推动主动型商业发展的四大层级

生成式人工智能目前仍主要处于研究和探索阶段，最终的点击行为——至少在现阶段——仍发生在传统网站上，且由人类而非机器完成。  

然而，尽管公众讨论仍主要聚焦于面向消费者的购物助手，但支撑智能代理大规模搜索、谈判及支付辅助的基础设施转型已然启动——尤其在B2B流程中，因其涉及的授权与责任归属更为明确。  

早期试点表明，电子商务正迅速演变为多代理生态系统，其中公共代理、商户代理、支付服务提供商（PSP）代理和卡网络代理实时进行协商与交易。该基础设施的每个层级都承担着不同的角色并拥有控制点。随着代理型商业规模的扩大，理解这些要素将成为决定商户竞争优势的关键。  

第一层——消费级人工智能平台：探索领域的新守门人

到2025年底，ChatGPT、Perplexity等公共人工智能平台已成为重要的发现渠道，对消费者选择光顾哪些商户的影响日益显著。  

2025年9月29日 ，万事达卡的Agent Pay服务登陆 ChatGPT平台，使美国持卡人能够直接在聊天界面完成购物。与此同时，OpenAI与 Shopify 达成合作 ，让 超过百万家商户可通过ChatGPT的对话 流程被发现并支持购买。  

与人类不同，人工智能代理不会进行有意义的滚动或解读视觉层次结构——ChatGPT、Perplexity、Claude、Gemini等模型从信息流、API及schema.org标记中提取结构化数据。  

尽管商家品牌团队倾注了诸多心力，但对于代理商而言，主页、横幅广告和用户体验流程都显得微不足道。

相反，代理将产品视为表格中的行，其中包含可解析的属性、价格、库存和政策等信息，从而能在多个商家间进行比较。  

商品信息中属性覆盖率接近100%（95%以上）的商家，其产品在AI推荐中的可见度通常高于属性信息不全的商品目录。另一方面，智能助手会自动跳过缺少关键数据的商品，例如配送时效、尺码信息或退货政策等。  

提升代理式商业中商家曝光度的最佳实践：

1. 通过数据源和schema.org产品标记，完整呈现结构化的产品数据，确保代理商能可靠地解读产品目录。  
2. 将目录结构与自然语言查询（例如“200英镑以下且周五前送达的防水徒步靴”）保持一致，而不仅遵循内部商品陈列逻辑。
3. 保持库存与价格的实时同步。平台会对过时或不准确的数据进行惩罚，并迅速降低那些经常虚报库存或配送承诺的商家的排名。
4. 提供支持智能代理的结账API，或采用新兴的智能代理商务标准，以避免脆弱的HTML抓取。
5. 在维护欺诈防御机制的同时，将可信代理加入白名单，从而区分有益的自动化操作与滥用型机器人。

随着智能商业规模的扩大，那些专注优化机器可读数据而非（仅）依赖人工说服力的商家，将在智能驱动的发现领域占据主导地位。而那些仍只为页面浏览量而建的商家，则可能面临被市场忽视的风险。

第二层——商户代理：从 可见性到成果掌控

即使你已为公共代理商完成所有优化，或许能掌握可见性，却仍无法掌控品牌体验、交易结果或信息传递给消费者的方式。近期行业调查显示，多数大型零售商预计代理支付将在三年内成为主流，但许多企业尚未明确其系统如何处理大规模运营中的代理商发起购买、售后修改或退款事宜。

这正是定制化商户代理发挥作用之处。  

当公共代理提出"周二前送达120英镑跑鞋"等请求时，商家代理会解读需求意图，将其映射至商家的商品目录和物流系统，并整合商家能提供的最佳方案。 该系统能实时管理库存限制，选择满足时效要求的配送与履约方案，并在产品展示与权衡方案中保持品牌调性。商户代理不再被动依赖公共代理抓取页面信息进行猜测，而是成为主动的谈判方——当无法精准匹配时，可灵活调整捆绑方案、应用促销活动并提出替代方案。  

随着时间推移，商户代理很可能成为支付服务提供商（PSP）和支付方案的默认接口。它们将实时呈现库存、定价和风险信号，供基础设施其他层级据此进行优化。在此模式下，公共或经纪代理负责协调多家商户，而各商户代理则专注于最大化自身业务的转化率、利润率和客户体验。  

在智能代理商务领域，未来属于那些构建智能代理的主体——他们能主动塑造交易形态 并掌控结果，而非被动等待被发掘。

第三层 - 支付代理：运营智能中心  

Agents within PSPs and global payment platforms like Nuvei increasingly handle the operational intelligence that turns intent into money movement at scale. They own fraud detection, routing optimization, authorization performance, dispute management, reconciliation, treasury decisions, and compliance enforcement across thousands of merchants and millions of transactions. As these control points converge in agentic commerce, the intelligence payment platforms will generate can compound across the entire value chain.  

A payment agent can operate within one of the following control points:

• Checkout and funding. Determining whether a transaction should be approved, which funding source to use, and which fraud signals to apply in real time. Intelligence here means learning new patterns of agent behavior and adapting as agents evolve. ‍
• Authorization and routing. Deciding whether to send a transaction through 3D Secure, which acquirer or route to choose based on live performance, and whether to bid dynamically on fees and routing paths. Recent case studies show that AI‑driven routing and risk optimization can reduce fraud losses by more than half and lift approval rates enough to deliver high‑single‑digit revenue uplifts for some merchants, especially in cross‑border and higher‑risk segments. ‍
• Post‑payment controls. Optimizing chargeback evidence, fund release timing, and liquidity management, with decisions that compound across thousands of transactions per day.

PSPs and payment platforms with globally distributed data, multi‑acquirer global setups, and embedded AI decisioning will be better positioned to train agent‑aware models across geographies and use cases.

The intelligence received from such PSPs can become a shared asset for merchants who want to benefit from agentic commerce without building every capability themselves and the entire infrastructure powering agentic commerce.  

Layer Four - Scheme Agents: Encoding trust and standards

Card networks are evolving beyond passive payment rails into intelligent orchestration layers that differentiate agents from human transactions and apply context‑specific security models.  

For example, in October 2025, Visa unveiled its Trusted Agent Protocol (TAP) - developed with Cloudflare - to provide cryptographic verification for AI agents during browsing and checkout. With it, merchants and PSPs can distinguish trusted agents from malicious automation with minimal changes to their infrastructure.

Meanwhile, Mastercard is working with partners including Microsoft, IBM, and Google to scale agentic commerce globally, and has announced plans to expand Agent Pay across Latin America.  

Scheme agents are beginning to detect and classify “agent-present” transactions via: ‍

• ‍Trusted Agent Protocol (Visa): Cryptographic signatures verify agent identity during browsing ‍
• Agent Pay (Mastercard): Special credentials prove "I'm ChatGPT acting for Alex B"

Scheme agents spot agent traffic through cloud IPs, automation fingerprints, and protocol flags. These distinguish legitimate agents from humans and malicious bots, then apply agent-specific rules such as lower friction for trusted agents, tighter scrutiny for unknown ones. They also coordinate authentication by preserving agent identity signals through the PSP → issuer → settlement flow.  

What remains undeveloped is broad, interoperable adoption of these standards, as well as clear liability rules that define what happens when an authorized agent makes a harmful decision on behalf of a consumer – or a business.  

As agentic commerce scales, scheme agents will increasingly be responsible for coordinating authentication, applying context‑aware fraud models, and enforcing emerging “agent‑present” categories that sit alongside today’s card‑present and card‑not‑present distinctions.  

At the same time, schemes are exploring how card network protocols (such as TAP and Agent Pay) could extend toward APM interoperability.  

Is your payment infrastructure ready for agentic commerce?

Industry surveys suggest that close to 60% of banks and large corporations expect agentic payments to be mainstream within the next three years, with early adoption clustering around recurring billing and B2B purchasing flows.  

It is likely only a matter of time before these patterns extends more broadly into B2C, particularly for repeat and low-risk purchases. eCommerce is evolving into a multi-agent ecosystem, where human intent is increasingly expressed indirectly rather than through clicks.

For merchants, the question is less whether agents are coming and more whether their infrastructure will be ready when agents become a primary interface for purchase.

‍

Asaf Ben Gal is Director of AI & Analytics at Nuvei, leading the company’s AI strategy and applied machine learning initiatives to turn advanced technologies into measurable business impact.