Context
Before building anything, I ran an experiment. I called 15 real-estate companies - not to catch anyone out, but to see what actually happens when a lead who is ready to spend money picks up the phone.
The results were worse than I expected:
- 9 calls went to voicemail or rang out, during business hours
- 4 were answered by someone clearly in the middle of a job, who took my number on a piece of paper (one of them was not especially polite about it)
- 2 ended with an actual appointment in a calendar
I was not asking difficult questions. I was calling to buy.
The problem
Every one of those missed calls is a customer who simply dialled the next company on Google. The two businesses that won my appointment did not win because they were better at their job - they won because they picked up.
That is the whole gap. It is not a quality problem, it is an availability problem:
- Agents are on viewings, in cars, on other calls. The phone rings anyway.
- A number written on a piece of paper mid-job is a lead that gets lost by Friday.
- Everything outside office hours goes to voicemail, and most people do not leave one.
What I built
A voice AI agent that sits in front of the phone line and handles the part that never needed a human in the first place.
- Answers on the first ring, 24/7. No hold music, no voicemail, no “please call back during business hours.”
- Qualifies the caller. What they are looking for, budget range, timeline, which listing they are calling about.
- Books the viewing. Straight into the agent’s calendar, with the slot confirmed on the call.
- Hands over the context. The agent receives an SMS with the caller’s details and everything the conversation surfaced.
- Escalates properly. If the caller asks for a human during office hours and someone is available, the agent transfers the call. Outside hours, it takes contact details and sets a clear expectation for when someone will call back.
That last point matters more than it looks. The feature callers actually want is not a perfectly human-sounding bot - it is the ability to reach a real person when the matter feels too important to leave with a machine. Forcing someone through five conversation nodes before they can talk to a human is a great strategy if your goal is to make them stop calling.
Why voice is harder than chat
A chatbot that works well on a website usually falls apart the moment you put the same logic on a phone line. Text gives everyone a buffer - the customer types, the system thinks, the answer appears. Voice gives you nothing.
- Latency has to stay under ~500ms or the conversation feels broken
- Interruptions have to be handled naturally, not with a robotic “I’m sorry, please repeat”
- Background noise, accents and unclear speech are the normal case, not the edge case
- Tool calls - checking a calendar, looking up a listing - have to happen mid-conversation without a dead pause
This is why the build is roughly 25% building and 75% testing. Mapping the call scenarios, wiring the calendar and CRM, preparing an AI-adjusted knowledge base and picking the voice is the fast part. Running real conversations, watching where it breaks, fixing the prompt and running it again is what turns “technically working” into something a customer does not hang up on.
Timeline
- Week 1 - discovery and mapping. Every call scenario the business handles, where callers get stuck, what has to reach a human.
- Week 2 - build and connect. Telephony, calendar, CRM, knowledge base. First real test calls.
- Weeks 3-4 - testing and launch. Edge cases, awkward pauses, the strange things people say on real calls.
The agent does not have to be perfect to go live. It can start handling real calls while the refinement continues in the background.
The result
I listened to the agent handle a real call at a real estate agency.
60 seconds. Appointment booked. The agent got every detail by SMS. No missed call, no voicemail, no “please call back during business hours.”
On disclosure
Since 2 August 2026 the EU AI Act requires companies to tell callers when they are talking to AI, so hiding it is not an option - and it turns out not to matter much. Once people know it is AI, they stop expecting a human and start expecting speed. That changes what you optimise for: no fake coughing, no acting, just an agent that understands intent, never loops the same unhelpful line, and always captures the request even when it cannot handle it.
What it proves
The competitor beating you is rarely beating you on quality. He is beating you on availability.
A narrow agent that does one thing reliably - answer, qualify, book, hand off - delivers more than a general-purpose assistant that impresses in a demo and disappoints on the 50th call. And it ships in 3-4 weeks, because most of the six-month projects spend five of those months deciding what to build.