Before Giving AI Agents Authority, Decide Who Handles Their Mistakes

Imagine this scenario. An online retailer’s AI agent inadvertently sends an order to a customer’s old address. By the time the mistake is discovered, the parcel is already on its way. The operations team is trying to intercept it, customer service wants to send a replacement, and finance wants to know whether the first parcel can be recovered before approving another dispatch.

In this case, the AI agent is a software allowed to make decisions and trigger actions in company systems, rather than simply making recommendations to a person. Before granting an agent that authority, a company should appoint someone who can settle such disagreements and approve the remedy. That person needs the power to cover the consequences of the agent’s actions, including decisions that remain outstanding after IT has corrected the records.

When permitted actions go wrong

Across industries, protocols and standards vary but, the fundamentals remain the same where accountability is placed on a human. The Monetary Authority of Singapore (MAS) and industry partners have examined what happens when AI agents are allowed to take actions that affect customers, money or operations in their July white paper, Safeguards for Agentic Finance at Runtime. Developed for financial institutions, it proposes checks before actions are executed, supported by records for review, accountability and remediation. It is an industry reference approach rather than regulatory guidance.

A retailer faces different obligations, but the underlying issue is similar: once a company allows an AI agent to make changes or trigger instructions on its own, someone must be able to resolve the problems those actions create. An outdated address or incomplete inventory record can produce a wrong decision within the agent’s permissions, leaving the company to trace what followed and decide how to respond. If it applies the same mistaken assumption to several orders, the company may have multiple deliveries to trace before discovering the first error.

Employees can also send parcels to the wrong address, and conventional automation can spread errors quickly. The MAS paper describes agents as entities that can plan intermediate steps, select tools and initiate actions without continuous human direction, sometimes without human review of each action. This combination matters because a mistaken decision can determine several subsequent steps before someone intervenes.

Restoring records without erasing valid changes

Recovery then becomes more complicated than restoring an earlier version of the customer database. During the same period, other customers may have legitimately changed their addresses. Simply restoring the database to an earlier point in time could remove those valid updates alongside the agent’s errors, creating a second set of incorrect records.

The company must distinguish the agent’s changes from legitimate activity and establish which warehouse instructions have already been executed. Without that visibility, organizations may know an agent made a mistake but still struggle to identify its full operational impact. Full visibility also help the incident owner give operations and customer service a reliable account of which orders require action.

An address can be corrected within minutes, but the customer’s remedy may still be stuck in an internal disagreement. When nobody has authority to settle it, the consequences of the agent’s mistake last longer than they need to.

The expansion of business AI adoption provides context for examining these arrangements. The Infocomm Media Development Authority (IMDA) recently released the Singapore Digital Economy Report 2026 which shows AI adoption among all enterprises rising from 14.7 per cent in 2024 to 23.5 per cent in 2025. Separately, its 2026 deployment findings show that two in five AI-adopting firms were expanding deployment or had fully deployed or integrated AI across the company. These figures cover AI generally and do not establish how widely agents are used.

Who has the authority to decide?

For firms delegating consequential actions to agents, existing incident management may provide much of what is needed. Many companies already have managers who coordinate technical and operational responses. Those managers can authorize remedies, settle competing priorities and retain responsibility through resolution.

The difficulty arises when the manager’s authority ends before the consequences are resolved. In the retailer scenario, an incident manager might be able to suspend the agent and arrange correction of the records, while replacement orders above a spending threshold require another executive’s approval. Therefore, the incident owner needs agreed decision-making and spending limits, with a named approver for matters beyond them, so someone can choose a course of action before all the facts are known.

Legal and compliance specialists should continue to assess obligations, while the owner coordinates their work with operational and customer decisions. Leading this response should be defined separately from investigating responsibility for the original failure, so the person arranging recovery does not inherit blame for deployment decisions made elsewhere.

A tabletop exercise can expose the gaps by requiring a decision under uncertainty. The exercise should establish who can approve the expenditure, how disagreement is escalated and which unresolved issues would prevent closure.

For companies extending agents’ responsibilities across systems, these decisions form part of the authority they are delegating. They should also understand how they would investigate, contain and recover from an authorized action that produces the wrong outcome. An address can be corrected within minutes, but the customer’s remedy may still be stuck in an internal disagreement. When nobody has authority to settle it, the consequences of the agent’s mistake last longer than they need to.

The views and opinions expressed in this article are those of the author and do not necessarily reflect those of CDOTrends. Image credit: iStockphoto/Andrii Dodonov