Amazon’s $180 billion quarter revealed AI shopping assistants are working, but CEO Andy Jassy admitted the data they depend on isn’t ready. Inside the catalog systems that will determine whether the $5 trillion opportunity materializes.
Inside an AWS data center in eastern Oregon. Amazon’s $125 billion in 2025 capital expenditures is building the physical infrastructure for AI, but the data systems inside may matter more. (Credit: Amazon)
Amazon’s Q3 2025 earnings told two stories.
The headline numbers were impressive: $180.2 billion in revenue, up 13% year-over-year. AWS accelerated to 20% growth. “A pace we haven’t seen since 2022,” CEO Andy Jassy said on the October earnings call.¹ Rufus, Amazon’s AI shopping assistant, had reached 250 million active users. Customers who engage with it are 60% more likely to complete purchases. The chatbot is on track to generate over $10 billion in incremental annualized sales.²
But buried in Jassy’s remarks was a more revealing admission. Asked about “agentic commerce,” shopping mediated by AI agents acting autonomously on behalf of consumers, he acknowledged Amazon would need to “partner with third-party agents,” comparing the moment to “the beginning of search engines many years ago being sources of discovery for commerce.”
Then came the candid part:
“Right now, I would say the customer experience is not [good]. There’s no personalization, there’s no shopping history, the delivery estimates are frequently wrong, the prices are often wrong.“
The problem isn’t the AI. It’s the data the AI needs to work with.
The Stakes: $5 Trillion by 2030
OpenAI’s Instant Checkout lets users purchase products without leaving the ChatGPT interface. (Credit: OpenAI)
McKinsey projects that agentic commerce will orchestrate $3-5 trillion in global revenue by 2030.³ Morgan Stanley estimates AI shopping agents could drive $190-385 billion in U.S. e-commerce spending alone, capturing 10-20% of the market.⁴ By 2030, the firm projects AI agents will grow from near-zero to 126 million users, nearly half of all online shoppers.
The race is already underway. OpenAI launched Instant Checkout in September, letting ChatGPT’s 700 million weekly users buy directly from Etsy and Shopify merchants without leaving the chat.⁵ Google released its Agent Payments Protocol, backed by Mastercard, PayPal, American Express, and Alibaba. Amazon launched “Buy for Me” in April, an AI agent that purchases from third-party websites directly within the Amazon app.
“Agentic will be a paradigm shift for e-commerce. With greater digitization of consumers’ wallets, this could shake up the e-commerce funnel with implications across retailers and digital advertising players.”
— Nathan Feather, Morgan Stanley analyst
Washington has taken notice too. President Biden’s January 2025 Executive Order on AI Infrastructure warned that AI “is too important to be offshored.”⁶ President Trump’s AI Action Plan, released in July, declared it “a national security imperative for the United States to achieve and maintain unquestioned and unchallenged global technological dominance.”⁷ The $500 billion Stargate initiative reflects the same bipartisan conviction: AI infrastructure is strategic infrastructure.
But here’s what gets lost in the policy debates and earnings calls: all these AI agents share a common dependency. Before they can negotiate prices, compare products, or complete transactions, they need accurate, structured data about what’s actually for sale.
They need the catalog. And at Amazon’s scale, keeping that catalog trustworthy is one of the hardest problems in e-commerce.
The Catalog Problem
I spent the past month talking to engineers across the industry, at Amazon, Stripe, Shopify, Walmart, and several startups building agentic commerce infrastructure, trying to understand what actually has to work for AI shopping agents to deliver on the hype.
The answer I kept hearing surprised me. It wasn’t about better models or smarter agents. It was about data.
Amazon’s catalog contains hundreds of millions of products. Without active management, the number of active ASINs (Amazon Standard Identification Numbers) would have ballooned from 62 billion to over 90 billion items. That volume creates compounding problems: duplicate listings, stale data, orphaned products from terminated sellers, inconsistent pricing, inventory records that don’t match reality.
When Jassy told analysts that third-party agents provide “delivery estimates [that] are frequently wrong” and “prices [that] are often wrong,” he was describing what happens when AI systems encounter catalog data that hasn’t been properly maintained. The agent isn’t hallucinating. It’s reading bad data.
“Everyone focuses on the model. But the model is only as good as what it’s querying. Garbage in, garbage out. Except now the garbage is being served to AI agents that will make purchasing decisions. The stakes are completely different.”
— Principal Engineer, Fortune 100 e-commerce company
One Engineer’s Fingerprints
Several engineers I spoke with pointed me toward Ausaf Qazi, a Senior Software Development Engineer within Amazon’s Items & Offers Platform. His name came up repeatedly when I asked who was doing the most critical infrastructure work on catalog systems.
Qazi’s day-to-day work involves designing and implementing the data pipelines that process item updates across Amazon’s catalog, building validation systems that catch inconsistencies before they reach customer-facing applications, and architecting the security controls that protect catalog data from manipulation.
“Ausaf doesn’t just maintain systems. He identifies failure modes before they happen. Last quarter, he redesigned our merge workflow to handle edge cases that were causing data inconsistencies downstream. It wasn’t glamorous work, but it prevented what could have been significant customer-facing issues.”
— Staff Engineer, Amazon
Qazi is a named contributor to Amazon’s Catalog Systems, part of a VP-level initiative that delivered striking results in 2024: 20.5 billion ASINs tombstoned, 43.5 billion SKUs deleted, $8.53 million in annualized infrastructure cost avoidance. His specific contributions included building automated data extraction workflows that replaced manual querying across 15+ tables, implementing validation systems that prevent accidental deletion of active products, and engineering improvements that increased processing capacity from 500 to 3,500 transactions per second. That’s a 7x improvement.
“The catalog work required someone who could think across the entire system. You’re touching fraud detection, seller management, inventory systems, search indexing. Get any of it wrong and you’ve got customer impact. Ausaf has the judgment to navigate that complexity.”
— Senior Manager, Amazon
He also holds code review authority across multiple teams and change management approval authority for production deployments affecting Amazon’s global infrastructure. That’s not a designation you hand out casually.
What Other Companies Are Building
To understand how Amazon’s work fits into the broader landscape, I reached out to engineers at Stripe, Shopify, and Walmart.
At Stripe, infrastructure teams have been building the payment rails that AI agents will use to transact. The company’s Agentic Commerce Protocol, co-developed with OpenAI, enables transactions directly within AI chat interfaces.
“Stripe is building the economic infrastructure for AI. That means re-architecting today’s commerce systems.”
— Will Gaybrick, President, Stripe
Shopify has taken a different approach, developing what they call “agent-friendly APIs” that standardize how AI systems query inventory.
“Shopping is changing fast. People are discovering products in AI conversations, not just through search or ads. We had to rethink how product data flows.”
— Vanessa Lee, VP of Product, Shopify
Walmart’s commerce platform engineers have been restructuring product data to be agent-readable since early this year. Google, Visa, Mastercard, and PayPal are all developing agent-specific payment capabilities.
The picture that emerged from these conversations: an entire industry retooling its infrastructure simultaneously, with engineers at every major player racing to solve overlapping problems.
The Open Source Work
What distinguishes Qazi from many infrastructure engineers is that his work extends beyond his employer.
Sentinel, his open source edge computing project, addresses a gap in how Kubernetes handles network partitions. Traditional cloud systems assume reliable connectivity. That assumption fails catastrophically at the edge, particularly for retail store locations with intermittent networks running edge infrastructure.
The retail application matters for agentic commerce. When a store’s network connection drops, inventory systems go blind. AI agents querying product availability get stale data, or nothing at all. Qazi’s implementation keeps edge systems operational during outages, ensuring inventory counts and pricing remain synchronized even when connectivity fails.
“I evaluated Sentinel for our 5G edge deployment. What struck me wasn’t just the technical implementation. It was the design sensibility. He’d anticipated failure modes that most engineers don’t think about until they’re debugging production incidents. The partition-resilient consensus work is genuinely novel.”
— Distinguished Engineer, telecommunications company
The project has attracted contributors adapting it for retail point-of-sale systems, industrial IoT, and autonomous vehicle infrastructure.
Qazi also sees another infrastructure gap looming: the economics of AI queries at scale.
“When you have millions of AI agents making product queries, even cheap API calls compound into real costs. And a lot of those queries are semantically identical. ‘What’s the return policy?’ and ‘How do I return this?’ should hit the same cached response, but traditional caching only catches exact string matches.”
He’s been working on a semantic caching layer that matches queries based on meaning rather than exact text. Early benchmarks suggest 40-70% cost reductions for applications with repetitive query patterns. That’s exactly the profile of agentic commerce workloads.
“It’s not ready for production yet. But as the agent scale increases, someone’s going to need to solve this.”
Engineers at Anthropic and OpenAI have been working on similar inference optimization problems. Startups like Baseten and Modal are building infrastructure for high-volume AI workloads. The economics of agentic commerce will depend on solving these scaling challenges across the stack.
The Competitive Picture
Amazon’s position in agentic commerce is complicated.
The company has blocked 47 AI bots from crawling its website, protecting its proprietary data from competitors.⁸ But it’s also building aggressively: AWS launched AgentCore for deploying AI agents at scale, and the SDK has been downloaded over 1 million times. Rufus keeps expanding. It can now auto-purchase items when prices hit user-defined thresholds.
CNBC recently characterized Amazon’s situation as “the leader’s dilemma.”⁸ Jordan Berke, founder of retail consulting firm Tomorrow, put it bluntly:
“Their market share is so significant that they have the most to lose.”
Whether Amazon blocks external agents or partners with them, both paths require the same foundation: accurate, structured, trustworthy product data.
What Happens Next
Jassy’s Q3 call previewed where Amazon thinks this is heading.
“I still think if you look at the worldwide market segment share of retail, still 80% to 85% of it lives in physical stores. And that equation is going to flip over time. And I think AI is going to only accelerate that.”
The flip requires infrastructure. Not just AI models and chatbots, but the foundational data systems those models query. When Morgan Stanley projects 126 million AI shopping agents by 2030, each agent will need reliable answers to basic questions: What’s in stock? What does it cost? Can it ship today?
The answers come from infrastructure, the unsexy, unglamorous systems that engineers like Qazi are building and refining. The quality of those systems will determine whether agentic commerce delivers on McKinsey’s $5 trillion projection or stumbles over the same data quality problems Jassy admitted are plaguing current implementations.
“I’m very excited about, and as a business, we’re very excited about in the long term, the prospect of agentic commerce.”
— Andy Jassy, CEO, Amazon
Amazon’s $125 billion in 2025 capital expenditures reflects that excitement. But capital builds data centers. Engineers build the systems that make the data useful.
The AI agents are coming. Whether they actually work depends on infrastructure most people will never see.
References
- Amazon Q3 2025 Earnings Call (October 30, 2025)
- Fortune, “Amazon says its AI shopping assistant Rufus is on pace to pull in an extra $10 billion in sales” (October 2025)
- McKinsey & Company, “The agentic commerce opportunity” (October 2025)
- Morgan Stanley Research, “Agentic Commerce Outlook” (October 2025)
- OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol” (September 2025)
- White House, “Statement by President Biden on the Executive Order on Advancing U.S. Leadership in Artificial Intelligence Infrastructure” (January 14, 2025)
- White House, “America’s AI Action Plan” (July 2025)
- CNBC, “Amazon faces a dilemma: fight AI shopping agents, or join them” (October 2025)
- Modern Retail, “Amazon CEO expects to ‘find ways’ to partner with third-party AI shopping agents” (October 2025)
Related Topics: Amazon, AWS, Agentic Commerce, AI Infrastructure, E-Commerce, Catalog Management, Data Quality
