An artificial intelligence (AI) system that answers a question incorrectly presents one kind of risk. A system authorised to act on a bank’s data, transactions or customer relationships presents another. The shift from generating responses to executing tasks makes controls, accountability and operating economics central to AI adoption.

On the fourth day of The Asian Banker Beijing AI Innovation Study Tour 2026, delegates visited NetEase and Pony.ai. The discussions ranged from supervising AI agents and running driverless vehicles to connecting customer information with banking workflows. Across these different applications, the question was not simply whether AI could perform a task, but whether it could do so safely, consistently and at an acceptable cost.

NetEase tests how banks can control AI agents and their outputs

The morning visit to NetEase focused on how financial institutions can manage the security risks created by wider use of generative AI and autonomous agents.

NetEase has developed enterprise technology services from capabilities originally supporting its gaming, entertainment and other digital businesses. Its GuardEase security business provides content moderation, AI security and related services to external organisations.

The company described how AI adoption in banking is progressing from predictive models and employee assistants towards agents that can retrieve information, interact with systems and perform operational tasks. This transition changes the nature of the risk. An inaccurate chatbot response may misinform an employee or customer. An agent with access to business systems could also disclose confidential information, initiate an unauthorised action or execute a workflow incorrectly.

NetEase grouped the main challenges into data security, unreliable model decisions, attacks against AI systems and inappropriate agent behaviour. Its proposed controls include monitoring inputs and outputs, protecting sensitive information, detecting suspicious instructions and evaluating actions throughout the AI lifecycle.

The approach extends conventional content moderation into the operation of AI applications. Rather than checking only whether generated text is acceptable, banks must also consider what information an agent can access and which actions it is permitted to take.

NetEase cited a large Chinese bank that introduced automated screening of AI-generated and other digital content. The presentation reported high review accuracy and a substantial reduction in the content-review workforce following automation. However, the company did not specify the accuracy benchmark, the period of the staffing change, changes in employee responsibilities or the resulting financial savings.

Delegates questioned whether AI could reliably supervise another AI system, particularly if both models shared weaknesses. NetEase said its security controls were designed to make bounded decisions—allowing, rejecting or escalating an interaction—rather than hold unrestricted conversations. Narrowing the task can make performance easier to evaluate, although the supervising system still requires independent testing.

Another exchange concerned the balance between detection accuracy and false positives. NetEase acknowledged that stricter thresholds can block legitimate activity and said banks ultimately determine their preferred risk settings.

For banks, the relevant test is whether additional security controls detect harmful activity without excessive false positives, delay or interruption to legitimate workflows. That requires clear escalation procedures and measured performance under operational conditions.

Pony.ai demonstrates commercial driverless operations, with profitability still unproven

At Pony.ai, delegates experienced a driverless vehicle operating along a designated route. During the journey, the system performed driving functions without a human at the controls, providing a practical example of AI taking action in a physical environment.

Unlike a controlled simulation, autonomous transport requires AI to interpret its physical surroundings and respond to road conditions in real time. Its performance depends on the integration of perception, navigation, vehicle controls and safety mechanisms.

The successful journey established an operating capability on that route, not reliable performance under every road or weather condition or the commercial viability of a wider fleet.

Delegates questioned the continued presence of conventional vehicle controls, emergency procedures and how passengers could obtain assistance if something went wrong. These questions reflected a broader concern about transferring responsibility from a human operator to an automated system.

The discussion also covered the limits of autonomous driving in difficult environments, including roads with limited infrastructure, unusual obstacles and adverse conditions. The ability to complete a journey on a designated route does not establish equivalent performance in every operating environment.

Delegates also discussed the difference between positive economics for an individual vehicle and profitability at company level. Pony.ai reported that its Guangzhou and Shenzhen operations had reached city-wide unit-economics breakeven, in November 2025 and February 2026 respectively. In its second-quarter 2026 results, the company reported 1,975 robotaxis in its fleet as of 30 June and robotaxi services revenue of $12.1 million for the quarter. These company-reported operating results establish a commercial business beyond the demonstration, but do not establish company-wide profitability. Fleet costs extend to sensors, computing equipment, maintenance, insurance, customer support and remote supervision.

Pony.ai’s development of autonomous driving also raises questions about how the technology will be updated. Hardware replacement cycles, changing sensor configurations and software improvements can affect the cost and useful life of a vehicle.

Delegates raised questions about insurance and responsibility in the event of an accident. The discussion pointed to the need for insurance arrangements suited to driverless operations, alongside conventional coverage considerations.

A further concern was cybersecurity. Centralised systems supporting multiple vehicles create operational advantages but also raise questions about the consequences of unauthorised access and the safeguards needed to prevent remote interference.

For banks, Pony.ai offered a comparison with AI agents performing financial tasks. Both involve systems moving beyond recommendations into actions with real-world consequences. Their governance depends on clearly defined operating boundaries, intervention mechanisms and evidence of reliable performance.

The visit did not establish comparable fleet-wide accident rates, utilisation or the cost of driverless transport against conventional services. These measures remain relevant to assessing safety and sustainable profitability as commercial operations expand.

Banks must measure what AI changes in production

The visits exposed three distinct tests of AI deployment. NetEase presented a banking example in which automated review coincided with a smaller manual workforce, without a quantified financial benefit or fully defined performance benchmark. Pony.ai demonstrated an existing commercial capability while leaving company-wide profitability and comparative safety performance unresolved.

The common constraint is the distance between technical capability and repeatable operational results. Banks need to know whether security tools reduce harmful activity without obstructing legitimate customers, whether autonomous systems remain controllable when conditions change, and whether AI-driven engagement outperforms established analytical methods.

The more consequential question is where an AI system should be permitted to act. Permissions, human intervention, audit trails and consistent data definitions matter as much as model performance when automation reaches customer-facing or regulated processes.

As the technology moves into production, measurable outcomes—not the sophistication of a demonstration—will determine its value. For banks, the case for AI must be made in lower costs, better customer outcomes, stronger controls or improved productivity, with those gains assessed against the risks and costs of implementation.