From Code to Sales: My AI Got a Meeting with Barcelona's Largest Gym Chain in 3 Hours
The pivot
Yesterday I built Casandra Lite — a multi-tenant Telegram bot with MiniMax M2.7, escalation system, reminders, supervision. A complete MVP in one day.
Today I froze it.
Not because it doesn't work. It works. But because I have too many fronts open and there's one that needs real traction: OhanaSmart — robotic açaí and natural yogurt stations for gyms, hotels and coworkings.
Code can wait. Clients can't.
The prospecting machine
Instead of writing code, today I built something more valuable: an automated sales machine.
My agent (Harvie, running on OpenClaw with Claude Opus) became a marketing specialist: with its own corporate email, signature with logo and defined personality. In under 2 hours:
- Researched the Barcelona market: premium gyms, 4★ hotels, coworkings
- Compiled 7 leads with real contact emails
- Drafted personalized emails — in Catalan for local gyms, Spanish for hotels, English for international coworkings
- Sent all 7 emails with PDF dossier attached and corporate signature with logo
- Set up automatic follow-up: check every 2h, follow-ups at 3 days, weekly report
All from a €4/month VPS.
The result
3 hours after the first send, Barcelona's largest gym chain replied. 20 clubs. +75,000 m² of facilities.
"Gràcies per posar-te en contacte. Et sembla que organitzem trucada per veure opcions de col·laboració?"
Call scheduled for the following week.
What the AI did (and what it didn't)
What Harvie did:
- Searched for leads on the web
- Found contact emails
- Drafted emails personalized by language and sector
- Sent them with signature, logo and dossier attached
- Configured DNS (SPF, DKIM, DMARC) so they don't land in spam
- Set up automatic follow-up crons
- Detected 2 bounces and resent to correct emails
What it did NOT do:
- Confirm the meeting — I do that
- Decide pricing — I decide that
- Close anything — AI opens doors, I walk through them
The lesson
I've spent weeks building tools. Bots, benchmarks, integrations. All useful, all interesting, all… with no clients.
Today, in 2 hours, my AI did what I'd been putting off for weeks: knocking on doors.
You don't need a perfect product. You need a client who says "yes, I want to try it". The code gets written after.
Prospecting stack
For those who want to replicate it:
- OpenClaw + Claude Opus as orchestrator
- Custom email (domain + SMTP) — not Gmail, not Mailchimp
- SPF + DKIM + DMARC configured (essential to avoid spam)
- Python for SMTP sending with HTML signature + attachments
- Crons for automatic follow-ups
- Brave Search for researching leads
- Total cost: ~€4/month (VPS) + Claude tokens
Next steps
- Prepare the call (script, economic proposal, pilot)
- Keep prospecting: 5-8 new leads per day
- First goal: LOI (Letter of Intent) — a paper that says "we want a machine"
Yesterday's bot can wait. Next week's meeting can't.
— I, Johnny — configured agent: Harvie. AI doesn't take your job: it takes your excuses.