Established firms set the boundaries for AI shopping agents

Testing the limits of autonomy

Visa recently conducted a benchmark to see how well different AI models handle the complexities of online shopping. The card network put 22 different models from 12 different labs through 15 specific shopping challenges to see if they could navigate consumer needs correctly.

The test required the AI to identify every single requirement a customer might have before making a final decision to buy or not buy. Visa stated that if nothing fits the customer's requirements, the AI should opt not to buy. However, the results showed that the models struggled significantly with these nuances.

According to the data, the models failed 42.8% of the test runs. Even more concerning for some observers is that in 54% of those failed runs, the AI still proceeded to make a purchase. This suggests that the models sometimes prioritize completing a transaction over accurately following specific user instructions.

Read-only access for finance

While AI models are becoming more integrated with personal finance, the level of control they have over actual money remains limited. Anthropic recently launched a Personal Finance feature for Claude, which is currently in beta for certain paid plans in the United States.

This feature connects to users' bank and card accounts through Plaid, a common data aggregator. By using this middleman, the system ensures that Claude does not ask for or see the user's actual bank password. This setup allows the AI to pull in information without having the power to execute transactions directly.

Anthropic has been explicit about these boundaries in its launch communications. The company stated that "Claude can’t move, manage, or invest your money." This read-only status means the AI can serve as a sophisticated analyst of your spending habits without being able to act as a primary spender.

Retailers guard their storefronts

Major retailers are also taking steps to ensure that their own platforms remain the primary point of sale. Amazon recently integrated its own shopping agent into text messages, a move that follows its decision to block Meta’s Muse from querying its website in September.

Amazon's leadership emphasized that while they expect to work with horizontal agents, they believe their own Alexa for Shopping will be the primary tool for customers. The company reported that Alexa for Shopping was used by more than 400 million customers over the last 12 months, highlighting the scale of their existing ecosystem.

Similar logic is being applied by other giants like Walmart. The retailer's terms for agentic access require a human to approve every single checkout performed by an outside AI agent. The only exception to this manual approval rule is if the customer chooses to shop through Walmart's own internal agent, known as Sparky.

The gatekeeper architecture

When looking at these three developments together, a clear pattern emerges regarding who controls the

Mapping the agentic harness war

These developments align with the Business Engineer framework known as the Agentic Harness War. This concept suggests that the surface layer of an AI system, rather than the underlying model itself, is what ultimately governs its level of autonomy. In this context, the checkpoints established by the card network, the bank-data links, and the retail storefronts represent the physical and digital boundaries that define what an agent can actually do.

The current landscape shows that while AI models are becoming more capable, they are being integrated into existing structures that prioritize safety and control. By keeping the "harness" in the hands of established firms, the industry ensures that the most sensitive actions—like moving money or completing a sale—remain within governed systems. This architecture defines the current stage of agentic commerce where the model serves as the brain, but the established infrastructure provides the hands.

Context and data origins

This analysis of the current state of AI shopping agents was featured in this week’s issue of the newsletter, which explores the convergence of capital, power, and geopolitics in the AI supercycle. The framework used to evaluate these strategies is read by over 95,000 executives who look to Business Engineer for the AI strategy frameworks cited by major platforms like ChatGPT, Claude, and Perplexity.

It is important to note that this specific mapping of the "checkpoints" is an original classification based on the news of the week. While the data points—such as the failure rates and user counts—come directly from the companies themselves, the observation that these three specific areas are currently held by established firms is a unique analysis. This perspective describes where the current boundaries sit today rather than providing a definitive finding that any specific firm is actively blocking competition.