The system interface may label an output as a “recommendation,” but managers may still require every department to follow it. Employees who choose a different course of action may have to provide additional explanations or even worry that their performance evaluations will be affected. In such circumstances, it becomes difficult for anyone in the organization to treat an AI recommendation as merely a reference. Taiwanese companies may encounter the same issue when using AI Agent for sales quotations, procurement comparisons, quality-control decisions, or customer service.
Whether AI output is truly a recommendation or effectively a requirement depends on how decision-making authority is allocated in the actual workflow. Before discussing model accuracy, companies must clarify who can reject a recommendation, who is responsible for approval, and whether rejecting it will affect performance evaluations, approval speed, or subsequent operations. If employees do not have a reasonable opportunity to reject a recommendation, “human decision-making” may amount to nothing more than a button on the screen.
When Does an AI Recommendation Become a De Facto Requirement?
Business Next’s summary of a Reuters investigation reported that McDonald’s uses machine learning to analyze store transactions and make pricing recommendations for individual locations and menu items. Factors considered include local customers’ price sensitivity and willingness to pay, as well as competitors’ publicly available prices. The company stated that the tool provides recommendations and that franchisees retain control over final pricing. However, the report also noted that McDonald’s had incorporated the pricing tool into its operating standards and tracked instances in which franchisees deviated from recommended prices.
The central issue is the management system. When an organization continually tracks who has not followed a recommendation—and records of those deviations may affect performance evaluations, contract renewals, or management reviews—the resulting pressure can influence decisions even if the system continues to label its output as a “recommendation.” Conversely, recording the reasons for deviations can help assess the model and workflow. The key question is whether those records are used to improve the system or enforce compliance.
Companies must also clarify who benefits from the system’s definition of “optimal.” Headquarters may prioritize overall revenue and brand consistency, while individual stores or distribution channels directly bear costs such as labor, inventory, and rent. When the parties have different objectives and costs, the same recommendation may benefit one party while leaving another to bear the risk. Before deployment, the implementation team should identify everyone affected by a recommendation, the costs borne by each party, and the metrics used by the system. It should also assess whether deviating from a recommendation could lead to unreasonably adverse consequences.
Map the Decision-Making Process Before Establishing Human Approval Gates
Companies can begin by mapping the workflow from data input and AI recommendations through human review and actual execution. Each step should specify the responsible role, the available actions, and the conditions under which the process must stop. The workflow diagram should answer practical questions such as: Will an AI recommendation be automatically entered into a quotation? Can the person handling the case modify it? Who approves any modifications? If a manager does not respond, will the workflow wait, send the case back, or proceed with the recommendation?
The team should then establish human approval gates based on risk. For example, if a sales quotation exceeds the authorized range, a procurement recommendation involves a single supplier, a quality-control decision could result in the return of an entire batch, or a customer-service resolution involves compensation, AI Agent may first compile the supporting information and make a recommendation. However, the workflow must stop and wait for a person with the appropriate approval authority to decide. The system should not automatically treat “no response” as approval; otherwise, human approval can easily become a mere formality.
The design of the rejection process is equally important. Employees should be able to choose a different option and record their reasons, such as an inventory discrepancy at the site, separate contractual terms with the customer, or data that has not yet been updated. Managers should review the reasons and associated risks rather than focusing only on the number of deviations. The project team can begin by piloting a workflow with clearly defined risks and decision-makers, then test whether employees can reject recommendations, whether the process actually stops when required, and whether approval responsibility is assigned to the correct role.
Records Must Show Who Made the Decision and Why It Was Approved
Governance records should enable management, legal, and audit personnel to determine at least four things: what the AI recommended at the time, which data informed the recommendation, what final decision was made by a human, and who approved it and when. If a human chose a different option, the reason and subsequent outcome should also be retained. Only then can the data be used to evaluate recommendation quality, resolve internal disputes, and reconstruct the decision-making process when external questions arise.
Permissions should also be segregated. Access to transaction data, authority to modify recommendation rules, authority to approve outcomes, and audit responsibilities should not all be controlled by the same role. If a pricing system uses competitors’ publicly available prices and makes recommendations to multiple stores or distribution channels, the company should also have legal or compliance personnel review the data sources, recommendation methodology, and scope of use, as legal experts continue to hold differing views on the associated competition-law risks.
EgentWrX can connect tasks into workflows and establish human approval gates before handoffs, allowing a process to pause until a person authorizes it to proceed. The platform can also define who may execute, view, and manage tasks based on organizational hierarchy while retaining verifiable audit records. However, implementation teams must still define the approval conditions, role responsibilities, and purposes for which records may be used. The platform provides the mechanisms needed to enforce governance rules.
Before going live, the project owner should select a controlled test case and separately simulate accepting, modifying, and rejecting an AI recommendation. The team should confirm that all three paths can be completed and that rejecting a recommendation does not prevent subsequent work from proceeding. If the team cannot use the records to identify the decision-maker, the basis for approval, and the outcome of execution, it should first address the gaps in the workflow and permissions before expanding the scope of use.
FAQ
Does retaining a human approval button mean that a person is responsible for the decision?
No. The company must also ensure that the approver has sufficient information, authority, and time to exercise judgment and that rejecting a recommendation will not lead to unreasonably adverse consequences. If the button only allows the approver to accept the default answer, or if the system automatically adopts the recommendation after a timeout, human approval remains a mere formality.
Can a company record instances in which employees deviate from AI recommendations?
Yes, but it should first explain the purpose of the records, who may access them, and how they will be used. Deviation records are useful for identifying missing data, problems with rules, and exceptional scenarios. If they are to be incorporated into performance evaluations, managers must separately review the reasons for each deviation and avoid treating adoption rates as a measure of compliance.
Which AI recommendations should require mandatory human approval?
If an AI recommendation involves an amount exceeding the authorized limit, customer rights and interests, supplier selection, the return of an entire batch, or regulatory risk, the workflow should stop and wait for human approval. Companies can classify decisions according to the potential severity of losses and their reversibility, prioritizing decisions with major impacts that would be difficult to reverse.
What should a company do before using competitors’ publicly available prices as inputs for AI pricing?
Legal or compliance personnel should first review the data sources, recommendation methodology, intended recipients, and record-retention rules. If the system makes pricing recommendations to multiple distribution channels, the company should also clarify whether each party retains independent decision-making authority and preserve records of human decisions and the basis for approval.
References
- Are Burgers More Expensive If You Are Willing to Pay More? McDonald’s in the U.S. Reportedly Introduces “AI Pricing,” with Big Mac Prices Differing by 20% Just Three Kilometers Apart — Business Next’s summary of a Reuters investigation, company statements, and expert opinions, covering AI pricing, franchisee autonomy, and competition-law concerns.