Can voice AI agents really sell on a call

• How many times should a voice AI agent attempt to call a lead before stopping?
• Can a voice AI agent handle sales objections, and who is right, the sales manager or the CTO?
• How to identify a time-waster lead within the first 90 seconds of a call
• What happens when a lead switches from English to Hindi mid-conversation

Three attempts over seven days. Seven calls in one day is not persistence. It's a spam score waiting to happen.

What went wrong in this case is the difference between a retry loop and a retry strategy. A retry loop redials every time the CRM marks a call as "not reached", which can mean anything, including a human forgot to update the field. A retry strategy moves to the next attempt only on a verified outcome: voicemail detected, confirmed no-answer, or a specific silence duration that signals the call wasn't picked up.

The business risk here is real. TRAI's guidelines classify repeated same-day automated dials to the same number as potential harassment, which can get a DLT-registered number flagged and blacklisted. Once that happens, the entire outbound calling operation is affected, not just one lead.

The fix is a Max Attempt Cap with minimum inter-call intervals enforced at the telephony layer, not in the CRM. At Thinkly AI, retry cadence is configured as part of every outbound deployment: one call on day 0, one on day 2, one on day 5. After three verified non-responses, the lead either needs a human to call with context or is marked dead and removed from the active queue.

Both the sales manager and the CTO are right, for different parts of the problem.

Voice AI in 2026 handles scripted objections reliably. "I'll think about it," "the price is too high," "send me the brochure": these are finite, predictable objections. A well-trained agent can counter each with a prepared response and a logical next step: schedule a callback, offer a comparison, send the brochure link to WhatsApp. At Thinkly AI, these first-layer objections are mapped during the agent scripting phase so the agent has a prepared path for every common response a real estate lead gives.

What AI cannot handle is emergent objections, the ones nobody anticipated. "I heard the developer has a pending court case." "My wife just changed her mind about the location." These require contextual judgment, emotional reading, and real-time improvisation that a scripted agent isn't built for.

The correct deployment model: AI handles the first two objection layers and qualifies intent, then flags the call for a senior human closer the moment something novel surfaces. It is not AI versus humans. It is AI for volume and predictable friction, humans for complexity and relationship.

This is exactly what a qualification score is for, and it should kick in within the first 90 seconds, not after 8 minutes.

A properly configured voice AI agent runs a live intent scoring model in parallel with the conversation. It's listening for specific signals: Does the lead ask about possession dates, or just say "general inquiry"? Did they specify a budget when asked, or deflect? Is the number already in the CRM as a dead lead from six months ago? Each signal adjusts a confidence score in real time.

If the score drops below a threshold, say, below 40 out of 100 by the 90-second mark, the agent wraps the call politely: "I'll have someone from the team follow up on WhatsApp" and moves on. The lead isn't dropped; they're deprioritised and handed to a lower-cost follow-up channel.

The maths on why this matters: 8 minutes per lead across 4,000 leads a month is 533 hours of wasted call time. Thinkly AI builds qualification scoring into agent logic from the start. It's not a feature that gets added later. Most platforms don't have it configured by default, which is why teams end up running 8-minute calls on leads who were never going to buy.

With most global voice AI platforms, the agent has something close to an existential crisis.

The Speech-to-Text engine locks into the language it was initialised with. The moment the lead switches, from "Hi, I was interested in the 3BHK" to "bhai, price mein kuch ho sakta hai?", the STT either produces a transcription error or a high Word Error Rate, and the LLM responds with something that has nothing to do with what was just said.

In the Indian real estate market, this is not an edge case. This is how your buyers actually talk. Code-switching between English and Hindi, Hinglish, is the default mode of conversation for a large portion of the buyer base across Mumbai, Delhi, Pune, Bengaluru. An agent that can't handle it is not India-ready; it's only ready for the first half of the call.

The fix is Hinglish-native STT, specifically models trained on code-switched Indian speech, not US-accented English. Models like Sarvam AI are built for exactly this: they're trained on real Indian conversational data where English and Hindi switch mid-sentence, mid-phrase, sometimes mid-word. Thinkly AI deploys Hinglish-native STT as the default for Indian real estate clients, not as an add-on configuration, but as the baseline. Because in this market, a lead who only speaks English throughout a property enquiry call is the edge case, not the other way around.

The questions in this episode cover four things that break quietly in real estate voice AI deployments: retry logic that gets numbers blacklisted, objection handling that sends complex leads to an agent that can't respond, qualification scoring that's never been configured, and STT that falls apart the moment a lead says "bhai."

Thinkly AI is built for Indian real estate specifically: Hinglish-native agents, TRAI-compliant retry configuration, intent scoring built into the call flow, and a deployment process that maps objection handling before the agent goes live. Clients like Emaar, Runwal, and Sattva Group run Thinkly AI across their presales operations.

New episode every week. Subscribe on YouTube to get Agent Support in your feed.