## How to Structure AI for a Marketplace: Turning Customer Data into Timely Decisions

technical-article · 2026-10-06 · Victoria · 7 min · en

> How marketplaces can connect customer data, specialized agents, and timely decisions.

Marketplaces collect years of data about what people search for, view, and buy. But having that data doesn’t guarantee a useful customer experience.

It’s standard: we buy a sweater and keep seeing ads for it. Someone even commented recently that she had bought the same shirt three times because she forgot she already owned it. Each time she saw it, she thought, “Aw, this seems like me.”

It’s kind of a funny story, but it points to a real challenge for the teams working on those companies: customer behavior, logistic efficiency, and the data a marketplace collects don’t always line up.

Of course, we all have an idea now that AI can help connect those signals, but the solution needs to focus on the best usage of business knowledge, how to treat new inputs, how quickly it needs to respond, and how to retrain the solution fast enough.

Structuring an AI Solution for a Marketplace

In a talk by Nancy, a solution architect at AWS, the problem had three parts: fragmented customer data, a high volume of new interactions, and everyday situations that change what a customer needs.

A model trained on past behavior may not account for new interactions. A customer might usually shop for herself, then look for a Mother’s Day gift. If the marketplace keeps treating that purchase as a signal about her personal style, she may see recommendations for the gift item long after the occasion has passed.

One way to approach this is with a system of specialized AI agents: an orchestrator receives customer signals and calls agents responsible for specific tasks, such as recommendations, ad copy, or loyalty offers.

The orchestrator also needs to keep the company’s broader goal in view. While a specialized agent might have as a main metric improving loyalty-program activation, the bigger goal is to build a lasting customer relationship (that's what the orchestrator is responsible for).

<figure class="ma-diagram" role="img" aria-label="High-level flow from customer signals to an orchestrator and specialized agents"><figcaption>Customer signals flow to an orchestrator and specialized agents</figcaption><div class="ma-flow"><div>Customer signals</div><div>Orchestrator agent</div><div>Specialized agents<br>recommendations · ad copy · loyalty offers</div><div>Marketplace decisions and actions</div></div><p class="ma-loop">Outcomes from each action become new signals</p></figure>

The system also needs to learn from what happens after it acts. Did the recommendation help? Did the customer respond to an offer? Those outcomes become new signals that can inform later decisions.

<p class="ma-caption">The full architecture will be available in the upcoming downloadable guide.</p>

How AI Agents Can Respond to Marketplace Behavior

Consider Sophia, a potential buyer who has visited a marketplace several times without making a purchase. The system knows e.g. her region, age, and has signals that she’s interested in workwear and premium brands.

She’s interested. How can the marketplace help her complete a purchase?

<figure class="ma-diagram" role="img" aria-label="Flow 1 — Sophia’s first signals"><figcaption>Flow 1 — Sophia’s first signals</figcaption><div class="ma-flow"><div>Searches for blazers</div><div>Clicks on three blazers</div><div>Interest in workwear increases</div><div>Orchestrator calls the recommendation specialist</div></div><p class="ma-loop">Sophia receives “complete look” recommendations</p></figure>

If Sophia adds an item to her basket but leaves without buying, the marketplace might send a reminder by email. An agent (specialized in offers) could also choose to test free shipping, because that kind of campaign is showing results in her group of customers (not Sophia specifically, but the group of people that have the same overall characteristics) .

That agent’s role is to learn which offer helps customers like Sophia convert at that point in the journey. And here is where knowing how to structure the data (not only having it) makes the difference: the agent will need clear boundaries for what it can test and what it should consider, in with a scope big enough to be specific, but small enough to avoid hallucination.

<figure class="ma-diagram" role="img" aria-label="The orchestrator calls a specialized agent and provides the required brief"><figcaption>The orchestrator prepares the context for the task</figcaption><div class="ma-arch"><div><strong>Orchestrator</strong>Chooses the specialized agent</div><div><strong>Task</strong>What the agent needs to decide</div><div><strong>Signals and context</strong>Relevant information about the customer and the moment</div><div><strong>Boundaries</strong>What the agent can test and consider</div><div class="wide"><strong>Specialized agent</strong>Receives the context it needs to produce a useful response</div></div></figure>

Scaling AI Agents for Real-Time Decisions

In Nancy’s example, the system needed to support up to 300 million users. At that scale, the marketplace has to consider how long a decision takes and what it costs to make it. Nowadays, customers expect the experience to feel immediate. Sophia won't be there waiting if a marketplace she never bought from feels slow or unresponsive.

To keep up with the speed needed in an experience like this, the system can use different paths depending on the information available. If it has recent signals about a customer, it may be able to use information already stored for that user. If that isn’t enough, as mentioned, it can also draw on patterns from a relevant group of customers.

We saw a related idea in a food marketplace Bleu worked with, connecting suppliers with mid-market buyers such as supermarkets. The products also needed a quality check. [Read the case study.](https://bleu.builders/cases/animal-protein/)

Also in that case, a new supermarket may not have browsed enough for the marketplace to understand its preferences. But the system can look at what similar supermarkets in the region buy, then combine that with the new buyer’s signals, such as clicks on premium products.

<figure class="ma-diagram" role="img" aria-label="Flow 2 — Choosing a faster decision path"><figcaption>Flow 2 — Choosing a faster decision path</figcaption><div class="ma-branches ma-three"><div><strong>Recent signals for this customer</strong>Use information available for that customer</div><div><strong>Limited customer history</strong>Combine observed behavior with patterns from similar buyers</div><div><strong>Relevant context sent to the agent</strong>Receive a recommendation without running the full decision process</div></div></figure>

Evaluating AI Agents in a Marketplace

There are several ways to check whether agents are making good decisions. In a scenario of big scale, agents can evaluate other agents. For a smaller business, a human feedback loop may be more practical: a person reviews a proposed action, corrects it when needed, and sends that feedback back to the model.

Bleu used a human feedback loop while creating a second-brain model for a consultancy firm. [Read that case study.](https://bleu.builders/cases/second-brain/) A marketplace serving millions of customers has a different volume, so its feedback process also needs to account for more decisions and more opportunities to catch problems.

<figure class="ma-diagram" role="img" aria-label="Marketplace signals and agent evaluation"><figcaption>Business signals and agent evaluation</figcaption><div class="ma-arch"><div><strong>Marketplace signals</strong>Clicks · journey steps · basket additions</div><div><strong>Agent action</strong>Recommendation or offer shown</div><div><strong>Observed outcome</strong>Customer response · conversion · abandoned step</div><div><strong>Evaluation</strong>Another agent or a person reviews the decision</div><div class="wide"><strong>Feedback to the system</strong>Corrections and outcomes inform future decisions</div></div></figure>

A marketplace doesn’t need to start with a complete agent architecture. It needs to identify which customer decision is worth improving, what information that decision requires, how quickly it must happen, and how the team will know if the result is good.

Where to Start with AI in a Marketplace

Deploying agents across a marketplace operation without prioritizing where they can help can lead to systems that don’t work well in practice. And when the marketplace is already running, introducing agents without careful planning can put logistics and the customer experience at risk.

A safer starting point is to identify one decision worth improving, understand which data can support it, and define how the result will be evaluated.

If your team is exploring where AI could make a difference, the [Bleu AI Guide](https://bleu.builders/guia-ia/) can help you compare opportunities and define a first project to test.

Themes: artificial intelligence · marketplaces · AI agents