There’s a lot of noise right now about AI agents going off-script, hiding their reasoning, taking actions nobody approved. Worth paying attention to if you’re building frontier models. Less relevant if you’re running customer support for a logistics company in Indore, where the actual problem is simpler: can the agent understand what your customer just said in Hindi, mid-call, with a Tamil accent layered on top, while a kid is screaming in the background?
That’s the problem we built ConvoZen around. Not lab-condition AI safety scenarios, the un-glamorous reality of how Indians actually talk to customer support.
Our agents run across voice, WhatsApp, email, and chat, and they don’t pause to figure out which Indian language they’re hearing before responding, they’re built for it natively. Hindi, Tamil, Bengali, and more, handled the same way an English-only agent handles English. They update CRMs mid-conversation, carry context across channels so customers don’t repeat themselves, and close out workflows without waiting on a human to confirm every step.
The results aren’t theoretical. Akshara and Ragini, our in-house speech models, now power over 90% of quality audits automatically and have pushed efficiency up by 30% across the deployments running on them. NoBroker, the company that built us, runs ConvoZen on its own support operations across thousands of conversations a day. If it didn’t actually work at that volume, it wouldn’t still be running internally.
Akhil Gupta, our founder, has a useful way of framing the human side of this: the agents cut dependency on people for routine queries by up to 40%, but the moment something gets complicated, a human takes over, backed by ConvoZen Copilot pulling context and suggesting next steps in real time. Meanwhile our Supervisor agents quietly audit every single call, which is part of why manual QA hours have dropped 90% for clients without anyone losing visibility into what the agents are doing.
So no, the agents aren’t lying or going rogue. They’re doing something less headline-worthy: handling Bharat’s customer conversations in Bharat’s languages, freeing up people to handle the parts that actually need a human.
Read more here.