I gave this job to an AI agent. It worked. More or less.
From demo video hype to reality: when AI saves you 8 hours and when it burns 12 fixing hallucinations.
The Problem
Everyone promises 'autonomous agents working while you sleep'. In reality, unconstrained agents hallucinate, message the wrong people, or loop infinitely on API quotas.
Baseline
2 hours per day spent manually scanning Telegram channels, job boards and feeds for relevant business signals.
Hypothesis
A lean Node.js assistant with hybrid models and a strict draft-first policy can filter 90% of noise without ever sending an unapproved message.
Practical Intervention
Built igor-telegram-assistant with SQLite WAL, deterministic heuristic risk classification (high vs low risk), and manual `/send` & `/skip` controls via private chat.
Measurement & Outcome
How it was measured: Daily hours saved, zero unintended external messages sent, signal filtering precision.
Running 24/7 in production on PM2. Zero external mishaps, manual scan time cut from 120 minutes down to 15 minutes.
What Failed
Allowing it to negotiate or reply fully autonomously produced overly diplomatic, out-of-touch messages. Full automation is a fantasy; assistive drafting is powerful.
Practical Conclusion
Don't buy 'magic autonomous agents'. Build deterministic automations where AI handles specific extraction and you hold the final trigger.
Relevance to Your Business
Directly applicable to WhatsApp customer care, booking intake, quote sorting, and lead qualification for small businesses.
FROM NOTES TO ACTION
Have a similar problem in your business?
I don’t sell generic consulting. Tell me where you are stuck: I’ll reply with an honest, straightforward diagnosis.