Summary

• G2A says Dave handled around 15,000 tickets during its first 63 days and reduced human support demand by 27–30%.
• Access to transaction and policy data is more consequential than the system’s conversational interface.
• All performance measures are company supplied, while the published material differs over whether Dave handled 14,400 or approximately 15,000 tickets.

G2A has rolled out an autonomous support agent to marketplace sellers across 180 countries, connecting the software to transaction records and commercial policies so that it can resolve routine cases before they enter a human queue.

G2A.COM says the agent, named Dave, now manages roughly 10,000 conversations each week and operates around the clock. The global deployment follows a two-month trial intended to help independent sellers handle sharp increases in support demand.

According to G2A’s figures, Dave resolves or streamlines 54% of cases and has reduced the volume requiring human seller support by between 27% and 30%. The company reports routing accuracy of 93.8%, a satisfaction score of 4.16 out of five, and negative feedback on 0.81% of responses.

Those measures have not been accompanied by an independent audit, methodology paper, or control group. The release headline refers to 14,400 tickets during 63 days, while the body describes approximately 15,000, a discrepancy that should be resolved before either total is treated as precise.

Transaction access changes the system’s role

Dave connects to transaction history, payment status, and seller policies, allowing it to handle cases involving vouchers, digital keys, software licences, and seller-managed refunds. Access to verified records distinguishes the system from a chatbot limited to producing conversational answers.

Grounding an agent in transactional data can reduce some forms of hallucination, but it cannot remove error. The system may misunderstand a request, retrieve the wrong record, apply an outdated policy, route a case incorrectly, or produce an explanation extending beyond the evidence available.

Data quality therefore becomes part of support quality. Sellers need accurate policies, properly recorded payment states, and consistent refund information, since an agent connected to incomplete or stale records can answer quickly while producing the wrong result.

G2A says each response is assessed automatically across 13 quality and safety dimensions. More detail would be needed to establish whether those checks rely on deterministic rules, another model, human sampling, or a combination, and how often automated scores disagree with customer outcomes.

Fewer tickets do not remove the workload

Customer support automation has progressed from decision trees and simple chatbots towards retrieval systems and agents capable of completing bounded actions. Dave continues that established progression rather than creating a new direction for enterprise service technology.

A reduction of 27–30% in tickets reaching human teams could produce meaningful operational savings, although it does not establish an equivalent fall in labour or cost. The remaining cases may become more difficult once straightforward enquiries have been removed, while staff must also monitor the agent, maintain policies, investigate failures, and manage appeals.

G2A says cases requiring manual intervention are transferred with an analysis and summary so users do not have to repeat their account. A functioning handover can prevent the familiar experience in which a customer completes an automated conversation before starting again with a person.

Routing accuracy of 93.8% still leaves a material minority of cases initially sent to the wrong seller or support function when applied across large volumes. The operational effect depends on how quickly errors are identified and whether misrouting delays refunds, reveals information, or obscures responsibility between G2A and a merchant.

Marketplace responsibility remains divided

G2A connects buyers with independent sellers, creating boundaries between marketplace operations, merchant decisions, payments, and customer rights. An agent interpreting seller policies must preserve those distinctions without sending users indefinitely between different organisations.

Seller autonomy also limits central automation because Dave can use policy and transaction information while merchants remain responsible for many commercial decisions. The system needs clear authority boundaries covering actions it can complete, actions requiring approval, and cases that G2A must handle because they concern the operation of the marketplace.

Availability across 180 countries introduces differences in language, consumer law, payment systems, tax, and acceptable remedies. Multilingual capability may translate a conversation, but it cannot assume that legal rights or seller obligations are identical across jurisdictions.

The metrics disclosed by G2A provide more operational detail than many enterprise AI announcements because they cover ticket reduction, routing, satisfaction, quality scoring, and negative feedback. Stronger evidence would include repeat contacts, resolution durability, escalation time, refund errors, cost per case, seller outcomes, and comparisons with human-only handling.

The early figures suggest that a transaction-grounded agent can absorb a meaningful share of marketplace support demand. Its longer performance will depend on whether accuracy holds during major traffic spikes, across jurisdictions and languages, and when cases move beyond the clean records and defined policies on which automation works best.