Welcome to The Median, DataCamp’s newsletter for November 7, 2025.
In this edition: Amazon clashes with Perplexity over agentic browsing, Microsoft’s new study reveals AI agent failures, OpenAI and Amazon ink a $38 billion cloud deal, Coca-Cola’s holiday AI ads spark new debates, Sora’s Android app sees a massive debut while Meta expands “Vibes” in Europe.
This Week in 60 Seconds
Amazon Sends Legal Threats to Perplexity Over AI Shopping Agent
Amazon issued a cease and desist letter to Perplexity, demanding the removal of its AI-powered shopping assistant, “Comet,” from its online store. The core of the dispute is whether the AI agent must identify itself as a bot. Amazon insists it must operate openly, while Perplexity argues that since the agent acts on a human’s behalf, it automatically has the same permissions as a human user. Perplexity fired back with a blog post titled “Bullying is Not Innovation,” bringing the growing tension between e-commerce sites and new agentic browsers into sharp focus. We’ll explore this high-stakes battle and what it means for the future of e-commerce in our Deeper Look section.
Microsoft Tests AI Agents in Simulated Marketplace, Finds Surprising Failures
Microsoft Research built an open-source simulated “Magentic Marketplace” to see how AI agents would behave in a real-world economy. Agents like GPT-4o and Gemini 2.5 Flash were found to be vulnerable to manipulation from business agents using tactics like prompt injection and fake social proof. Researchers also identified a “Paradox of Choice,” where agents became overwhelmed when presented with too many options, causing their performance and overall consumer welfare to decline. The study also found agents struggled with collaboration and showed systemic biases, such as accepting the first offer they received. We’ll connect these findings to the new agentic economy in the Deeper Look section.
OpenAI and Amazon Announce $38 Billion Cloud Deal
OpenAI announced a seven-year strategic partnership with Amazon, committing to purchase $38 billion in cloud computing services from AWS. The deal provides OpenAI with immediate access to AWS infrastructure, including hundreds of thousands of NVIDIA GPUs, to run and scale its advanced generative AI and agentic workloads. This move, which follows a restructuring that allows OpenAI to buy compute from firms other than Microsoft, is part of a larger mission to massively grow its computing power. All capacity from the new deal is targeted for deployment before the end of 2026.
Sora Android App Sees 470K Day-One Installs as Meta Launches “Vibes” in Europe
Meta announced it is launching its new Meta AI app across Europe, with the “Vibes” feature at its core. Vibes is a dedicated feed for creating, sharing, and remixing short-form, AI-generated videos, designed as a collaborative experience that can be cross-posted to Instagram and Facebook. This move comes just as its key competitor, OpenAI’s Sora, launched its Android app to a massive reception—an estimated 470,000 downloads on its first day. The Sora Android app’s release was limited to the U.S., Canada, Japan, South Korea, Taiwan, Thailand, and Vietnam. We’ve discussed AI-based feeds in depth in a previous newsletter.
Coca-Cola’s 2025 Holiday AI Ads Spark New Debates
Coca-Cola has released its new “Holidays Are Coming” commercials, which are once again generated by AI but show significant technical improvements over last year’s criticized attempts. The company’s CMO, Manolo Arroyo, has told the Wall Street Journal that AI makes the campaign “cheaper and speedier,” cutting production time from a year to about a month. However, the move continues to fuel debate over AI’s impact on creative jobs. While Arroyo insists the “engine of this is human storytellers” and that a team of artists works “frame-by-frame” to refine the images, one behind-the-scenes film noted that only five AI specialists were needed to generate over 70,000 video clips for an ad.
Free Access Week: November 3–9
Access DataCamp’s entire platform for free, including all 600+ courses, for an entire week.
A Deeper Look at This Week’s News
The Agentic Economy’s First Big Fight: Why Amazon vs. Perplexity Is Not Only About User Experience
This week, we got a glimpse into the near future of online commerce, and it looks more like a legal battlefield than a business race to improve the user experience.
The opening salvo was fired when Amazon issued a cease and desist letter to Perplexity. Here’s the dispute in a nutshell:
Amazon’s public stance: Amazon claims Comet violates its Terms of Service (TOS) because the AI agent fails to identify itself as an automated bot. Amazon insists this creates a “significantly degraded shopping and customer service experience.”
Perplexity’s public rebuttal: Perplexity argues that its agent is merely “delegated labor” acting solely on behalf of a human user. Inherently, the agent possesses the “same permissions” as that human shopper.
While the public-facing fight is about TOS and user experience, this dispute transcends a simple legal squabble. It is the first high-stakes battle over who will control the user journey—and the profits—in the new “agentic economy.”
The real stakes: Agentic walled gardens vs. the open web of agents
This conflict is the first major corporate clash between two fundamentally opposed futures for the internet, a choice that Microsoft researchers, in their paper “The Agentic Economy,” call the “agentic walled gardens” versus the “web of agents”:
The “agentic walled garden”: This is the model Amazon is building. It’s a closed ecosystem where the platform controls its own proprietary “siloed service agent” (like Amazon’s Rufus) to interact with customers inside its own store. In this model, Amazon dictates all the terms and, critically, protects its core business model.
The “open web of agents”: This is the model Perplexity represents. It uses an end-to-end agent that operates on the user’s behalf, roaming the entire web to find and execute tasks.
How the open web impacts advertising revenue
The Microsoft researchers explain that agents aim to provide general-purpose help, such as aggregating research or navigating websites to make reservations on a user’s behalf. However, the researchers highlight a fundamental problem with this approach.
This functionality is currently achieved through “computer use models” that merely “simulate a user pointing and clicking” on existing websites. They note this gives the illusion of agent cooperation, but because the business isn’t actually participating with its own agent, the model has limits.
Most importantly, the paper warns that by “mimicking human users,” these agents “risk creating adversarial relationships with businesses.” This is especially true for businesses that “rely heavily on advertising revenue” and will naturally “resist having their websites accessed by agents instead of humans.”
And Amazon relies very heavily on advertising revenue.
Amazon’s $17.7B fear: The “zero-click disruption”
Amazon’s retail media and advertising business is a juggernaut, reportedly generating $17.7 billion in Q3 this year. This entire advertising model is built on the battle for the conversion funnel—controlling what you see, and in what order, to expose you to sponsored products, curated lists, and upsells.
AI agents are engineered to bypass this entire “influence economy.” An agent’s job isn’t to browse sponsored listings. Its job is to analyze data and find the objectively best product at the best price, regardless of who paid for top placement. This is the “zero-click disruption,” and it threatens to annihilate Amazon’s highest-margin revenue stream.
Neutralizing dynamic pricing
Another pillar of modern e-commerce is information asymmetry—the simple fact that a human shopper cannot possibly compare prices across 100 different websites in real-time.
Platforms use this fact to implement dynamic pricing, adjusting margins based on user behavior, demand, and competitor prices. Agents eliminate this asymmetry instantly, collapsing a platform’s ability to strategically manage prices.
Amazon’s public fears are also valid (at least according to early research)
While Amazon is clearly protecting its ad business, its public fear of a “degraded” platform isn’t just a convenient excuse. As it turns out, new research from Microsoft shows that a world run by today’s AI agents would be chaotic, unpredictable, and easily manipulated.
Microsoft Research, in collaboration with Arizona State University, built the “Magentic Marketplace”—a simulated e-commerce world where “assistant agents” (customers) were prompted to buy from “service agents” (businesses). The results were alarming and prove Amazon’s public fears are valid:
Agents suffer a “Paradox of Choice”: Agent performance and consumer welfare decreased as the number of options increased. The agents became “overwhelmed” by too many choices, failing their core optimization task.
Agents are vulnerable to manipulation: “Service agents” (businesses) easily tricked “customer agents.” Models were consistently fooled by simple tactics like fake social proof (e.g., “Join 50,000+ satisfied customers”) and prompt injection attacks, validating fears of a chaotic, untraceable, and manipulation-filled marketplace.
Agents accept the first offer: Instead of optimizing, agents demonstrated a strong “first-offer acceptance” bias. They frequently accepted the first offer they received without shopping around for better options, defeating their primary purpose.
More options reduced customer welfare, revealing a Paradox of Choice effect (Source: Microsoft)
The future: An agent transaction fee?
An emerging solution to Amazon’s public problem is the Agent Payments Protocol (AP2). It’s designed to solve the trust issue by using cryptographically-signed “Intent and Cart Mandates.” These mandates would provide an auditable, verifiable record that proves a user actually gave an agent permission for a specific task (e.g., “buy these shoes, up to $50”).
However, this does not solve Amazon’s unspoken, multi-billion-dollar problem: AP2 is a payment protocol, not an advertising one, and agents would still bypass the “influence economy.”
However, the AP2 protocol opens up the possibility of a new revenue stream. Instead of charging advertisers to influence human choice, platforms would charge for the service of providing a secure, auditable, and trusted transaction for AI agents. This could mean a new “agent transaction fee” on every AP2-compliant purchase.
Industry Use Cases
Google AI Helps Scientists Model and Monitor Ecosystems
Google DeepMind announced new AI-powered initiatives to model the Earth’s biosphere. The research includes using vision transformers to predict deforestation risk from satellite data, deploying a Graph Neural Net (GNN) to map species distributions at scale, and applying the Perch 2.0 foundational model to classify animal vocalizations. These tools are already being used in the field, helping scientists in Hawai’i, for example, monitor endangered honeycreepers by identifying their calls to understand population health.
Lloyds Bank to Launch AI Financial Assistant for 21 Million Customers
Lloyds Banking Group announced the launch of a large-scale AI-powered financial assistant for its 21 million mobile app customers. The new tool will use agentic AI to act as a financial companion, providing round-the-clock, personalized financial guidance, spending insights, and savings support. Built on the bank’s internal generative AI framework, the assistant can transfer customers to human experts when needed. Lloyds plans to expand the tool’s capabilities to cover mortgages, car finance, and protection services from 2026 onwards.
Chime Outlines Agentic Model for Marketing
In a case study published by OpenAI, fintech company Chime outlined its strategy for an agentic marketing model. CMO Vineet Mehra explained that his team is moving beyond simple tasks, using AI for real-time media optimization and customer insight analysis. The company also built a custom “Chime Content GPT,” trained on its best-performing content, to maintain brand voice and accelerate creative production.
Tokens of Wisdom
Human in the loop is essential. Both as a quality gate, but also to build that trust with users, who are the ultimate consumers of your solution, who have a say in whether it is reliable.
—Shane Murray, Senior Vice President of Digital Platform Analytics at Versant Media
Listen to stories of AI disasters and AI successes in our latest DataFramed podcast with Shane Murray.