AI Game of Credit Cards: Competing for Users Today, Fighting for Access Tomorrow

AI is entering the bank card industry chain from two directions.

According to incomplete statistics from Hub, as of July 23, at least six institutions—including China Merchants Bank, Ping An Bank, SPD Bank, Agricultural Bank, Bank of Communications, and MYbank—have intensively launched credit card, debit card, or account products equipped with AI computing power rights within just over a month.

But in many products, AI is still only a card-opening benefit.

For example, the activity rules for the AI platinum card of a leading major bank show that after a new cardholder spends six times and a total of 5,888 yuan, they can first receive a 20-inch trolley case, and then choose one among a $50 Token voucher, a 300-yuan Starbucks card, and a two-year WPS membership.

In actual selection, Token ranks alongside coffee vouchers and office membership; what the bank has changed is only the card-opening gift, not the transaction chain.

Notably, on the other end of the bank card industry chain, some institutions have already begun experimenting with having AI initiate transactions directly.

In June this year, Visa, the world's largest payment card organization, announced it would connect payment capabilities to ChatGPT. In late June, American online broker Robinhood launched the Agentic Credit Card, allowing users to create independent virtual cards for AI agents, set budgets and approval rules, and have the AI place orders within prescribed limits.

The first route is banks handing AI to cardholders; the second route is cardholders handing cards to AI.

Although these two explorations seem to move in different directions, both contend for the same position: the transaction entry point in the AI era. The question is no longer whether bank cards carry AI, but who can enter AI's decision chain and become its default interface when invoking account, payment, and credit capabilities.

Once product searching, consumption decision-making, and payment initiation are gradually handled by AI, banks may still provide accounts, credit lines, and funds, but lose the direct channel to customers, retreating to backstage roles in the transaction chain.

Shifting Benefits Track

For domestic banks, the first step to compete for the entry point of the AI era is still to attract a batch of AI users into their account system.

Before the current wave of AI cards emerged, the credit card industry had shrunk continuously for more than three years.

According to central bank data, the nationwide stock of credit and quasi-credit cards dropped from 807 million in Q3 2022 to 687 million in Q1 2026, declining for 14 consecutive quarters.

Some independently operated credit card businesses have been reintegrated into branch and head office retail lines.

Past business models relying on card-opening gifts and offline promotions have become increasingly difficult to cover customer acquisition and operating costs, while coffee, audio/video memberships, suitcases, and cashback have become quickly replicable standard configurations.

Tokens have thus become new carriers of benefits; compared with generalized consumption benefits, computing power is more likely to target developers, researchers, content creators, and frequent users of large models.

Banks’ ideal vision is to use Tokens to identify a category of consumers.

Product design is layered according to different business objectives: for example, China Merchants Bank, SPD Bank, and Agricultural Bank of China strengthen their tech customer positioning via model tokens, cloud services, or membership benefits; Ping An Bank embeds AI benefits in account opening, spending, and asset threshold processes; MYbank integrates AI tools into microbusiness scenarios.

Judging from actual results, such benefits already have a screening effect on customer groups.

Liu Pengfei, a researcher at Postal Savings Bank, told Hub that AI benefit cards attract young customers aged 25-35 much more efficiently than traditional card-opening gifts, and compared to customers drawn by traditional physical gifts, AI card customers have higher median education and income.

But Liu Pengfei also noted that many users’ activity drops significantly the month after redeeming benefits, indicating the “freebie hunting” tendency remains obvious.

No issuing banks have yet disclosed actual issuance, benefit redemption rates, or retention at maturity for AI cards. Whether Tokens can continue to drive customer card usage after attracting new clients, and further transform into deposit, wealth management, or credit customers, lacks data validation.

Costs are similarly lacking comparable standards; banks and model providers have not disclosed actual procurement prices or settlement methods, and the claimed hundreds of millions or even billions of Tokens cannot be directly converted into cash value and may shrink rapidly with model price reductions.

More critically, can the data flow back?

Liu Pengfei noted the mainstream model currently is that banks purchase Tokens or membership quotas, and customers jump to the model provider’s app to redeem and use them.

“The data banks can access is usually limited to ‘how many vouchers were issued and redeemed.’ Conversation content, usage frequency, preference tags, and other core behavioral data on the provider side—most banks do not get this.” Liu Pengfei said, “Only a small fraction of banks can obtain desensitized aggregate data via data return agreements, but granularity is far from enough to support precise marketing.”

What banks get may only be a single benefit redemption, not a set of new customer data.

Zeng Gang, Deputy Director of the National Finance and Development Laboratory, told Hub that, based on current implementation, the essence of this round of domestic AI cards is still a new packaging of traffic competition.

Zeng Gang said institutions like CMB, Ping An, SPD deliver Token computing quotas as card-opening gifts, which is no different logically from giving airport lounge access or a trolley case—the AI benefits serve as a customer acquisition hook, not a product core.

Currently, banks face widespread pressure of shrinking credit card scale and falling activation rates; the AI concept provides a rare window for differentiated narratives, which is naturally prioritized by marketing departments.

This does not mean embedding AI into credit cards is an empty proposition right now.

Zeng Gang pointed out that major banks’ intelligent risk control transformation predates this AI card boom by years. Dynamic limit adjustment and real-time anti-fraud run in the background. The real bottleneck is the regulatory boundary and data separation; personal credit decisions demand strict model interpretability, and the black-box nature of large models creates structural tension with existing regulation logic.

“If in the future, regulations loosen their standards for credit-granting AI models, it will be technically feasible to have AI permeate from frontend benefits into backend lending engines—it’s just that the time window hasn’t opened yet.” Zeng Gang said.

Looking at the public product forms, this round of AI cards mainly renovates frontend benefits and marketing.

Tokens improve how banks find specific customer groups but have not yet connected benefit usage, customer data, and credit business into a closed loop. AI benefits help banks compete for users more precisely but have not yet converted these user relations into control of transaction entry points.

AI Starts Swiping Cards

While current domestic AI credit cards are still applied, consumed, and repaid by people, global payment card organizations Mastercard, Visa, and Robinhood have begun advancing something else: letting AI become the actual user of payment tools.

As early as April this year, Mastercard partnered with DBS Bank and UOB in Singapore to complete the first real agent transaction: an AI booked a car for the user to Changi Airport.

In such models, users need not complete search, price comparison, and payment step by step. Instead, they give a task and restrictions, and AI understands their needs, chooses products, and invokes payment credentials within authorized boundaries, thus making the agent a new role in the transaction chain.

Domestic banks start from benefit side, overseas card organizations directly rebuild payment protocols—this seems a divergence of paths.

“They are not mutually exclusive, but the difference is more about stage.” Zeng Gang said.

Zeng Gang noted that Visa and Mastercard launched frameworks like Agent Pay because they themselves are not licensed lenders, and their involvement across the consumption decision chain does not involve credit risk exposure. Modifying payment protocols aligns with their natural capability boundaries.

“Domestic banks play both issuer and funder roles.” Zeng Gang noted, “Entering from the benefit end is a product of both path dependency and regulatory caution, not strategic conservatism.”

The difference between the two paths lies not just in technical progress but also in industry roles: issuing banks first compete for customers via benefits, while card organizations try to gain agency rights for AI to initiate and orchestrate transactions.

Meanwhile, this stage difference is shrinking.

Hub notes that China UnionPay has already released the APOP open protocol framework for agent payments, so the domestic market is also shifting from benefit pilot to infrastructure layout.

Zeng Gang emphasizes: once AI agents can independently complete price comparison, ordering, and payment, the main risk for banks is not fraud but the total mediation of customer relationships.

“An AI layer will be inserted between consumers and banks, and banks’ ability to sense customer behavior, cross-sell touch points, and brand stickiness will be systematically diluted by AI platforms.” Zeng Gang says, “At that point, banks won’t lose to rivals, but to a total reconstruction of the interaction layer.”

This may also change competition between payment credentials: previously, users would actively choose cards for cashback, points, or usage habits; in the future, transaction attribution may depend on which payment credential AI calls by default.

Which card is called first may depend on cashback ratio, available limit, transaction success rate, and platform partnership or recommendation rules.

Banks need to compete not only to be users’ regular cards but also to enter AI’s default payment sequence.

At that point, issuing volume and card exposure may become less important; whether one is the default account, limit, or payment credential AI invokes will directly determine how much business a bank can get.

“‘Degenerating into a funding channel’ is a realistic forecast, but not the only outcome.” Zeng Gang said.

Every revolution in payment infrastructure—from magnetic stripe to chip, from POS to QR codes—has triggered bank card extinction theory, but the result was evolution, not elimination.

“Under the AI-driven payment paradigm, bank cards are most likely to shift from proactive consumer tools to passive assets managed by AI agents; card number, limit, and benefit parameters will become variables agents call in decision trees.” Zeng Gang said.

Cards will still exist, transactions will still go through banks, but transaction entry points and customer interaction may shift to the AI platform.

Who Takes Responsibility

If one day banks truly retreat to the background, a more practical issue will arise: when AI spends money for users and transactions go awry, who is responsible?

In traditional payments, users confirm transactions via password, fingerprint, or face recognition; whether “the person operated it” is often a key basis for judging transaction responsibility.

Agent payments will change this premise.

Users may simply tell AI: buy me a pair of shoes under $200 suitable for long-distance running;

The brand, merchant, delivery, and which card to use may all be decided by AI.

A single agent transaction at minimum needs to answer: who executes the task, whether the user authorized it, to what extent, and whether the final transaction matches original intent.

If AI bought the wrong product, can the transaction be cancelled? If the agent is hacked or exceeds authorization, do losses fall on the user, platform, or bank? If AI shifts recommendations for cashback or commissions, does this require disclosure? When users deny transactions, who proves agents did not overstep?

UnionPay APOP divides agent payments into modules for agent identity, user identity, intent management, and payment authorization;

Mastercard proposes Agentic Token, generating independent digital credentials for different agents and recording/verifying user intent with Verifiable Intent.

These solutions try to make agent transactions identifiable, constrainable, and traceable, and able to be reconstructed in dispute. But all are still at pilot stage and not yet universal standards covering banks, merchants, and AI platforms.

Risk has already shown up at earlier stages.

In the first half of this year, an engineering firm manager in Guangzhou consulted AI about group accidental insurance; AI finally returned an Alipay collection QR code not from an insurer, and 1,618 yuan was paid out.

In another test, a Washington Post reporter tested OpenAI's Operator, asking only to find affordable eggs; AI skipped further verification and placed the order directly. OpenAI admitted the guardrails failed.

Neither incident was strictly agent payment, but they show that even before AI gets full payment rights, misalignments can occur between product identification, merchant verification, user intent, and ultimate execution.

For banks, this is not just a passive exit.

Zeng Gang believes transformation is both threat and opportunity for banks.

“Truly capable banks can evolve into the ‘trusted financial execution layer for AI,’ providing real-time credit, compliance filters, and risk control at the moment of agent-initiated payment, which pure payment accounts cannot replace.”

Zeng Gang pointed out, the card itself will disappear from view, but its underlying credit, funds scheduling, and compliance capabilities may find new reasons for existence in the AI era. The key is whether banks can participate actively in protocol layer formulation, not just wait to be integrated.

This brings the two previously mentioned paths to convergence.

In AI benefit cards, Tokens stand alongside trolley cases and coffee vouchers; the challenge is whether banks can turn one customer acquisition event into lasting customer relationships.

With agent payments, cards recede to the transaction background; the test here is whether banks can embed their credit, risk control, and compliance ability into the new payment rules.

The former decides why customers stay; the latter decides how banks continue to stay in the transaction chain.

In the AI era, the biggest change for bank cards may not be disappearance, but no longer being directly visible to users. True competition will shift from card-vs-card to who becomes the default interface AI calls for payment, account, and credit capabilities.

Risk warning and disclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice, nor does it take into account the special investment goals, financial circumstances, or needs of individual users. Users should consider whether any opinion, viewpoint, or conclusion in this article fits their particular situation. Invest accordingly, at your own risk.