Building the AI workforcefor Bangladeshi commerceBuilding the AI workforce for Bangladeshi commerce

Everyone is suddenly talking about conversational commerce. Bangladesh has been doing it for ten years.

Almost nothing here goes through a checkout. Someone sees a product on Instagram, asks the price in a DM, moves to WhatsApp, and pays cash when the rider turns up. Nine in ten online buyers here prefer paying that way. There's no cart to abandon and no card to decline. The conversation is the whole business.

It works, and it has a ceiling. A brand grows until it hits the number of messages one founder can read in a day, and then it just stops. Meanwhile every parcel ships before anyone knows whether the buyer will take it, and the ones that come back cost the delivery twice.

What's changed is that a machine can now hold the conversation. Two years ago there was no voice model speaking Bangla well enough that someone in Sylhet wouldn't hang up. Now there is, and a conversation costs almost nothing. That's the opening, and I don't think it stays open for long.

I spent a while wondering why the big platforms wouldn't just take it. They score fraud on card signals, which a cash order never produces. The couriers will tell you how many parcels a buyer accepted from them, but only from them, so someone who refused delivery on one network looks clean on the next. The chat tools answer the message and leave the order for a person to type in later. Everybody has a piece of it. Nobody has the order.

So we built that. Omnistra is an AI workforce sitting on one record of every order, buyer, product and parcel. Tania takes the call in Bangla, Riya Text works the DMs, and all of them answer from what is actually true rather than what someone typed into a prompt last month, including whether Pathao or Steadfast is the one that delivers in that area.

The part I keep coming back to is what happens as it runs. Every delivery teaches the system something about a buyer, an area, a courier, and a merchant joining next year inherits all of it. The score that tells a seller whether to trust a cash order gets better with every merchant on the platform, so it is worth more to each of them the more of them there are. A single-tenant tool cannot copy that, and it is why this compounds instead of just growing.

I built voice AI for years before this, and I have sat with enough people running these businesses to know what holds them back is not ambition. They are doing by hand what a machine has quietly got good enough to do.

We're early. Some of what is on this site we built last month, and we are taking our first merchants slowly. But I have never been more sure of a direction. Bangladesh worked out how this kind of commerce runs. We want to build the thing it runs on, here first, and then in the markets that work the same way.