AI agents won’t just automate tasks. They are starting to buy, compare and negotiate on behalf of their users. Tomorrow, the customer’s agent will look for the best price; the company’s agent will have to convert without eroding margin. The real question will no longer be what AI can do, but what we are willing to let it decide.
Take a simple example. A consumer asks their agent: “Find me a reliable washing machine under €700, delivered this week, at the best price.” The agent shortlists a model at €679. On the other side, the retailer’s agent knows the stock, the margin and the commercial rules: it can grant up to an 8% discount or offer free delivery.
What should it do? The problem is no longer only technological. It becomes economic: how far do we allow a machine to arbitrate between conversion, margin and loyalty?
Agentic commerce is leaving the lab
This scenario still sounds forward-looking, but the infrastructure is already being put in place. In Q1 2026, Shopify reported that AI-driven traffic to its stores had grown eightfold year on year. Visa says it has already processed agent-initiated transactions and, like Mastercard and Ant International, is working on mechanisms to identify and authenticate them. According to a Visa study from spring 2026, 53% of executives surveyed would be ready to let their agents negotiate directly with other companies’ agents.
Technology makes negotiation possible. What remains is to give it a framework.
Securing a negotiation between agents
Three moments must be distinguished. Before the negotiation: knowing who has the right to act on behalf of the customer or the company. During the exchange: keeping a record of every offer — otherwise it becomes impossible to understand why an agent granted 4% rather than 8%. Finally, once the transaction is closed: setting the conditions under which a human can still take back control.
These guardrails don’t remove the agent’s autonomy. They make it controllable.
Nespresso: what you delegate is the envelope
At Nespresso, where I led operational excellence for the customer contact centers, advisors did not mechanically apply a grid specifying which compensation to grant in each situation. We had defined limits, escalation rules and, above all, a monthly envelope within which they kept real freedom of decision.
The result was counter-intuitive: this autonomy gave better control over commercial gestures than a rigid grid, because advisors reasoned across all their decisions rather than interaction by interaction. The envelope created the constraint; within it, judgment created the performance.
The analogy has its limits with AI: an agent lacks an advisor’s relational intuition, so its latitude must be translated into explicit objectives and rules. But the envelope principle still holds: rather than assessing each decision in isolation, you measure average margin, retention, churn or conversion rate over time, then adjust the mandate accordingly.
Customer autonomy won’t be binary
Technology can also move faster than trust. According to Visa, 58% of US consumers are comfortable with AI comparing prices. But only 38% would let it complete the purchase directly, and 27% would let it spend autonomously.
This sketches the likely shape of that autonomy: not an on/off switch, but a slider.
“Compare for me.” Then: “Negotiate up to this limit.” Then: “Buy if the price drops below €620.”
Companies will have to build the same gradation on the agent side. A delivery date can be changed automatically; a discount directly commits margin; a sensitive dispute warrants human validation.
There is also a subtler risk: that an effective negotiation fails to pass its value back to the customer. An agent that simply announces “I got this product for €629” withholds part of the picture: what did the company actually concede? Was the initial price €679? Was delivery free? Was an extra warranty negotiated?
The mandate must therefore cover not only what the agent may negotiate, but also how the outcome is reported.
How do two agents shake hands?
At what point does a negotiation between two agents actually become an agreement?
Offers and counter-offers can remain flexible and iterative. No standard has emerged yet — better to admit it. But the moment both parties commit must be unambiguously identifiable: timestamped, auditable, and reversible only within the window provided for that purpose.
Flexibility can be tolerated during the negotiation. Not at the moment of commitment.
Back to the washing machine
The customer’s agent may negotiate up to €650 before asking for confirmation. The retailer’s agent, for its part, reasons on its monthly envelope rather than on this single sale: it knows its stock level, available margin and conversion targets. So it offers a partial discount and free delivery rather than the maximum.
The deal takes shape at €652. Since this exceeds the threshold set by the customer, their agent asks for confirmation — exactly as the mandate required. Once approved, the agreement is timestamped and logged.
None of this is really science fiction. It is a question of mandate, not technology.
Who will design the mandate?
If agents negotiate thousands of transactions tomorrow, someone will still have to decide the strategy they apply: how far to defend margin, when to favor loyalty, which situations deserve an exception.
New roles will emerge at the intersection of sales, pricing, customer relations and data. Their mission will be less to negotiate each transaction than to design, test and adjust the rules by which agents negotiate.
A deeply human territory will also remain: the out-of-bounds — situations the rules did not anticipate, where someone will still have to decide.
The sales profession could thus gradually shift from negotiating transactions to designing negotiation strategies.
The real battle will be over the mandate
The question will soon no longer be “can agents negotiate?” — they will. It will be: what have we given them the right to negotiate?
For the customer: what budget, what autonomy, what confirmations. For the company: what minimum margin, what level of discount, when to take back control.
Agentic commerce won’t just be a new interface. It will be a new form of delegation — and its performance will depend less on the power of the one we delegate to than on the quality of the mandate we give it.