What separates a B2C presales workflow where voice AI actually produces site visits from one that gets quietly switched off after a month is almost never the technology. It is the implementation. The model that sounded impressive in the demo and the agent that fumbles on live leads are often the same product. The difference is what happened in the two weeks between signing and going live: what the agent was taught, how the script was built, how voice and WhatsApp were connected, and how the pilot was measured.
The failed implementations share a pattern. The agent goes live knowing the brochure but not the answers to the thirty questions buyers actually ask, so conversations stall on carpet area and payment plans. Voice and WhatsApp run as two disconnected systems, so the buyer receives a brochure on chat and then a call from an agent that does not know it was sent, and every touchpoint restarts the relationship instead of continuing it. Retries fire at the same hour that already failed. And because nobody defined upfront what working means, the pilot ends in a debate about vibes instead of a decision about numbers.
The successful implementations do the same five things in the same order, every time: load the project knowledge properly, build the script with the sales head in the room, connect the CRM and the WhatsApp layer so context flows both ways, pilot on a defined lead slice with rep-style metrics, and only then scale. That process is knowable, repeatable, and short. It fits on one page, and this post lays it out stage by stage. Our note on the 6 things that go wrong with voice AI in real estate covers the failure modes in more detail.
Companies like Thinkly AI run this exact implementation for high-ticket B2C presales teams, real estate and EdTech foremost, deploying multilingual, human-like voice AI agents with connected WhatsApp journeys that qualify buyers and book site visits and consultations, live on real leads within one week.
What does setting up an AI voice agent involve?
Setting up an AI voice agent for a real estate project involves five stages: loading project and inventory data, configuring the conversation script and languages, connecting the CRM and lead sources, running a supervised pilot on live leads, and scaling to full volume. On a platform built for real estate, the full process takes roughly one to two weeks, and the pilot stage is where most of the tuning happens.
Stage 1: Project and inventory data (days 1 to 2)
The agent is only as good as what it knows. The setup starts with the material a new presales rep would be trained on: configurations and carpet areas, price bands and payment plans, possession timeline, amenities, micro-market context, RERA number, and the answers to the twenty questions buyers actually ask. Most of this already exists in the project's sales kit, so the work is assembling it, not creating it.
Two inputs teams forget and regret: the objection list, meaning what buyers push back on and the approved responses, and the disqualification rules, meaning budgets or timelines the project should politely release rather than book.
Stage 2: Script, persona, and language configuration (days 2 to 4)
This is where the agent stops being generic. The qualification flow gets defined: greeting, project introduction, the qualification questions in order, the booking offer, the close. So does the agent's persona: name, gender, tone, and crucially, language behaviour. For most Indian projects that means Hinglish as the spine with clean switches into Hindi, Marathi, or the buyer's language mid-conversation, the pattern we describe in Hinglish AI calling for real estate sales. Thinkly AI configures this per project, because a Virar project and a Bandra project should not sound like the same caller.
The script review with the sales head is the single highest-value meeting in the whole process. Thirty minutes of a buyer would never say it like that saves weeks of pilot iteration.
Stage 3: CRM and lead source connections (days 3 to 5)
The agent needs leads flowing in and outcomes flowing out. Inbound: portal integrations or webhook feeds from 99acres, MagicBricks, Housing.com, and Meta or Google lead forms. Outbound: structured writeback to the CRM, whether LeadSquared, Salesforce, or Sell.do, with the transcript, qualification answers, intent score, and next action on every record. Calendar sync for site visit slots completes the loop.
This stage runs in parallel with scripting and is mostly an integration checklist rather than an engineering project when the platform has the connectors already built.
See the setup process on your own project
Bring one project's sales kit and we will show you what the configured agent sounds like.
Book a demoStage 4: Supervised pilot on live leads (week 2)
The pilot is where the deployment earns trust. A defined slice of live leads, often after-hours enquiries where the comparison baseline is nobody called at all, goes to the agent while the team listens to recordings daily. The first days always surface fixes: a mispronounced locality name, a question buyers ask that the script did not cover, a retry timing that lands during office hours when buyers cannot talk.
The discipline that makes pilots work is measuring the agent like a rep: connect rate, qualification rate, visits booked, and show-up rate, reviewed against the human team's numbers on the same lead types. Sales call analytics across every pilot call makes this review fast, because every conversation is scored, not sampled. Our guide on how AI call scoring works explains the rubric.
Stage 5: Scale and steady state (week 3 onward)
Once pilot numbers hold, volume ramps: more lead sources, retry cadences switched on for the backlog, reminder and feedback calls added to the flow. Steady state is not set and forget. It is a monthly rhythm of reviewing flagged calls, updating the script when the project's pricing or inventory changes, and adding campaigns such as a new tower launch or a dormant database reactivation as sales priorities move.
The one-page version
| Stage | Duration | Owner | Output |
|---|---|---|---|
| Project data | Days 1 to 2 | Sales head plus platform | Agent knowledge base |
| Script plus languages | Days 2 to 4 | Platform, sales head review | Approved conversation flow |
| Integrations | Days 3 to 5 | Platform plus CRM admin | Leads in, outcomes out |
| Pilot | Week 2 | Presales lead | Validated numbers |
| Scale | Week 3 plus | Presales lead | Full-volume operation |
The pattern to avoid is the reverse order, months of integration planning before anyone has heard the agent speak. Hearing configured calls on your own project in week one is what turns internal sceptics into sponsors, which is why AI agents for real estate deployments that start with the script ship, and ones that start with an architecture document do not. Our full guide to AI voice agents for real estate in India covers what to expect after go-live.
Ready to run the process on live leads?
Thinkly AI takes a project from sales kit to supervised pilot in about a week.
Book a demo