How AI Is Transforming Retail, Brands & Commerce Worldwide | Sreeraman Mohan Girija, Fynd

He runs Fynd’s AI. He would still hire a human to design a logo.

Sreeraman Mohan Girija, known throughout the company as SMG, founded Fynd in 2012 and leads its AI work. His position on what AI is for is narrower and more useful than most.

The problem he keeps returning to is not a shortage of intelligence. It is that intelligence cannot move through a business whose systems cannot exchange context. A customer sees a product page, a shelf, a delivery promise, a returns button. Underneath sits product information, inventory, warehouses, transport, payments, marketplaces, stores and support, run by vendors who do not share a version of reality. That gap is what Fynd has spent more than a decade on, moving from omnichannel retail to a broader unified commerce platform.

Omnichannel stopped at the interface

The omnichannel promise was a consistent experience across web, app and store, and SMG argues that in most of the world it was only ever half delivered. The front end looks joined up while order, warehouse, transport and inventory systems carry on in separate lanes.

Unified commerce goes underneath that. It asks whether the whole stack can behave as one system with common data and coordinated decisions, so the experience is consistent because the operations are connected rather than because every channel got the same coat of paint.

You cannot rip out a hundred-year-old core

In mature markets Fynd meets retailers that are a century or two old and still running mainframe-era systems. Replacing that in one movement costs years, creates lock-in risk, and threatens exactly the continuity the technology exists to protect.

So the approach is layered. A modular, composable architecture sits above what exists, and the retailer takes the capability it needs without committing to a rebuild. Transformation becomes a sequence of controlled improvements instead of a cutover with a date on it.

Phase one is boring, and that is the point

SMG is deliberately unexcited about autonomous commerce as a near-term reality. He calls the present moment phase one: use AI to automate known workflows and take manual effort out of operations that already exist.

Cataloguing is the obvious example, because teams still lose serious time to preparing images, attributes, formats and marketplace-ready content. In Fynd AI Studio, he says work that once took a day and several people can compress into minutes once the workflow is set up. The value is not a store that runs itself. It is a team getting hours back.

The difference between a chatbot and an agent

A chatbot answers a question. An agentic system helps a customer discover a product, evaluate it, buy it and get post-order support, across a website, an app or WhatsApp. The distinction is action rather than more natural conversation.

Which is why an agent cannot be a disconnected interface. It needs permissioned access to accurate catalogue, inventory, customer, order, payment and fulfilment context before it can do anything useful reliably. Everything in the AI conversation eventually comes back to whether the data underneath is trustworthy.

Giving one person the capabilities of a team

Starting a fashion label conventionally requires designers, trend research, tech packs, manufacturers, sampling, quality checks, photography and a commerce operation. Fynd Create is built to compress that for a founder who has one designer, or none.

Trend inputs support the concept, the platform generates design variations and tech packs, and it connects the founder to a manufacturing network. Nobody is pretending the physical supply chain vanished. The claim is that a small operator can reach coordinated capabilities that used to require an organisation. Digital samples do related work: seeing how a product drapes before physical units travel between manufacturer and retailer shortens review cycles, cuts shipping, and kills weak directions before materials are committed.

A cart abandonment flow that knows who it is talking to

The standard version waits and sends everyone the same discount. A connected workflow can tell that one shopper is a loyal customer who needs a reminder while another is new and might justify a real acquisition incentive.

Boltic combines order history, customer profile, cart data and communication tools to produce different actions for different people. That protects margin as well as experience, because personalisation is only worth anything when it changes the decision, rather than inserting a first name into a template.

Turning COD risk into a workflow

Cash on delivery fraud and repeat returns are an expensive operational problem for D2C brands, and SMG’s answer is a good illustration of what practical AI actually looks like.

A new order gets checked against customer and address history, unusual return behaviour is flagged, the support team is alerted, and an automated verification call goes out. If nobody answers, the order pauses for manual review. No single signal decides anything. The workflow gathers evidence and routes the exception to a person, which is the version of automation that survives contact with a real business.

Generation is not the finish line

Early generative imagery struggled to hold the same face across a campaign, and humans spot facial inconsistency immediately. Brands spot product inconsistency faster still. Better models, dedicated avatars and more disciplined prompting have improved that control considerably.

But commerce imagery has to maintain product truth, brand consistency, marketplace specifications and customer trust across every variation. AI becomes useful at scale only once the system can reproduce quality, not merely produce novelty.

The tool is not a genie

Two designers using identical software produce completely different work, and SMG thinks AI is no different. The output depends on judgement, inputs, iteration and understanding of the problem far more than on the existence of the tool.

Which explains the apparent contradiction in someone who runs AI across a retail stack and would still hire a human to design a brand identity. Automation accelerates execution. Identity requires interpretation. The mature question is not whether AI can make something, but where human judgement is still the source of the value.

One last thing, offered as a note to his daughter: he says he would have refused this podcast two years ago, preferring to keep his head down and let the product talk. Expansion and public work forced him to learn that builders also have to explain and sell what they have made. Identity does not have to harden with age. You keep the core and change the interface.

Listen to the full conversation with Sreeraman Mohan Girija on FyndOutWithRagini.