving work from email to Telegram."
Fair. If every message needed careful review, it would be. But here's how we solve that.
The system tracks which action types get approved without edits. After enough consistent approvals of the same type, that type auto-promotes:
| Level | What happens |
|---|---|
| Manual | Every message → Telegram approval |
| Semi-auto | Draft shown, auto-sends after 2 hours if no action |
| Auto | No notification, sends immediately |
Routine nudges — "Hey, just following up from Tuesday" — hit semi-auto fast. Approve them five times without changes and they graduate.
High-stakes messages — Q&A replies with specific deal numbers, booking proposals, anything touching financials — stay manual forever. We hard-coded that. It's not a setting you can accidentally change.
Within a week, the sponsor is only reviewing the messages that actually need a human eye. The routine stuff runs itself.
The emergency stop
We also built a /stop command.
Type it in the bot and everything freezes. No outbound messages, no scheduling, no AI decisions.
They can also kill specific threads — /pause james stops all activity for that one contact without touching anyone else.
This matters more psychologically than technically. Knowing you can stop everything at any time makes it easier to trust the system in the first place. I've noticed clients visibly relax once they see that button exists.
What we learned
The approval UX matters more than I expected. We originally showed the full email body in the Telegram notification. People weren't reading it — they just tapped Send. Switched to a 3-line preview with a "Show full draft" link, and approval quality went up immediately. The sponsor actually caught a mistake on the third draft.
Trust builds fast for boring messages and slow for important ones. That's exactly right. The system behaves like a good employee: handles routine stuff automatically, escalates the judgment calls.
One thing I didn't anticipate: the approval loop makes the AI better over time. Every time the sponsor edits a draft, we can fold that feedback in. The loop goes both ways.
The pattern, not the product
We built this for investor follow-ups, but it works anywhere you're running AI communication at volume: client onboarding, lead follow-up sequences, support ticket drafts, internal updates.
The core idea is simple. AI drafts, human approves, trust builds, routine messages auto-promote, humans stay on the hard stuff.
Full automation is fragile. Full manual is a bottleneck. The leash is the setup people actually stick with.
We're working with a handful of capital raise clients right now. If you want to see the full system — including stage-aware information gating and deal Q&A — reach out.