What does winning a sale mean when the shopper isn’t a person at all, but a piece of software buying on their behalf?
For most of the internet’s history, that question didn’t need asking. You won by persuading a human — a sharper headline, a cleaner landing page, one less step at checkout. That assumption is starting to wobble. As AI agents take on discovery, comparison, and sometimes the purchase itself, the thing a brand has to convince may not have a pulse.
Aniket Deosthali has spent more time than most sitting with that shift. He’s the CEO and co-founder of Envive, a Seattle company building AI agents for retail brands, and before that he was Head of Product for Conversational Commerce at Walmart, where he built the retailer’s LLM- and generative-AI-powered shopping experiences from zero to scale. Envive now works with brands including Spanx, Supergoop!, Coterie, and Wine Enthusiast. His read on where this is heading is blunt.
“Winning used to mean persuading a person to click,” Deosthali says. “Increasingly, it will mean earning the confidence of a machine acting on that person’s behalf.”
That single change moves the unit of competition. An agent won’t page through ten landing pages, and it won’t fall for the cleverest bit of copy. It’s trying to work out which product best satisfies a customer’s intent, weighing things like relevance, price, availability, delivery, trust, and what that shopper has preferred before. So the job of marketing widens. “Your product data, customer signals, inventory and service policies all become part of your marketing,” he says — which is a strange sentence to read if you’ve spent a career thinking of marketing as the stuff a customer sees.
Persuasion doesn’t die in this world, though. It relocates. “The brands that win will combine machine-readable truth with a differentiated experience,” he says. An agent may narrow the field to a handful of options, but there’s still room to earn preference once a customer is in front of the product. “The rewrite is significant, but it is not the death of marketing. It is an expansion of what marketing must influence.”
Getting picked by a system you can’t see
If an agent is doing the choosing, the obvious next question is what actually drives that choice. Deosthali is careful not to pretend there’s a single formula — different agents run on different models, data, and commercial deals. But underneath, he says, an agent is really asking three things: “Can I understand this product? Can I trust the information? Is it a strong fit for what this customer wants right now?”
Answer those badly and you disappear, which is why he treats product data as foundational rather than back-office plumbing. Reviews, return policies, price, inventory accuracy, fulfillment history — all of it feeds an agent’s confidence, and much of it lives in the machine-readable signals most marketing teams have never had to think about. Context matters just as much. Someone asking for “the best running shoes” might mean the fastest, the most supportive, the most sustainable, or just the cheapest. “The winning product is not necessarily the one with the most keywords,” Deosthali says. “It is the one whose attributes can be matched most clearly to the shopper’s actual intent.”
His analogy for the stakes is a new kind of shelf visibility. In a physical store, a great product can fail because it’s sitting where nobody looks. “In agentic commerce, you can have a great product that effectively does not exist because machines cannot interpret or validate it.”
Does the agent flatten the brand?
The fear a lot of brand leaders carry is that agents grind everything down to price and spec, commoditizing the identity and loyalty a brand spent years building. Deosthali thinks that fear is real, but conditional. “The fear is warranted if brands allow someone else’s agent to become the only interface between them and their customers,” he says. In that setup, sure, products get reduced to comparable attributes.
But he flips the usual worry. “Agents do not eliminate brand. They make a strong brand even more valuable because brand itself is a trust signal.” Reputation, product quality, consistency — those help an agent judge whether a recommendation will land well. And loyalty becomes a direct input, not a soft one. Ask an agent to “find me a jacket like the one I bought from this brand last year,” and the brand is suddenly doing real work inside the decision.
His time at Walmart sharpened this for him. The best AI shopping experiences, he found, do more than spit out an efficient answer. “They understand context and reduce uncertainty.” A beauty brand can help someone build a regimen. An outdoor retailer can figure out which gear suits a specific trip. Those are hard to commoditize precisely because they run on knowledge and relationship, not just a spec sheet.
Who should own the agent
That leads to the question Deosthali seems most animated about: when the interaction starts somewhere like ChatGPT, who owns the shopping agent — the platform or the brand? His answer leans hard in one direction. Customers should be able to start wherever’s convenient, he says, and brands need to show up inside those third-party surfaces the way they once had to show up in search and social. What they shouldn’t do is hand over the whole relationship.
A platform’s agent, after all, is built around the platform’s goals and its view of the customer. Cede control of the checkout and the customer record entirely, and a brand can lose sight of why a product got recommended, what the shopper wanted, and the chance to learn from any of it. “Let platforms help customers find you,” he says, “but make sure you retain the ability to serve, understand and build a lasting relationship with them.”
So what should a CMO who knows this is coming actually do first? Don’t start with “we need an agent,” Deosthali warns. Start with a high-friction journey where customers genuinely need help — choosing between complex products, replenishing routine buys, finding the right option on a budget. Then fix the foundation: audit whether your catalog, inventory, policies, reviews, and first-party data are clean enough for an agent to trust. Measure the things that matter — conversion, basket size, repeat purchase — not how many conversations a bot handled.
And the waste of time? “Launching a generic chatbot, calling it an agent and measuring success through engagement.” A conversational layer that can’t read intent or take real action, he says, is just one more thing standing between the customer and the product.
Which is maybe the cleanest way to state the whole shift. “The goal isn’t to make commerce more conversational,” Deosthali says. “It’s to make commerce more effective.”
