AI agents are no longer just automating tasks: they are starting to purchase, compare, and negotiate on behalf of their users. In this op-ed, the author argues the decisive question will not be whether these agents are capable, but how far they are allowed to decide.
Take the example of a consumer who asks their agent to ‘find a reliable washing machine for under 700 euros, delivered this week, at the best price.’ The agent selects a model at 679 euros. On the other side, the retailer’s agent knows the stock, margin, and commercial rules, and can offer up to an 8% discount or free delivery. The problem then becomes economic: how far can we let a machine arbitrate between conversion, margin, and customer retention?
Agentic Commerce Leaves the Lab
This scenario may still seem speculative, but the infrastructure is already being put in place. In the first quarter of 2026, Shopify reported that traffic from The AI / news to its stores had increased eightfold year-on-year. Visa states it has already processed transactions initiated by agents and is working, like Mastercard or Ant International, on mechanisms to identify and authenticate them. According to a Visa study from spring 2026, 53% of surveyed executives would be willing to let their agents negotiate directly with those of other companies.
Technology thus makes negotiation possible. The remaining task is to provide a framework for it.
A Mandate to Frame
The author distinguishes three stages. Before negotiation, it must be clear who has the right to act on behalf of the customer or the company. During the exchange, each proposal must be traceable; otherwise, it becomes impossible to understand why an agent granted 4% rather than 8%. Once the transaction is concluded, conditions under which a human can take back control must be defined. These safeguards do not eliminate the agent’s autonomy; they make it controllable.
To illustrate this point, the author recalls his experience at Nespresso, where he oversaw operational excellence in customer relationship centers. Advisors did not mechanically apply a compensation grid: they had limits, escalation rules, and above all a monthly envelope within which they decided freely. In his view, this autonomy led to better mastery of commercial actions than a rigid grid, because the advisor reasoned across all their decisions. However, the analogy has its limits: an agent does not have the relational intuition of an advisor, and its latitude must be translated into explicit objectives and rules. The principle of the envelope remains valid, measured over time through average margin, retention, churn (customer attrition rate), or conversion rate.
A Slider Rather Than a Button
Technology can move faster than trust. According to Visa, 58% of US consumers are comfortable with the idea of an AI comparing prices, but only 38% would accept it making the purchase directly, and 27% would accept it spending autonomously. Autonomy will likely take the form of a slider: ‘Compare for me,’ then ‘Negotiate up to this limit,’ then ‘Purchase if the price drops below 620 euros.’
In the washing machine example, the customer’s agent can negotiate up to 650 euros before requesting confirmation. The retailer’s agent offers a partial discount and free delivery rather than the maximum. An agreement emerges at 652 euros, an amount exceeding the threshold set by the customer: their agent therefore requests confirmation, as the mandate stipulated. The agreement is then timestamped and traced.
