Retail, 30 September 2026
G2A.COM: An Autonomous Support Agent That Resolves Seller Tickets Without Human Review
G2A.COM — an online marketplace for game keys, software licenses and digital vouchers, used by independent sellers in 180 countries.
Problem
Seller-support demand spikes unpredictably around major game launches, when thousands of buyers report voucher, key-activation or refund issues within hours of each other. A support team sized for normal volume can’t absorb that kind of spike without service quality collapsing — and every case first requires checking the specific seller’s real transaction history and policy terms before it can be resolved correctly.
Result
54%
- 54% of seller support cases resolved by Dave with no human involved, out of roughly 14,400-15,000 tickets handled during the 63-day trial.
- Ticket volume reaching human Seller Support fell by 27-30%.
- Routing accuracy of 93.8% and an average quality score of 91.2%, with only 0.81% negative feedback; users rated the experience 4.16/5, and 74% called the process easy or very easy.
- Runs 24/7 across all 180 countries where G2A operates.
Solution
G2A built “Dave,” a proprietary in-house support agent — not a generic chatbot — that connects directly to each seller’s transaction ledger. Rather than resolving a case from generated text alone, Dave first pulls the seller’s real payment status, voucher/key state and applicable policy, and only then decides and executes an action: refund, reissue or escalation to a human. The company says grounding every decision in verified transaction data, instead of letting the model answer from the prompt alone, is what keeps it from hallucinating on financially and legally sensitive cases involving vouchers, licenses and refunds. It ran a two-month internal trial before its global rollout on July 30, 2026.
Could a small team do this?
Partially. The principle behind Dave — ground every decision in a live lookup against the real system of record (order history, payment status, policy rules) instead of letting the model freelance from a prompt — is exactly the kind of agent a solo operator can build: a support bot wired to a database or API via function-calling tools, with clear rules for what it resolves automatically versus hands off to a human. That’s realistically a few days of work for one narrow, well-defined ticket type (e.g. “was this voucher already redeemed — refund if not”) backed by a handful of API integrations. What isn’t replicable solo is the scale: 180-country, 24/7 coverage across an unknown number of transaction systems and languages, plus enough historical ticket data to tune routing accuracy above 90%. A realistic solo version would pick one narrow, high-volume, data-verifiable ticket type — refund status, subscription cancellation, order tracking — and prove the same “check the record, then act” pattern on that alone before trying to generalize further.