Inside Product Agents: AI in ecommerce

At Hello Retail, we’ve spent years building what we call Product Intelligence, developed quietly and iteratively across thousands of stores and millions of interactions.

When we launched Product Agents, we wanted to explain the mechanics rather than just announce it. So we sat down with Sarah Miguel Cournane, our data scientist, in February, as part of our Hello Retail Conversations series. This is what she walked us through.

Watch the Hello Retail Conversation with Sarah here →

The shift: From predicting behavior to acting on it

Ecommerce has used AI for years, just not always under that name. Recommendation engines, search algorithms, collaborative filtering: these are all forms of machine learning that have shaped how shops work for a decade.

What’s changed is speed.

“We’ve gone from using AI to support decisions to actually verifying what AI produces for us.” (Sarah Miguel Cournane)

That shift matters. Where AI once supported decisions, it now acts on them in real time, at scale, and in formats customers actually see.

That’s where Product Agents come in.

What are Product Agents?

A Product Agent is an autonomous AI system that monitors your product catalog and customer behavior, then independently decides who to contact, when, with which product, and what message to send. Unlike manual email flows built around fixed triggers, each agent acts on live signals: a viewed product drops in price, a consumable is due for replenishment, a preferred item goes out of stock. The decision layer is predictive AI. The writing layer is generative AI.

Product Intelligence: The foundation most people overlook

Before you can generate anything meaningful, you need to understand what drives behavior. That’s what Product Intelligence does, and it starts from a different place than most personalization systems.

Rather than relying on who a user is, it focuses on what they interact with: which products get viewed together, bought together, or never overlap, and which get replaced, replenished, or abandoned.

This creates a different foundation for personalization.

“If we don’t know who you are, we focus on what you engage with. That’s enough to understand intent.” (Sarah Miguel Cournane)

Behind the scenes, every product is mapped into a high-dimensional space, with each item carrying a set of characteristics and relationships rather than only an ID.

The diagram above shows the transformation: a product (here, a bag) contributes its image, title, and description into a multi-dimensional vector representation. That vector is how Product Intelligence tracks each item’s relationships across the catalog, even before any purchase data exists.

So when a new product enters a catalog, it doesn’t start from zero. It inherits knowledge.

That solves one of ecommerce’s biggest problems: the cold start.

The reality most stores don’t see

Two numbers came up in the conversation that frame the challenge: roughly 30% of products drive 70–80% of revenue, and roughly 50% disappear within six months.

Half your product knowledge resets constantly. Product Intelligence addresses this by linking new inventory to historical behavior, so the system’s understanding carries forward even as the catalog changes.

Where Product Agents change the game

Most email setups depend on flows you build manually: setting triggers, guessing timing, choosing content. The maintenance never stops.

Product Agents remove that layer. Instead of building flows, you define intent. The system handles the rest: who to contact, when, which product to show, and what message to send. It runs on predictive and generative AI.

The real innovation: Combining two types of AI

Generative AI gets a lot of attention right now, but in isolation it isn’t enough for what Product Agents do. The value comes from combining both:

• Predictive AI: what people are likely to want and when
• Generative AI: how to communicate it effectively

“It’s not a competition between the two. The combination is where things get interesting.” (Sarah Miguel Cournane)

That’s what Product Agents are built on.

Timing matters, and now you get both

Replenishment is a clear example. Buying a sofa is a one-time event; buying dog food is not. That difference shapes everything: frequency of communication, message type, timing.

Product Agents understand that distinction automatically, the same way they handle price drops, product alternatives, and inventory changes. The result is messages tied to actual intent rather than a campaign calendar.

The hidden complexity: Tone of voice across languages

One unexpected challenge turned out to be linguistic rather than technical.

Tone of voice is not universal. Friendly in Spanish reads differently than friendly in German; formal in Danish is a different register than formal in English. What sounds natural in one market can feel off in another.

That makes tone of voice a matrix of language, brand identity, context, and audience expectation, all working together at scale. The system has to find the right register across thousands of emails it never manually reviews.

Merchant-side agents

Most of the current conversation around AI in ecommerce focuses on shoppers: shopping assistants, AI copilots, chat interfaces. Product Agents take a different angle. They are built for merchants, but they serve the shopper.

“They help the shop understand what the customer wants, and help the customer find it faster.” (Sarah Miguel Cournane)

Instead of waiting for the customer to act, the store becomes proactive.

What this actually means for ecommerce

Ecommerce is shifting from static flows to dynamic decisions, from targeting segments to addressing individuals, from campaigns to continuous optimization.

And from shops you visit, to shops that come to you, in a way that feels relevant rather than intrusive.

Final thought

The easy framing is AI as a layer you add on top of existing systems. What’s happening is more structural: a different way of organizing how decisions get made in ecommerce.

Product Agents are one expression of that shift, and the infrastructure underneath is still early.

Full details and a breakdown of each agent are available on the Product Agents page.