Merchants who let AI agents act autonomously should first assign responsibility for mistakes or extreme risks.
“Once an agent can update a product catalog, issue a refund, or change a campaign, it becomes operational software with delegated business authority. The company needs to treat it that way,” wrote David Greenberg, chief marketing officer at BlueRock Security, a cybersecurity company.
Businesses, including ecommerce shops, are evolving from experimenting with AI to deploying agents in production or customer-facing applications.
Delegated Authority
Consider an ecommerce merchant that instructs an AI agent to improve conversion rates for underperforming products.
The agent might examine inventory, identify slow-selling items, reduce prices, and shift advertising spending toward those products. Each action could be reasonable, but the combined result might undermine margins, exhaust inventory, or create unexpected costs.
The question then is how much authority the merchant should delegate.
David Greenberg
The Massachusetts Institute of Technology’s Center for Information Systems Research identified two factors — ambiguity and risk — to determine which decisions AI should make.
Routine tasks with predictable outcomes and limited consequences are good candidates for automation. But matters involving uncertainty, significant financial exposure, or competing objectives warrant greater human involvement.
Greenberg, in an email interview with Practical Ecommerce, suggested additional considerations, including monetary value, reversibility, customer impact, data sensitivity, and cumulative exposure.
“A small refund repeated thousands of times can become a significant problem,” Greenberg wrote.
Merchants should establish guardrails that define what an agent can access, change, or spend. Those restrictions are enforceable as the agent works, not through instructions alone.
“If your visibility ends at the prompt, keep authority narrow,” Greenberg wrote.
Accountability
Delegating a decision to an AI agent does not shift responsibility for its consequences.
A department “accountable for the process should remain accountable when an agent performs it,” Greenberg stated. “Marketing owns campaign outcomes. Customer service owns its customer-resolution process. Procurement owns purchasing decisions.”
That principle applies regardless of the agent.
“A shopper experiences the company’s promise, price, or refund. An internal explanation about which AI vendor generated it will not repair a poor experience,” noted Greenberg, adding that every operating agent should have an “owner,” someone responsible for configuration, access, monitoring, and incident response. In smaller companies, those responsibilities might belong to the same person.
Merchants should also be able to reconstruct what happened after an agent acts.
For a refund, that means identifying the:
- original request,
- order,
- applicable policy,
- proposed amount,
- authorization,
- payment transaction,
- and customer communication.
These records make it possible to distinguish between a bad decision, an execution error, and a customer error. Ultimately, adjusting to what went wrong should improve performance henceforth.
Commercial Judgment
Perhaps the top risk is an agent that performs exactly as instructed but produces a poor outcome.
Greenberg offered a hypothetical example of a merchant authorizing an advertising agent to maximize revenue within a fixed weekly budget.
The agent identifies a product with a relatively high conversion rate and reallocates spending toward it. Sales increase. The agent stays within budget and advertises only approved products. The AI automation appears successful.
“But the product has thin margins, expensive fulfillment, and a high return rate. The agent has optimized the metric it was given while reducing the business’s profitability,” wrote Greenberg, adding, “This is where marketing judgment becomes more valuable. Teams need to translate business strategy into measurable objectives and enforceable constraints.”
Connected Agents
The complexity increases when several AI agents operate within the same business.
Imagine a merchandising agent reducing prices, an advertising agent increasing spending, and a purchasing agent replenishing inventory.
“Each can behave reasonably within its own assignment while their combined actions create an unwanted feedback loop,” Greenberg wrote.
In a September 2026 Harvard Business Review article, researchers described how companies, including Walmart, are developing systems to coordinate AI agents across functions.
The researchers found that specialized agents can produce useful recommendations independently while generating conflicting outcomes across departments.
Coordinating those agents could automate entire processes, but it requires shared rules, management oversight, and a way to reconcile competing objectives.
Operating Agents
Increasingly, AI lets owners, managers, and contributors build custom tools and automate processes that once required developers, software, or repetitive manual work.
But building an agent is different from operating one.
The objective is not to require human approval for every routine action. It is to establish enough control to delegate useful work without surrendering responsibility.
The question for ecommerce companies is not what AI agents can do. It is what those agents are allowed to do, under what conditions, and who remains accountable for the results.
