Kiko AI Order Taker

Small merchants on ONDC were fielding orders by phone, exactly as they always had. Every order lived only in a phone call. There was no structured record, no price validation, no way to feed it into inventory or fulfillment.

The gap wasn't a lack of willingness to digitize. It was that any structured order-entry tool would be slower than just answering the phone. The product had to be at least as fast as talking, or merchants would ignore it.

Early transcription tests made it obvious that raw speech-to-text alone wasn't trustworthy enough to post directly as an order. Accents, background noise, and corrected orders ('actually, make that two') all broke naive transcription.

The real design question wasn't how accurate transcription could get. It was what happens when it's wrong, and how the merchant recovers in under two seconds without breaking their flow.

Structured the workflow as transcription, then structured extraction, then validation. Validation was designed as a lightweight, glanceable confirmation step rather than a full review screen.

Defined the ML product requirements for transcription and extraction, and partnered directly with data science on evaluation. Iterated on what 'good enough to trust' meant stage by stage, rather than chasing raw model accuracy in isolation.

Rolled out in stages, starting with a small merchant cohort to pressure-test the validation UX, then scaling as verification rates held up under real-world noise and accents.

Instrumented the pipeline so every low-confidence transcription routed to the validation step automatically, keeping bad orders from ever reaching fulfillment.

Scaled to 500+ active merchants generating ₹3.03 Cr+ in GMV, with a 92.6% verification rate. The vast majority of AI-transcribed orders needed zero merchant correction.

23.6% of customers became repeat buyers through the platform, a signal that the order experience held up well enough to bring people back.

Trust is built at the validation step, not the model card. Merchants forgave imperfect transcription as long as catching and fixing an error took less time than saying it out loud again.

Ship the safety net before the automation. The human-in-the-loop step should exist from day one, and get quieter over time as confidence earns its way down, not the other way around.