Editor’s note: You can watch an on-demand replay of the webinar here.Â
Key takeaways:
- CI&T research shows 90% of U.S. consumers have used or are open to using AI agents in their shopping journey.
- Inventory timeliness, zone pricing, and semantic data gaps are the biggest barriers to effective AI deployment in retail.
- Chatbot conversion rates can be misleading due to self-selection bias; overall site conversion is a more reliable measure.
Retailers have spent decades believing they introduce technology before consumers are ready for it. Agentic commerce may be the first exception to that rule.
That is the view of Melissa Minkow, Global Director of Retail Strategy and Insights at CI&T, and Nikki Baird, VP of Strategy and Product at Aptos, who spoke during a Retail TouchPoints webinar Wednesday focused on where AI is actually delivering results in the shopping journey.
Minkow said in her research, 90% of U.S. consumers surveyed said they have either used AI agents in their shopping journey or are open to doing so.
“That is a level of adoption that is unprecedented when it comes to retail technology,” she said in the session.
The shift, she said, is tied to a fundamental change in how consumers approach the path to purchase. Where shoppers once moved from discovery to research to buying, they now begin with research, discovering brands along the way.
“Consumers are approaching the retail space with a very clear picture in their heads, or a level of specificity they’ve never had before,” Minkow said.
Retailers, however, are not yet meeting that moment. Minkow pointed to a distinction between the AI retailers have deployed for more than two decades, largely pattern recognition, and what agentic and generative AI now require.
“Retailers really need to step up the way they are leveraging AI to match the energy that consumers have,” she said.
The Real Problem Is Problem Solving
Baird said the instinct to optimize product descriptions for keyword matching misses the point. What AI systems are actually keying in on is intent, specifically the problems consumers are trying to solve.
“Ecommerce websites never got to the point where they addressed solving the consumer problem,” Baird said. “They assumed the consumer identified what they needed to solve that problem and was coming to their site to learn more.”
She offered a pair of examples to illustrate the range. A consumer asking an AI agent to build a weekly meal plan around specific macros and then identify which stores offer the lowest prices on the ingredients is a different exercise from someone searching for a shampoo formulated for color-treated hair.
Both are problem-solving exercises. Neither maps neatly to how most retailers currently structure their product data or content.
“If you can get into more of that headspace and cover more of that territory from a content and presence perspective, you’re going to have a better opportunity to show up,” Baird said.
Data Quality Is the Real Bottleneck
Both speakers pointed to data infrastructure as a decisive factor in whether AI deployments work in practice. Baird said that the timeliness and accuracy of inventory data becomes especially fraught when feeding information to AI agents.
A customer who learns from a chatbot that a product is in stock expects that to be true when they arrive at the store or proceed to checkout.
“You don’t want to promise to somebody who logged in from an AI chat that you have this available when it’s actually sitting in somebody’s cart and they’re about to walk up to the counter in the store,” she said.
Pricing data compounds the problem. Zone pricing, limited-time promotions with deadlines, and loyalty discounts all create inconsistencies that can mislead AI systems and erode consumer trust.
“All of those things will undermine the trust that consumers have, either in you or in the LLM,” Baird said.
Semantic gaps between retailer product attributes and the language consumers use when querying AI systems present a further challenge.
A shirt that “looks great on Zoom calls” does not correspond to any existing product attribute in most retail catalogs, but it is exactly the kind of query consumers are now posing.
Lowe’s, Personalization, and the Privacy Line
Lowe’s was cited as an example of how retailers can use first-party data to build personalized digital experiences that serve actual consumer intent. The home improvement retailer is building toward a homepage experience driven by factors including a shopper’s location, climate, and past purchase history, with a goal of having no two customers see the same page.
Minkow said the approach aligns with how agentic commerce works at its best: not broadly educating consumers about a topic but helping them research a specific need to make a purchase.
Baird said relevancy matters more than personalization for its own sake. A shopper who has bought gardening supplies repeatedly but is now planning a bathroom renovation does not want a homepage built around gardening.
“You really have to be careful about personalization for personalization’s sake versus context and memory, which are terms that are going to become more and more important in the AI world,” she said.
On privacy, Minkow said consumer research consistently shows privacy as the top concern with any new technology, even as consumers increasingly appreciate the payoff when personalization works.
“When they see it work for them, that is when it wins and when the creepiness tends to subside a little bit,” she said.
Baird said pricing is the place where the line is clearest. Using behavioral data to charge more to consumers who appear especially motivated to buy crosses into territory that destroys trust. “As a retailer, you are selling certainty, you’re selling decision enablement,” she said. “And if consumers feel like you’re using that to charge them more, that is a dangerous spot.”
Minkow said dynamic pricing, which is different than surveillance pricing, when transparent and structured to benefit the consumer, is a legitimate tool. She pointed to retailers that held or reduced prices during periods of economic stress as an example.
“Dynamic pricing can be compassionate,” she said.
Where AI Risk Is Low and Where It Is Not
Baird said product recommendations and personalized homepages carry low risk because a wrong answer does not cause real harm. The consumer simply sees something less relevant.
“If you got it wrong, that’s still a better attempt than not attempting anything at all,” she said.
Customer service automation is a different matter. Retailers that deploy chatbots primarily as a cost-cutting measure often build in friction that drives customers away.
Baird said the critical failure mode is an AI that routes frustrated consumers in circles without a path to a human.
“If you do that over and over again, you’re destroying any loyalty or any trust that consumer has with you,” she said.
For smaller retailers without large technology budgets, both speakers pointed to AI-enhanced product recommendations through existing e-commerce platforms as the most accessible entry point.
Baird added a note of caution: the more retailers rely on platform-provided AI tools, the more their recommendations will converge.
“The more you rely on AI to drive your product description strategy or recommendation strategy, the more it’s going to recommend the same kinds of things it recommends for everybody else,” she said.
Chatbot Metrics and the Conversion Rate Problem
Baird said many companies are publishing chatbot conversion rates without accounting for selection bias. The consumers who choose to engage with a chatbot are often already close to a buying decision.
“If you’re just funneling more people who are closer to a buying decision anyway, you are not improving your conversion rate,” she said. “Nobody talks about overall site conversion rate. And I’m not convinced that the conversion rate is actually something that is moving the needle for retailers when consumers use these chatbots.”
Minkow pointed to time spent on site before converting as a useful metric, with shorter times indicating a better result.
“Conventional wisdom in retail says that the longer time you get a shopper on your site, the better,” she said. “But that actually correlates to frustration in today’s environment.”
She also pointed to returns rates as a meaningful signal, one that AI should be actively used to analyze and reduce.
The AI Cost Question No One Is Asking
Baird said consumers and retailers alike are not paying anything close to the actual cost of the AI services they are using. Providers are subsidizing usage at a scale she compared to the free shipping era of ecommerce.
“We are absolutely not paying what it costs today,” she said. If inference costs were passed through at anything close to actual levels, she said, the economics of AI-powered retail features would look very different.
“If you have to pay 10 times what you’re paying in inference cost today, is this chatbot still profitable for you?” Baird said. “You have to ask those questions.”
