Francesca Roche sits down with Simon Langevin, VP of Product at Coveo, to examine the trust challenge emerging as AI takes a larger role in product discovery and purchasing decisions.
AI shopping assistants promise faster answers and more personalized recommendations, but a confident response can quickly become a customer experience problem when the information behind it is incomplete, outdated, or contradictory. Product origin, availability, pricing, compatibility, and entitlement data all carry consequences when surfaced incorrectly.
For consumer brands, a flawed answer can damage the credibility that differentiates them from larger marketplaces.
“What makes you different than these big retailers is really your brand,” Langevin says.
“People trust that you know what you’re selling.”
The risk becomes more acute in B2B purchasing, where buyers may work with complex catalogs, negotiated price lists, and specific purchasing permissions.
An AI assistant that recommends an unavailable product or displays a default price instead of a customer’s agreed rate can create costly returns and difficult follow-ups.
Langevin says the issue extends beyond the behavior of the large language model itself. “It’s both,” he says, referring to hallucinations and poor data governance.
AI models may produce more refined answers through multi-step reasoning, but their outputs remain dependent on the information they can access. If a system draws on inaccurate or stale data, even a well-grounded answer can be wrong.
That raises a practical question for CX and commerce leaders: where should reliability begin?
Langevin points to retrieval, or the way an AI system finds and ranks the information it uses. Pulling data individually from commerce platforms, enterprise resource planning systems, and content management systems can add cost and create more opportunities for error.
He noted:
“Having one system where I, as an agent, I query, I retrieve data, I get a proper ranking as well of that data, depending on what I’m asking, will reduce drastically the potential errors.”
The interview also explores the governance choices behind trustworthy AI. In sectors with detailed product hierarchies and local requirements, business rules may need to determine whether a local source, a generic product listing, or another record takes priority. Those decisions affect whether an assistant can give a useful response with confidence.
When the available evidence is weak, an assistant should know when to stop. Langevin identifies low-confidence retrieval as an early warning sign that a customer should be handed to a human agent.
The stakes are rising because, as Langevin puts it, “the trust is broken.”
Customers increasingly want to understand not only what an AI recommends, but why it reached that conclusion and which sources support it.
Watch the full interview for a closer look at the controls that can help commerce teams make AI assistance more accurate, transparent, and dependable.
