Five questions for Jeannot Poisson, AI Transformation Principal at iubenda, on delegation, data, and what brands owe the customer behind the agent.
Takeaways
- Delegation will be gradual, and proportional to the stakes.
- When a purchase goes wrong, customers need one clear route to fix it.
- Personalization only earns trust if people can see and change what the agent remembers.
- Permission should be scoped to the task, time-limited, and easy to withdraw.
- When the agent handles discovery, brands must account for the fact that they have very few touchpoints and enter only at the order stage.
AI agents are starting to browse, compare, and buy on people’s behalf. The protocols that make it work are taking shape, and regulators have started to describe what they expect. What will become clearer in a few months and years is adoption. Will people actually hand an agent their money? If yes, what should a business owe them when they do?
PwC put the problem plainly in April 2026: people are increasingly using AI to find products and services, but many still aren’t comfortable letting an AI agent complete the purchase. They call that trust gap the next barrier to scale, and frame closing it as a growth move rather than a defensive one. Brands that get there first, PwC writes, “won’t just avoid risk. They’ll build the kind of loyalty that translates directly into long-term growth and competitive advantage.”
This time, we wanted a point of view from inside the technology, so we put five questions to Jeannot Poisson, who spends his days helping people across the company put AI to work. His answers are below, in his own words.
Is agentic commerce becoming a real channel?
We asked about timing, and about why handing over a purchase feels like a bigger step than saving a card and checking out in one click.
Jeannot: Yes, I think it will become a real channel, but people will hand over control gradually. From helping people use AI in their daily work, I’ve seen that interest comes quickly, while changing habits takes longer. People need to experience a clear benefit before they start relying on it.
I’d expect familiar, repeat purchases to gain traction by 2027. Giving an agent freedom to choose something expensive or personal feels further away. The willingness to delegate depends on how much the choice matters to you.
The jump from one-click checkout is significant because you’re delegating judgment. People will want to see the agent get things right, understand its limits, and know they can get help when it makes a mistake.
When an agent gets a purchase wrong, who answers for it?
We listed the failures we expected: an agent fooled by text planted on a product page, misreading a return policy, ordering the same thing twice.
Jeannot: From working with people using AI, I’ve learned that a confident answer can create more trust than it deserves. In shopping, an agent could sound certain while misunderstanding what someone wanted or overlooking an important condition.
My view is that the customer should have a clear route to put things right. They shouldn’t be passed between the merchant and the AI provider, with each saying it was the other’s fault. The exact responsibility will depend on what happened, but the process for getting help should be clear from the start.
It should also be possible to understand what was requested and what was actually agreed. People are more likely to try something new when they know a mistake won’t become their problem to untangle alone.
Agents get better the more they remember. Privacy works the other way. Is that a conflict?
Jeannot: There’s a tension, because remembering context is one of the things that makes an assistant useful. Having to explain your preferences every time quickly becomes frustrating. I see the value of continuity in how people use AI at work.
But sharing something once doesn’t necessarily mean you want it remembered forever or used for another purpose. That’s where expectations can diverge.
I’d want the agent to remember useful preferences while giving me a straightforward way to see, change, or remove them. Personalization should feel like a service I control. If the assistant seems to know things I don’t remember sharing, that can undermine trust very quickly.
When an agent buys on a brand’s site, who ends up holding the customer’s data?
We also asked what a brand loses in visibility compared with a normal checkout, and whether that makes future marketing harder.
Jeannot: Depending on how the service works, both the merchant and the agent platform may hold information. What interests me most is who ends up holding the relationship with the customer.
If I explain what I want to my assistant, compare options there, and ask it to buy, the brand may only enter the picture at the order stage. It has fewer opportunities to understand my needs or explain why I should choose their product or service.
That could make marketing harder, but it also puts more weight on the actual experience: accurate information, reliable delivery, useful support, and a product that meets expectations. I’d expect those things to matter even more when an assistant helps decide whether I buy from that brand again.
What does it take to trust an agent with your money?
Jeannot: In my experience, people become comfortable with AI when they understand what it will do and see it helping them consistently. For purchases, I’d want clear boundaries: what it can choose, how much it can spend, and when it needs to come back to me.
The amount of reassurance should match the purchase. Reordering an everyday item within an agreed budget could be simple. Booking an expensive trip should involve much more input.
I also wouldn’t assume that accepting a permission screen means someone understands how their information will be used. The explanation needs to connect to the action: what will be shared, with whom, and why.
Ultimately, I’d want to feel that I’m still making the important choices, even when the agent handles the work.
What we take from this interview
Read across the five answers and the same idea keeps surfacing: people will delegate the work, not the decision. They’ll hand over the searching and the ordering, as long as they can still see what it knows about them, set what it’s allowed to do, and reach a human when it gets something wrong.
PwC’s advice to brands lands in the same place: build platforms that let people grant, manage, and revoke the permissions an agent acts on, and explain plainly how those agents work. As they put it, “This isn’t just about compliance. It’s about defining and enforcing the boundaries of what an AI agent is allowed to do on a customer’s behalf.”
The emerging practice could be: treat an agent as a third-party application that needs explicit permission, give it only the access the task requires, keep that permission time-limited rather than standing, and let people withdraw it at any point.
That’s as much a design question as a legal one, and it’s where we’ve spent the last decade: making data use legible and giving people real control over it. Brands that get it right will be the ones people let their agents buy from a second time.