On 8 November 2026, Practical Ecommerce published “Who Owns an AI Agent’s Mistakes?” by Armando Roggio. It argues that merchants should decide who is responsible for an agent’s mistakes before they let it act on its own. It leans on David Greenberg, chief marketing officer at BlueRock Security, and on MIT’s Center for Information Systems Research. The article also cites Harvard Business Review research from September 2026.
Greenberg’s line is that an agent able to change catalogues, issue refunds or alter campaigns becomes “operational software with delegated business authority.” The article says delegating a task does not hand over responsibility for the result. Customers judge the company’s promise, price or refund. They don’t care which AI vendor was involved.
The owner is already known
Much of the excitement around agents comes with a question attached: when the AI gets it wrong, whose fault is it? I don’t think that needs a debate. The business owns it. If your agent promises a refund, a price or a delivery date, the customer is dealing with you.
So the useful question comes earlier. What was the agent allowed to commit to when it ran? That is a decision you make before the first conversation, and it is the one the article keeps coming back to.
What is actually new here
Is there a real need? Yes. Businesses are giving software the power to act, and acting costs money in a way answering does not. The article’s example is a merchant who tells an agent to lift conversion on weak products. The agent cuts prices and moves ad spend. Each step looks reasonable, but together they can hurt margin or empty the stock room.
The article also notes that a small refund repeated thousands of times becomes a large problem. The agent never did anything dramatic. It just never had a limit.
What hasn’t changed is the old rule for any employee with authority. You don’t hand over a company card and hope for good judgement. You set a spending limit. The only new part is that the employee is software, and it can repeat a decision far faster than a person can.
Put the limits in the system
Greenberg’s advice is that guardrails should be enforced while the agent runs, not just written into its instructions. If oversight stops at the prompt, he says, keep the agent’s authority narrow. I agree with that, and I would go a step further. Rules belong in the system, not in the prompt.
A prompt is a request. An instruction like “never refund more than S$50” asks the AI to remember and obey. A limit in the platform means a refund above S$50 cannot go through, whatever the AI decided. That cap is a setting you can read, test and change.
Take a small online shop. Set the maximum refund per order and the maximum discount. List the delivery dates the agent may quote, and decide what it must hand to a person. Then limit how much it can refund in total per day. That last one matters because of the cumulative problem the article raises. A cap on each action does not cover ten thousand actions.
The same goes for a clinic. The agent can book and move appointments inside the hours you set. It cannot promise a treatment price or a doctor’s availability that your system doesn’t show. A service business should list the same four things: money, promises, dates and prices. The system sets a ceiling on each, and the AI works below it.
None of this blames the AI. A well-built agent resolves most of what comes in. The business’s rules decide what it can commit to, so a wrong judgement call lands inside limits you chose.
Where this shows up in what I build
This is how I think about AeroChat. It can refund up to the amount the merchant sets and hands anything bigger to the team, with the order details attached. The limit is a merchant setting. Nobody has to hope the AI remembers it.
What to do
Before you switch on any agent that can act, write down four things. First, name one person who owns it, as the article recommends. Second, list what it may commit to in money, promises and dates. Third, set each of those limits in the platform and add a daily total. Fourth, check that every action is recorded, so you can reconstruct a refund from the request to the payment and the message the customer received.
Then test the limits by asking the agent for something over each one. If it gets through, the rule was in the prompt.
Source: Practical Ecommerce: Who Owns an AI Agent’s Mistakes?