G2A.COM on 14,400 cases and what its AI assistant taught it

G2A.COM, the digital entertainment marketplace, has rolled out its AI support assistant to sellers in 180 countries after a trial in which it helped resolve more than 14,400 seller support cases in 63 days. The company says the assistant cut the volume of tickets needing a human by 27 to 30 per cent, routed cases with 93.8 per cent accuracy and scored 4.16 out of 5 for satisfaction, while handling around 10,000 conversations a week.

Paweł Wróbel, chief growth officer at G2A.COM, answered written questions from Disrupts on what the trial taught the company that it did not expect, why the assistant is grounded in verified transaction data rather than generative text alone, where humans stay in the loop, and what an AI-native marketplace means beyond customer support.

The trial was designed around quality, Wróbel said, with the team testing scenarios that included the unintended ways users might interact with the assistant, "knowing that real-life use would not always follow the paths we had designed". The performance was known; the open question was acceptance. He said 91.2 per cent of users rated their conversation positively, based on feedback collected after each interaction, and that achieving this with the first version "was a strong validation of our approach".

The unexpected lesson was about preference rather than capability. "We started out thinking mainly about capabilities, what we could give our sellers to help them handle customer cases faster. We came out thinking mostly about choice. Some of the sellers' customers appreciate an instant, accurate answer at any hour; others want a human from the very first message, and no quality of AI response changes that. Respecting both is now a design principle."

The second finding was how clearly the boundary drew itself. For straightforward, informational cases the assistant supports sellers well. Complex or nuanced matters still call for human judgement, so those cases route onward with the AI preparing context rather than answering. Wróbel described both as things the company continues to observe and adjust with sellers rather than settle once.

The assistant is one of three AI tools G2A.COM now runs across the user journey. Buddy helps users discover the right seller offers, Stanley helps them navigate the platform's features, and the newest assistant helps sellers resolve customer issues. "Different use cases, one common ambition: to make digital commerce more seamless, relevant and trusted."

A wrong answer is worse than no answer

Unlike a generic chatbot, the assistant works from verified transaction data and seller policies. Wróbel's explanation of that decision was blunt. "In a marketplace, a wrong answer is worse than no answer. A generic chatbot generates plausible text; sellers need answers that are true for the specific order and case they are handling." Grounding every answer in facts the system can check, he said, rather than "a statistical guess about what an answer usually looks like", is what separates it from what is on the market.

What that rules out is the thing that erodes trust fastest: improvised answers. The assistant cannot invent an order status, promise a refund outside policy or describe a procedure that does not exist. If the data it holds does not support an answer, the case goes to a human. "That boundary is built in, not bolted on."

It also rules out shortcuts for G2A.COM itself. Grounding requires structured, governed data behind the agent, which is slower to build than plugging a model into a chat window. "But we see that effort as an investment. It means we are building a solution which can be fully controlled, one that can grow and evolve in the future."

Scaling without hiring in proportion

Demand spikes around launches such as GTA VI can arrive within hours. The assistant handles the queries and requests that do not need direct human intervention, which reduces pressure on support teams and lets specialists concentrate on cases where their judgement is needed. "From a scaling perspective, this means we can increase our capacity to support sellers without having to grow the team proportionally to every increase in demand," Wróbel said.

A 93.8 per cent routing accuracy still leaves cases that go astray. Wróbel did not claim otherwise. When a case is misrouted or needs judgement the system cannot provide, it is redirected to a human specialist; complex, unusual or sensitive cases stay under human oversight throughout. "The accuracy rate is therefore not the end of the process, but one measure of how effectively AI supports it. We continuously learn from cases that do not follow the expected path and use those insights to improve the system over time."

Human versus human plus AI

G2A.COM describes itself as building an AI-native organisation. For Wróbel that is not about adding tools to existing processes or running AI as a separate technology project. "It means fundamentally rethinking how we work and asking: if we were building G2A.COM today, with AI available from day one, which processes would we design differently?"

AI is already applied across security and fraud prevention, seller verification, data analysis, marketing, product development and repetitive workflows. The goal, he said, is to redesign processes so that people and AI work together rather than to automate individual tasks. "The future is not 'human versus AI'. The real competitive shift is human versus human + AI." Where AI lands next is therefore not one department: he expects it to become embedded wherever it helps the company make better decisions, identify patterns faster, personalise experiences, improve security and reduce repetitive work.

More intelligent, not more visible

His advice to other marketplaces weighing agentic AI is that "more AI should mean a more intelligent marketplace, not necessarily more AI visible to the user".

He drew on G2A.COM's Trust Algorithm research, published earlier this year with Juniper Research and based on more than 9,000 consumers in nine countries. Consumers in every market surveyed preferred marketplaces to buying directly from brands; 50 per cent said protecting them from fraud is primarily the platform's responsibility, against 20 per cent who placed it on themselves; and 39 per cent said they feel uncomfortable when AI moves from assisting with a task to making a purchasing decision on their behalf.

"Trust should ultimately be a consequence of a good experience, built into how the technology works, rather than something we ask users to give us upfront," Wróbel said. "The best use of AI is not always the one users see, it is the one that makes the experience more trusted, seamless and useful." That is also why G2A.COM is building its own capabilities rather than plugging in off-the-shelf tools: it gives the company control over reliability, data, governance and how the technology evolves with its users.

The assistant is available to sellers across all 180 countries in which G2A.COM operates, and the company says it is now handling around 10,000 conversations a week.