How AI Is Transforming Marketplaces into Intelligent Agents: The Future of Digital Platform Architecture

Over twenty years, Sergei Serzhan built several digital marketplaces across different industries — furniture marketplace DOM98, financial aggregator Credits.ru, and software distribution platform FreeSoft, which now reaches over two million users per month. In this article, he explains how AI is changing the fundamental logic of how marketplaces work — and why this applies to platforms in any industry.

The Scale Problem

Most digital marketplaces are built on the same logic: a vendor publishes a product, a user enters a query, and the platform returns a list. The problem emerges when the catalog grows to tens of thousands of items.

A FreeSoft user looking for a lightweight PDF editor without a subscription gets back dozens of apps with near-identical descriptions, an average rating of 4.2, and interface screenshots. A Credits.ru user arrives with a specific situation — a loan amount, a repayment period, specific terms — and receives a table of offers from dozens of banks. In both cases, the platform returned results. It did not solve the user’s problem.

The operational side looks the same. FreeSoft has 41,800 apps and 26,000 active developers, each publishing updates, submitting applications, and managing their pages. Every listing requires ongoing work: reviewing submissions, moderating content, enriching metadata, updating descriptions, and handling reviews. That volume grows with the catalog — linearly and predictably. The team size does not scale with it.

Adding filters or hiring more moderators does not fix this — it only buys time. The constraint is built into the model itself: the larger the catalog, the harder it is for users to find what they need, and the harder it is for the team to maintain quality across every listing.

This pressure — from both sides simultaneously — raised the question of changing the architecture. Not at the level of individual features, but at the level of how the platform interacts with users and how the work inside it is organized.

The Path to AI-First

When LLM models started entering the market, FreeSoft took a skeptical position for a long time. Text generation produced inconsistent results. AI-based moderation missed errors that a deterministic rule caught reliably. Many tasks were faster and cheaper to handle with traditional engineering approaches. About a year ago, model quality reached a point where they could be used in real production workflows — not in experiments, but in production.

The turning point came from observing a shift in user behavior. Children started interacting with devices through voice and AI assistants as their primary interface — not as an added feature, but as something entirely natural. The next generation of users will look for software not through a catalog search bar, but through AI agents. A platform whose data is not structured in a way that those agents can understand and recommend simply does not exist for that audience.

At FreeSoft, apps, categories, developers, and use cases began to be organized as interconnected entities rather than isolated database records — a unified knowledge system instead of a collection of separate pages.

AI as an Infrastructure Layer

At FreeSoft, AI is now embedded across most workflows — content generation, SEO, moderation, processing developer submissions, metadata enrichment, and user support. AI does not operate autonomously: it executes tasks within pipelines that the team defines and controls. The team focuses on setting tasks, configuring pipelines, and reviewing outputs — AI handles execution.

This required a shift to an entity-based data model. Previously, every app in the catalog existed as an isolated page. Now apps, categories, developers, features, and use cases are interconnected entities within a single system. RAG-based systems operate on top of this layer: when generating content, AI retrieves relevant context from the platform’s knowledge base — related apps, categories, historical data — and uses it to produce accurate, consistent output.

AI does not reduce the budget for development and operations — it increases productivity within the same budget. The same team processes a higher volume of submissions, ships features faster, and manages a growing catalog without proportional headcount growth. The resources freed up are reinvested in product development.

From Catalog to Agent

Search went through a similar evolution. Early search engines required an exact match between the query and the page text. Today, a user types “buy a car” and sees listings that say “sell a car” — the system understands intent, not just words. The next step follows the same direction: the user describes a task, the platform returns a ready solution.

At FreeSoft, platform data is being structured into unified knowledge layers that support AI retrieval and reasoning systems. A user will be able to describe a task — find the right software, configure a tool, work through a problem — and receive a direct answer without navigating the catalog. The interaction is not limited to a web interface: voice, messaging apps, and integration with other AI agents.

This also changes how platforms need to think about organic growth. Platforms used to optimize for search engines — URL structure, metadata, backlinks. Now there is another layer: structuring data so that AI agents can understand it. A platform that does not do this gradually disappears from AI system recommendations, the same way a website without SEO disappears from search results.

This applies to marketplaces of any type — software, financial products, or goods. A platform that structures its data as a connected knowledge system gains the ability to work as an agent: understanding the context of a request, not just its words.

AI is not a universal solution. Where a deterministic process is sufficient, traditional automation is more reliable and less expensive. AI adds real value where flexibility, reasoning, and work with unstructured data are required. Platforms that understand this boundary build more predictable and manageable systems — without unnecessary complexity where it is not needed.