Artificial intelligence (AI) is moving deeper into banking operations. Yet deploying it at scale requires more than computing capacity and sophisticated models. Banks must make fragmented enterprise data usable for business decisions, control AI-generated recommendations and protect digital customer interactions from increasingly sophisticated fraud.

These challenges framed the first day of The Asian Banker Beijing AI Innovation Study Tour 2026. Delegates visited China Construction Bank (CCB) to examine its data infrastructure and enterprise AI applications, followed by computer vision company MEGVII to explore biometric authentication, deepfake detection and intelligent monitoring.

CCB integrates real-time data processing with enterprise AI

Wang Lixin, general manager of CCB’s Data Management Department, welcomed delegates and outlined the bank’s use of technology across its businesses. CCB reported serving about 790 million individual customers and 14 million corporate customers. Its AI applications span wealth management, inclusive finance, risk management and technology development, as the bank moves from isolated use cases to deeper integration with business processes.

Representatives from CCB’s Data Management Department and CCB Fintech presented two initiatives: an integrated streaming and batch real-time data warehouse, and an AI-powered system known as the Digital Brain.

CCB described three generations of data infrastructure, progressing from traditional warehousing to big-data architectures separating computation from data access, and then to cloud architectures separating computing and storage. In 2026, it further developed its data supply chain through what it called a “water-infrastructure” approach, seeking to deliver information proactively to business functions rather than respond only to individual requests.

The real-time data warehouse combines real-time, micro-batch and T+1 processing modes through unified storage and access. This allows business units to obtain information at different speeds according to operational requirements rather than depend exclusively on overnight processing. CCB reported query response times of less than 500 milliseconds, with applications spanning credit cards, wealth management and deposits.

To support reliability, the architecture incorporates dual centres and dual channels, end-to-end monitoring and mechanisms for recovering or replenishing missing data. These controls are significant because faster access to information also increases the need to maintain data completeness and consistency across processing environments.

CCB’s second initiative, the Digital Brain, addresses limitations of conventional business intelligence tools. Traditional systems may require specialist technical knowledge, struggle to analyse data across business domains and take a month or longer to meet new data requirements. They may also present information without explaining its business significance.

The Digital Brain follows a five-stage closed loop covering event sensing, analysis, decision recommendations, business action and feedback. It combines large language models and agent frameworks with governed enterprise data, a semantic layer, isolated execution environments and data-item-level access permissions.

Business users can therefore conduct complex cross-domain analysis using natural-language requests rather than assemble data manually. According to CCB, the system had been operating for about six months and consumed around six billion tokens daily. The bank reported substantial improvements in analytical efficiency, although it did not provide a standardised measure of the resulting financial benefits.

Human oversight remains central to CCB's AI governance

CCB distinguished technological capability from permission to automate decisions. The bank reported implementing about 400 AI use cases in 2025 and described its continuing development of a unified agent platform and privately deployed AI infrastructure. It maintains human review of generative AI outputs, particularly where decisions may affect customers.

CCB representatives distinguished between internal AI applications and those delivering AI-generated outputs directly to customers. They said that, under the approval framework described during the session, customer-facing applications were subject to additional oversight and approval, reflecting the need for accountability when automated outputs may affect customers.

The distinction is also evident in fraud prevention. CCB explained that AI is used to develop and refine detection rules, which can subsequently be embedded in transaction-blocking processes. Generative AI does not directly make real-time blocking decisions. Graph computing is already used to identify suspicious relationships, although explaining these findings to customers remains a challenge.

Delegates from several countries and regions raised questions on the transition from T+1 to near-real-time processing, credit-decision model performance, no-code agent development, AI value measurement and the latency of graph-based fraud detection.

The exchanges raised a central question for banks: when should an AI recommendation be allowed to trigger a consequential action? Institutions must decide where established rules support automation, where human judgement remains necessary and how decisions can be explained and audited.

MEGVII strengthens biometric verification against generative AI fraud

At MEGVII, the focus shifted from enterprise data to digital identity and the security of customer interactions. Founded in 2011, the company develops computer vision technologies spanning facial recognition, mobile-device authentication, smart buildings and logistics automation.

The company introduced FaceID, its biometric verification solution used in electronic know-your-customer (eKYC) processes, remote account opening and higher-risk transactions. Its Face++ platform also provides facial recognition capabilities through application programming interfaces, allowing developers to integrate the technology into other services.

MEGVII demonstrated how facial verification combines identity matching with liveness detection to establish whether a person is genuinely present. Methods include prompted facial movements and randomised light patterns displayed on a smartphone screen. The system analyses facial responses and reflections to help distinguish a live user from photographs, prerecorded videos or manipulated images.

As generative AI makes manipulated images and videos easier to produce, MEGVII discussed threats involving AI-generated video, face replacement and altered camera feeds. It described combining biometric analysis with device-risk signals, behavioural information, dynamic blacklists and checks for suspicious patterns across authentication attempts. The session did not establish comparative detection accuracy or measurable reductions in banking fraud.

The company also discussed the practical limits of biometric systems. Delegates questioned how facial recognition performs when customers' appearances change over time, how confidence thresholds should be managed and what happens when legitimate users cannot complete liveness checks. MEGVII explained that facial landmarks can remain useful despite changes in appearance but acknowledged that significant changes and difficult capture conditions can affect matching performance.

Customer accessibility was another concern. Some retail users struggle to follow facial movement instructions, while changing lighting conditions and unstable camera positioning can complicate verification. MEGVII described manual review as one way to resolve cases that automated systems cannot determine confidently.

Beyond financial services, delegates viewed demonstrations of intelligent cameras, physical access controls and logistics automation. These applications showed how computer vision can be deployed across different operating environments, although the banking discussion concentrated on identity verification and fraud prevention.

Building reliable AI requires control over data, decisions and identity

The two visits illustrated different challenges in applying AI to banking. CCB is integrating real-time data and AI into enterprise analysis and decision-support processes, while retaining controls over consequential actions. MEGVII demonstrated technology intended to strengthen identity verification at the customer interface, alongside the practical limitations of biometric authentication.

Neither the number of AI applications nor the sophistication of a demonstration establishes commercial value. For banks, the more demanding test is whether these systems improve operational performance, reduce fraud and remain accountable at scale. CCB reported efficiency gains without a standardised measure of financial benefit, while the MEGVII presentation did not quantify fraud reduction. Those results will matter as much as the underlying technology.