The invisible tech revolution poised to make in-chat AI shopping real

For those who read my ‘I hate shopping’ manifesto in Eye on AI a few weeks back, where I shared how ChatGPT became more helpful for my style quests than Google, there was an obvious follow-up: What if I could make purchases in my AI chat? 

When I started writing about the early versions of AI agents two years ago, that was the holy grail developers pointed to—shopping. For some reason, though, they were always interested in how to have an AI agent buy your pizza, not pick out and order the best dress for that summer wedding next month. 

But pizza or not, those early efforts mostly fizzled, because there were so many seemingly insurmountable hurdles. Early agent efforts in 2023, like Auto-GPT and BabyAGI, were primarily focused on surface-level web manipulation—that is, having an agent operate a web browser and the user’s computer keyboard and mouse, and giving it access to a file with passwords and credit card data. But even the best AI systems can misinterpret website layouts, leading to transaction failures, and AI computer usage tools are relatively slow. Having a file sitting around with all your password and credit card data also creates cybersecurity risks. 

Giving the AI direct access to vendors’ backend systems would be a simpler, and better, way to enable agentic commerce. But every vendor’s backend—like a pizza place’s payment processing, a bookstore shopping cart, Amazon’s inventory—is its own walled garden that needs logins and authentication. There was the lack of persistent memory for AI agents, who might not remember where you left off while picking out that dress. Then there’s payment information—not surprisingly, most of us are reluctant to hand over our financial information to AI agents, and anyway, how could ChatGPT handle credit card or Venmo transactions anyway? 

This technology is changing the AI shopping game

That’s all changing now, thanks to an under-the-radar AI revolution called Model Context Protocol (MCP). Developed by Anthropic last November, the MCP architecture allows AI agents to “remember” context, execute transactions, and manage interactions in real-time across multiple platforms. With a standardized layer connecting to payment gateways, e-commerce inventories and financial services, AI agents are on the brink of being able to do far more than just window-shop. 

Of course, for this to work, all the different parties involved have to agree to use MCP. That’s happening: Over the past two months, Google, Microsoft and OpenAI have all agreed to use MCP. Payment processors have followed suit: Visa is also partnering with Anthropic, Microsoft, Mistral, OpenAI, Perplexity, and other leading tech players to bring seamless payment capabilities directly into AI chatbots, while MasterCard also announced the launch of its agentic payments program called Agent Pay. Stripe and Block (formerly Square) have gotten on board with MCP as well, and yesterday, PayPal announced a partnership with AI search startup Perplexity to power agentic commerce in its Pro platform. Starting this summer, the company said consumers can check out with PayPal or Venmo when they ask Perplexity to find products, book travel, or buy tickets. 

However, retailers have not jumped on board as fast as the payment processors. One company, Retail-MCP.com, has created a managed MCP service for retailers, but it’s not clear any major brand has launched an MCP server yet. But it will happen. 

For PayPal, agentic e-commerce is coming fast

I spoke yesterday with Prakhar Mehrotra, PayPal’s head of AI, who told me that Anthropic’s release of MCP marked the moment he realized agentic e-commerce was becoming a reality. 

“The unsolved problem was how two agents could communicate with each other,” he said. First, he explained, the AI needs to understand what you want to do—like booking a flight to Paris or reserving a table at a restaurant. But then other agents need to work together: One agent might handle flight booking, for example, while another handles a local dinner reservation and a third tackles payment processing for both. Until now, agents could not communicate with each other, so a chatbot could only recommend—not complete the actual purchase on multiple platforms. 

Now, six months after the release of MCP, the age of “identity commerce” is upon us, Mehrotra said. AI now has the ability to handle user transactions that are secure and personalized across platforms, so agents can complete real-time purchases without any manual logins (and without having to use a web browser; MCP works by directly transmitting data back and forth in a standardized protocol between the AI agent and an MCP server operated by a retailer that is seperate from the one hosting its e-commerce website). 

According to PayPal, the entire AI-powered shopping process—from payment and shipping to tracking and invoicing—will happen in the background. It will connect directly to your PayPal or Venmo account, using emerging technology that securely stores your payment details and lets you check out with just a single click or voice command—without the need for passwords. 

It’s still a little hard for me to imagine using Venmo or Visa to buy a new outfit using Perplexity or ChatGPT. But I’ll be an early AI shopping adopter, no doubt—and I’ll let you know how it goes.

With that, here’s the rest of the AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

AI IN THE NEWS

UAE close to AI chip deal with Nvidia, following similar Trump deal with Saudi Arabia. Reuters reported this morning that the United States has reached a preliminary agreement with the United Arab Emirates to permit the import of 500,000 of Nvidia's most advanced AI chips annually, beginning in 2025. According to two anonymous sources familiar with the matter, the agreement is expected to last until at least 2027, with the possibility of extending to 2030. This deal is set to accelerate the UAE’s development of data centers crucial for building advanced AI models. On Tuesday, I reported on a similar deal with Saudi Arabia: Trump granted Humain, a new AI company created last week by Saudi Arabia’s Public Investment Fund, as well as United Arab Emirates-based AI firm G42, increased access to U.S.-made advanced AI chips, particularly from Nvidia. The moves come as the Trump administration has scrapped export controls implemented under the Biden administration that limited the Saudis’ access to cutting-edge American technology. Humain also announced strategic partnerships with Nvidia and AMD to leverage their AI chips and infrastructure to establish Saudi Arabia as a global leader in AI.  

Google DeepMind introduces AlphaEvolve, a coding agent that can design its own advanced algorithms. Google DeepMind released AlphaEvolve, its latest AI-powered coding agent that is designed to autonomously discover and optimize advanced algorithms for both mathematical and practical computing challenges. AlphaEvolve uses Google’s Gemini models in combination with automated evaluators to iteratively generate, assess, and improve upon complex code solutions. DeepMind said AlphaEvolve has already been used to enhance Google’s data center efficiency, optimize chip design, and even improve AI training processes, including training the models that power AlphaEvolve itself. According to Wired, AlphaEvolve “demonstrates the potential for AI to come up with completely novel ideas through continual experimentation and evaluation,” which AI companies hope will lead to AI agents gradually learning to exhibit novel ideas in other areas besides coding. 

China restricts schoolchildren’s use of generative AI. According to CNBC, China is introducing new regulations to manage schoolchildren’s access to artificial intelligence technologies. Under the new guidelines, primary school students are banned from independently using unrestricted generative AI tools, though teachers are permitted to integrate the technology into classroom activities. Middle school students are allowed to explore how generative AI reasons and analyzes information, while high school students are granted broader usage rights. The national state media report did not specify exact restrictions but emphasized that the policies aim to encourage the “scientific” and “standardized” promotion of AI education tailored to different educational levels, as translated by CNBC.

OpenAI introduces HealthBench, a health dataset to evaluate AI models. On Monday, OpenAI debuted HealthBench, an open-source benchmark designed to evaluate the performance and safety of AI models in healthcare settings. Developed in collaboration with 262 physicians from 60 countries, HealthBench comprises 5,000 realistic, multi-turn medical conversations that span various specialties and languages. According to an OpenAI blog post, Improving human health “will be one of the defining impacts of AGI,” but to get there models must be useful and safe. “Evaluations are essential to understanding how models perform in health settings,” the blog post said. “Significant efforts have already been made across academia and industry, yet many existing evaluations do not reflect realistic scenarios, lack rigorous validation against expert medical opinion, or leave no room for state-of-the-art models to improve.” 

FORTUNE ON AI

Samaya AI, startup building AI for financial services, raises $43.5 million in VC funding —by Jeremy Kahn

Nvidia’s huge deal with AI startup Humain puts Saudi Arabia at ‘the front of the line’ of global chip customers, Dan Ives says —by Sasha Rogelberg

Saudi Arabia wants to build its post-oil future with massive AI data centers—Trump and U.S. tech have big incentives to oblige —by Sharon Goldman

AI CALENDAR

May 19-22: Microsoft Build, Seattle

May 20-21: Google IO, Mountain View, Calif.

May 20-23: Computex, Taipei

June 9-13: WWDC, Cupertino, Calif.

July 13-19: International Conference on Machine Learning (ICML), Vancouver

July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here.

Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.

EYE ON AI NUMBERS

60%

That’s the share of Gen Z adults who say generative AI is having a “significant impact” on their work lives, according to the second annual KPMG American Perspectives survey of 2,500 adults. Among millennials, 57% reported a significant impact, while only 41% of Gen X respondents felt the same.

Gen Z also leads in terms of personal life impact. 45% of Gen Z adults said generative AI is significantly affecting their day-to-day lives, up from 41% the previous year. Gen Z (57%) and millennials (55%) are the most likely to say AI’s influence is substantial, compared with just 40% of Gen X.