OpenAI’s efforts to turn ChatGPT into a personal shopper are running into obstacles with reading, analyzing, and processing product data, according to a report by The Information.
The company said in September that its in-app checkout would soon be available to millions of shops through its partnerships with Shopify and Stripe, but those promises have yet to materialize.
Obstacles include:
- ChatGPT needs to interpret information like pricing and inventory availability on demand, which is a lot of real-time data.
- Merchant data is not standardized and can be open to interpretation by AI agents. For example, a retailer may label something as “In Stock” in its system, even if its only available for a back order or preorder. So now it’s got to analyze that real-time data and process it for accuracy or it’ll make mistakes.
- Shopify says that building the checkout has required a lot of upfront work, but that selling on ChatGPT will be available for all its U.S. merchants very soon.
- Meanwhile Stripe has had to build a connection between the AI agents (ChatGPT in this case) and merchants’ backend systems (Shopify) so that the agents can properly read and understand product catalogs and send transaction information between the two parties. That’s hard to accurately when both parties aren’t quite ready.
Honestly, in recent months, it’s felt like OpenAI’s news and announcements around agentic commerce were wildly ahead of its readiness to implement on those promises. The industry has felt more like a headline race than an AI race.
OpenAI really needs its agentic commerce efforts to work. The company has told investors that it plans to generate around $110B in revenue from nonpaying users by 2030, and taking a cut of sales through its in-chat checkout is expected to play a big role in that mission.
Meanwhile Amazon is experiencing its own AI setbacks as it attempts to replace rival technologies with its own Nova AI models.
The company has spent the past three years building its own LLMs that offer a cheap alternative to leading models from third-parties, but aren’t quite good enough to fully replace them. For example, Amazon’s shopping assistant Rufus uses a mix of its in-house Nova models and models developed by Anthropic, while its vibe-coding product Kiro exclusively uses third-party models and not Nova.
Part of the reason for Amazon’s slow AI growth may have to do with its financially disciplined approach to developing its models. Most of the company’s AI team was compiled from pre-existing Alexa teams, and it has avoided throwing huge compensation packages at recruits like Meta and OpenAI, which has limited the amount of fresh talent it could bring in to the division.
An Amazon spokesperson said the strategy is working for them:
“The framing that this is a zero-sum game between Nova and other model providers misses what we’re actually doing. Our strategy is driven by the desire to offer AWS customers and our Amazon businesses as many AI options as possible, and that means investing heavily in the Nova models and our partnerships.”
Keep in mind that Amazon owns a significant stake in Anthropic, so it’s not exactly a losing proposition if it relies too heavily on their models.
