1 in 5 Times, AI Goes Rogue. Yet, ‘Automated Purchasing’ Continues

AI agents select products and complete transactions on behalf of humans. This is not science fiction; it is already happening in 2025.

At the same time, banks in the United States have begun blocking access to customer accounts by AI agents. The reason is simple: they go rogue. According to research data, when AI agents autonomously engage in purchasing behavior, they act contrary to the user’s intent about 20% of the time — that is, 1 in 5 times.

1 in 5 times. This is not just a level of “occasional mistakes.” If your store receives 100 orders a month, that means 20 of them could lead to trouble.

So, should small and medium-sized e-commerce businesses ignore this trend? Unfortunately, that is not an option.

What is Happening — The Background of Banks Saying ‘No to AI’

In the U.S., major banks like JPMorgan Chase and Wells Fargo have started detecting and blocking account access by AI agents.

The underlying issue is the “excessive autonomy” of AI agents. A user might instruct, “Please buy a coffee maker.” The agent then browses Amazon, reads reviews, adds what it deems the optimal product to the cart, and completes the purchase. Up to this point, everything is fine.

The problem arises when the agent also signs up for a subscription for premium coffee beans and even contracts for a filter subscription, resulting in a total cost that is three times the original amount. What was intended to be 3,000 yen ends up being 9,000 yen. Moreover, canceling subscription contracts can be cumbersome.

From the banks’ perspective, this is a reasonable decision. Customers report, “There’s a charge I don’t recognize.” Upon investigation, it turns out the AI agent did it autonomously. Does this qualify for chargebacks? It cannot be processed under the existing frameworks for fraudulent use. Hence, they stop it at the entry point.

There is a structural problem here. It is legally and systemically unclear whether AI agents are considered ‘the customer’ or ‘a third party.’

Large Corporations Are Already Operating Under the Assumption That ‘AI Agents Are Customers’

Spotify is revamping its API design in anticipation of an era where AI agents select music. When an agent requests, “Create a playlist that fits this user’s mood,” a personalized list of songs is returned instantly.

Uber has introduced ChatGPT Work to all its employees. Not only for internal operations but also envisioning a future where AI agents arrange rides.

OpenAI’s “Operator,” Google’s Project Mariner, and Anthropic’s “Computer Use” — tech giants are all investing in the development of AI agents. What they are witnessing is the “end of the era where humans operate browsers.”

According to Gartner’s predictions, by 2028, at least 15% of online purchases will be made through AI agents. Considering the scale of e-commerce sales, this could mean transactions worth trillions of yen in Japan alone will be conducted via AI agents.

This trend is irreversible. The question is not “whether it will come” but rather “when it will come” and “whether we are prepared when it does.”

The Real Issue for Small and Medium-Sized E-Commerce — The Rules of ‘Being Chosen’ Will Change

Now to the main point. What does the arrival of the AI agent era mean for small and medium-sized e-commerce businesses?

In short, the rules of competition will fundamentally change from ‘being chosen by humans’ to ‘being chosen by AI.’

Currently, e-commerce involves humans searching through browsers, viewing product pages, reading reviews, comparing, and making purchases. Therefore, design, the appeal of photos, and catchy slogans have been crucial.

AI agents do not consider such factors. They look at structured data. Price, specifications, stock status, shipping costs, return conditions, and numerical data from reviews. Whether these are organized in a machine-readable format will determine whether they are chosen or not.

This presents an opportunity for small businesses. Why?

The brand image and advertising that large corporations spend millions on are powerless against AI agents.

AI agents do not think, “I’ve seen this brand in a commercial.” They compare purely based on data. If the price is reasonable, specifications are clear, and return conditions are favorable, even an unknown small e-commerce business can be chosen.

Conversely, if product data is messy, there are no APIs, and price information is buried in PDFs, it won’t even be considered by AI agents. It’s as if it doesn’t exist.

Three Things Small and Medium-Sized E-Commerce Should Do Now

So, what should be done specifically? Large investments are not necessary. Start with the following three actions.

1. Make Product Data ‘AI-Readable’

This is the top priority. Product names, prices, specifications, stock numbers, shipping costs, return conditions — these should be described in structured data (JSON-LD or Schema.org).

In terms of cost, if you are using Shopify, there are free plugins available for this. Even for self-operated e-commerce, if you ask an engineer, it can be done for a few hundred yen per product, or a few tens of thousands of yen for 100 products. If outsourced, it would be around 100,000 to 300,000 yen.

By doing this, you will enter the ‘searchable target’ of AI agents. If you don’t do it, you won’t be included. It’s as simple as that.

2. Establish ‘Guardrails for AI Agents’ in Payment Processing

In a world with a 20% runaway rate, allowing payments to go through unchecked is dangerous. However, blocking all orders from AI agents would mean discarding future sales.

A realistic approach is to implement a tiered confirmation flow based on amount.

• Under 3,000 yen: Automatic approval
• 3,000 to 10,000 yen: Confirm with a push notification to the user
• Over 10,000 yen: Explicit user approval required

This system can be implemented simply by adding conditional branching to existing payment APIs. Development costs would be around 50,000 to 150,000 yen. Considering the costs of handling returns and chargebacks due to runaways (500 to 3,000 yen per case), the investment would quickly pay off.

3. Redesign the Return Flow with ‘Automation in Mind’

As purchases via AI agents increase, returns will also undoubtedly rise. While the 20% runaway rate does not directly translate to a return rate, it is certain that returns will increase compared to before.

The cost of processing returns is generally said to be 15-30% of the product price. For a 1,000 yen product, that’s 150-300 yen; for a 5,000 yen product, it’s 750-1,500 yen. If 100 returns occur in a month, that alone could lead to costs of 75,000 to 150,000 yen.

The key here is to automate the return flow rather than relying on manual labor. By systematizing the acceptance of return requests, refund processing, and restocking, the processing cost per case can be reduced to less than half. If you use Shopify, apps like “Return Prime” or “Loop Returns” can be used for a few thousand yen a month.

In fact, e-commerce platforms with easy return processes may be prioritized by AI agents as “safe purchasing destinations.” Enhancing return policies is an investment, not a cost.

‘E-Commerce Chosen by AI’ or ‘E-Commerce Ignored by AI’

To summarize.

The era where AI agents become the primary purchasers is coming. As the 20% runaway rate indicates, the technology is still immature, and there are risks significant enough for banks to take action.

However, that is precisely why now is the time to prepare.

Large corporations will invest millions in building infrastructure to accommodate AI agents. Small businesses do not need to mimic that. By structuring product data, establishing payment guardrails, and automating returns — with a total investment of around 200,000 to 500,000 yen, the foundation for ‘E-Commerce Chosen by AI’ can be established.

Conversely, if nothing is done, they won’t even appear in AI agents’ search results. Human customers will decrease, and customers via AI agents will not come. They will quietly disappear from the market.

I want to ask: Is your e-commerce site ‘existent’ in the eyes of AI agents?

Is your product data in a machine-readable format? Can your payment flow handle AI runaways? Is the return process automated?

If the answer to all is ‘No,’ I urge you to start with just one thing this week: structuring your product data. Just installing one Shopify plugin will suffice. It will take about 30 minutes.

The winners in the era of AI agents will not be the companies that are technically savvy. They will be the companies that acted first.