The world of agentic AI commerce is booming right now. PayPal, Amazon, the big credit card companies, Walmart and more are in on the trend and purportedly believe it’s not a trend at all, but the future of shopping.
Although it’s not just global behemoths. New Generation AI, a startup founded last year, is looking to build agentic AI products for brands and retailers of any size.
It’s an exciting time, said Jonathan Arena, one of New Gen’s co-founders and a former product lead at Patreon. The other two co-founders were product leaders at Stripe and Meta. But even while every company is adopting AI in some ways, there is in retail, especially, a “fear of rip and replace,” he said.
One client has 47 mar tech vendors that plug into its database, he added. All of those integrations must play nice with the AI services and vice versa.
But the vision is for an AI-based solution that “doesn’t need to be run through all of those kinds of vendor systems,” said co-founder Adam Behrens.
New programmatic
In conversations regarding AI-based commerce, the word “programmatic” often appears. But it doesn’t mean the same as data-driven programmatic advertising.
In AI parlance, “programmatic” is shorthand to describe connections made between two API solutions or AI agents. Another term that’s taken root for the idea is “headless commerce,” Arena noted.
In this vision of a programmatic or headless shopping experience, a human wouldn’t do so much of the slogging work of browsing sites or evaluating product reviews. That would fall to a personal agent. And the agent wouldn’t just browse the site, either. Rather than perusing Nike.com, for example, that shopping agent would ping Nike’s agentic AI. The shopper’s AI agent and the merchant’s agent would pass info back and forth, like where the person lives, what they’re shopping for, a potential budget and preferences based on similar past purchases.
If someone’s looking for new shoes because they’re considering running a marathon, Arena said, the Nike AI might pass back shoe models that seem the best fit, and perhaps a YouTube video and a Reddit thread that documents a long-distance runners’ shoe preferences.
This packet of product info and potential purchase options would be presented to the actual human. Ta-da!
Now, the tough part
This kind of headless commerce technology exists right now, Behrens said.
But it hasn’t yet become an actual shopping behavior beyond maybe some serious AI hardos in San Francisco.
For a SaaS and AI vendor like New Gen, the focus is on the business solutions side of the equation. Educating humans about AI shopper assistants and gaining adoption is somebody else’s problem, although that’s an equally important part of the equation. Otherwise, it’s like lifting only one side of a heavy barbell in a gym.
But getting businesses on track to even properly run new AI systems is a tall order, too.
For a retailer or manufacturing brand, it’s not just adding another vendor to the scores they already work with. Take that Nike example, for instance. When the shopper’s agentic AI visits the site, Nike shouldn’t even be presenting the same web experience it shows to actual people, Behrens said. There are vast costs that come with serving pages over and over again, especially now that there are so many LLMs and AI-based vendors that routinely, even many times daily, scrape every page and subdomain.
Currently, publishers and retailers tend to treat these site visits like malicious bots, he said, trying to filter them out by using captchas or some other bot detection tech. Instead, he said, there should be a different portal, like AI.Nike.com, where the AI tech goes to pass info agent to the agent. The costs could then be metered, akin to how people pay for data usage with enterprise AI models.
And then brands still need to wrangle the AI tech itself.
New Gen itself is constantly testing and switching out the various models it uses for different purposes. The company uses Google’s Gemini 2.5 model for image generation, having swapped it out with ChatGPT 4.0 because Gemini has a more robust API, Arena said. For code generation, though, New Gen uses ChatGPT 4.0. And for text generation the company uses Anthropic’s models, because it does a better job of preserving a brand’s distinct tone for creative copy.
Although, this conversation was from three weeks ago, so who knows how the product dynamics have changed in such a long time.
And the decisions about which models to use in which circumstances are further complicated because those decisions go beyond just quality. One model might generate superior videos but come at a greater cost in terms of, well, dollars, but also because it is slower.
Picking the models that are cheaper and faster is often just as important, Arena said.
Asking for data
AI commerce startups have another obstacle to adoption, which is the vast amount of data they require from business customers.
“Brands have never really had an incentive to expose an API on top of their product data, like pricing data and inventory data,” Behrens said.
Right now, brands do expose that kind of info to companies like Google and Meta, because businesses feel that they must adapt to how those platforms work, he said.
Although even that is a relatively new trend.
For example, earlier this month Meta announced new AI-based commerce products and beta programs that allow brands to optimize ad campaigns based on details like profit margin. But that means advertisers must provide a new data feed to Meta, which gives the ad platform profit-margin breakdowns, product by product.
“We’re starting to see, especially in San Francisco, a whole set of companies that now want access to this brand data in order to build net-new experiences,” Behrens said.
Now brands just need to get comfortable sharing that data to their AI vendors.
