What they are
Contacts tripled in a few weeks, and every one of them deserved an answer while they still
cared — on the phone, on WhatsApp, by email, on the website. Doing that with people alone
would have meant hiring, training, briefing and monitoring them for work that repeats
itself all day.
So I built two agents as real operators, not demos. The prompt and the configuration live
in git and are compiled into the agents; nobody edits them by hand in a dashboard.
Carla is the secretary, fast. Andrea is the assistant, he goes deeper.
The two of them
Carla
Voice · phone line
Carla is the voice agent on the phone line, built on ElevenLabs and Twilio. She opens every call by saying she is an AI assistant, gives the facts about the property, answers objections from the knowledge base, calls back the people she promised to call back, and books a phone consultation with a human consultant. About 29 tools mid-call: property details, free slots, book, move or cancel, visit requests, callbacks, urgency flags, WhatsApp and email, CRM updates.
She is measured on one thing: how many calls end with a booking.
Andrea
Text · WhatsApp, email, chat
Andrea is the text agent on WhatsApp, email and the website chat, with 22 tools. He builds a profile of what the person wants and proposes properties one by one, instead of throwing a catalogue at them. Since October 2026 he is the first contact for every new lead coming from the property portals and the website, with two WhatsApp reminders.
He goes deeper, because text gives him the time to.
The numbers
- ~3x contacts a week From about 1,250 in August to over 3,600 in September
- 79,857 past calls analysed 1.1 million turns, before Carla was designed
- 50.5% calls ending in a booking With the full property sheet. 18.2% with a saved search only
- ~14¢ per minute of voice The language model is most of it, not the voice
Measured September–October 2026. Carla took over the phone line around 15 September, Andrea went live between 21 and 24 September.
How they work
They are wired into the business the way a team member would be: the channels, the
property data, the consultants’ calendar, the messages and the CRM are all part of the same
flow. Under the hood: ElevenLabs for voice and text, Twilio for the phone line, Supabase for
the data and the tools, Gemini models doing the thinking.
-
Listen on the customer’s channel
Carla answers the calls. Andrea replies on WhatsApp, email and the website chat. Nobody is asked to move somewhere else.
-
Know the property
They read the property data and use tools when the conversation needs facts, not guesses. What the agent has in hand decides the result more than anything in the prompt.
-
Act
They book consultations, schedule callbacks, send a WhatsApp or an email, move the lead forward in the CRM, and set up saved-search alerts.
-
Hand over
When the case needs a person, they stop trying to be clever and raise the right flag, to the right human.
Handing over to humans, both ways
The handover is not an afterthought. When an agent cannot or should not go on, it opens one
of four kinds of S.O.S.:
- Urgent The auction is close and there is no free slot before it.
- We want to bid Three questions first — which house, what budget, mortgage or own funds — then straight to the head of the consultancy.
- Call me back Someone just wants to talk to a person.
- Staff task Admin follow-up for the office.
It works the other way too. If a consultant writes by hand in a thread, Andrea goes quiet
for 24 hours, 48 hours or forever — the consultant decides — and one button gives the
thread back to him. Every booking lands on a human consultant’s calendar.
The point is not replacing the team. It is making the boring part disappear and the handover
clean.
Want agents like these?
If your business has calls, messages, bookings and too many manual follow-ups, we can
probably build something useful. Tell me where the work is getting stuck.
Get in touch The repositories are private. Happy to walk through any of it.