On the two jobs of shopping, and why AI-led e-commerce needs new rails
Think about the last two things you bought online.
One might have been your weekly grocery order. You opened an app, tapped a few familiar items from your "previous orders" list, maybe added a new snack or two, and checked out in under three minutes. A task completed efficiently.
The other might have been a bigger purchase, say, a new television. That wasn't a 3-minute task; it was a project. For purchases with significant consideration, the journey is not as easy as buying groceries. It is a multi-stage process involving Awareness, Research, Consideration, Decision, Purchase, and the crucial Post-Purchase Experience.
It likely started weeks earlier with a vague search like "best 65-inch TV," which quickly spiraled into a dozen open browser tabs. And suddenly you are deep in technical specifications — OLED vs. QLED, refresh rates, HDR formats — and scrolling through endless user reviews, trying to decipher which ones were genuine and which were fake. These two experiences reveal a durable truth about digital commerce. Shopping has two jobs.
The first job is the transaction — the path to payment. Over the past two decades, companies like Amazon optimized this to near friction lessness with one-click checkouts, saved addresses, and reliable delivery windows. The second job is the decision — the research, comparison, validation, and ultimately the commitment to one option under uncertainty. For considered purchases, the second job is the real work, and the burden has largely sat with the shopper.
For years we accepted this division as the price of abundance. Today, that arrangement is changing. Artificial intelligence is moving from “helpful tool” to “capable delegate,” taking on the heavy lifting of decision-making and, in the process, rewriting the economics and experience of commerce.
You're Hired: The Unpaid Job of Being a Modern Shopper
Every considered purchase quietly turns you into an employee. You’re the project manager, the researcher, the analyst, and the final approver — all rolled into one. The only payoff is the hope of avoiding a costly mistake.
And it is real work. Google’s own data shows that a typical buyer spends nearly two weeks and consults more than a dozen sources before committing to a big purchase. Along the way, shoppers bounce between sites and reviews, trying to separate genuine insight from the fake and misleading. One recent survey found that the majority of consumers have run into fake reviews in the past year — which means every time you scroll, you’re also doing the unpaid job of a fraud detector.
No surprise, then, that uncertainty stalls many purchases at the last mile. A quarter of buyers abandon carts because they don’t fully trust the site; others back out over return policies that feel risky. The pattern is clear: buyers shoulder too much of the burden, while sellers spend heavily to push them across the line — often with diminishing returns.
The Great Handover: AI Takes on the Decision-Making Work
The most immediate shift AI is bringing to commerce is automating this unpaid, high-effort job of being a modern shopper by taking on the cognitive load of decision-making. Large Language Models (LLMs) can act as tireless research assistants, absorbing messy, unstructured information and returning structured, tailored guidance.
In place of opening 20 tabs to compare the camera specs on three different smartphones, you can now ask an AI agent to do it for you and summarize the key differences in a simple table.
Instead of you reading through 500 user reviews to see if a particular hiking boot runs narrow, you can now ask an AI to analyze them and report on the consensus.
Instead of you trying to understand the long-term cost of ownership for two different electric vehicles, you can now ask an AI to model it for you based on your driving habits, local energy prices, and projected resale values.
This is the great handover. The manual, time-consuming labor of information gathering and analysis is shifting from the user to the machine. AI is perfectly suited for this role. It can process and synthesize vast amounts of unstructured data from across the web in seconds, providing a curated, personalized, and digestible summary that directly addresses your needs.
This fundamentally changes your role from a researcher wading through digital noise to a director, asking precise questions and receiving tailored intelligence to inform your final choice.
An AI Assistant Isn't Enough: The Limits of Today's Tools
This new wave of AI assistants is undeniably powerful, and their impact is already being felt. Industry forecasts predict a tectonic shift, with Gartner anticipating a 25% drop in traditional search engine volume by 2026 as users turn to AI for direct answers. However, these first-generation tools are an intermediate step, not the final destination. Three limitations stand out.
First, generalist models are not specialized shopping agents. Researchers from Columbia Business School and Yale have found that they are trained to be competent at many tasks, not expert at the specific multi-objective tradeoffs in purchasing decisions. In controlled evaluations, state-of-the-art models showed non-trivial failure rates on simple economic rationality tests — such as choosing the lowest price among identical options — and performed inconsistently when weighing quality differences against price. A system that occasionally picks a more expensive, lower-rated item is a system a buyer will not delegate to at checkout.
Second, model heterogeneity is real. Different models make different choices given the same product set, sometimes dramatically so. If the “best” product depends on which model you asked, uncertainty for the buyer increases rather than decreases.
Third, hallucinations have concrete consequences. The now-famous incident in which an airline was held liable for a chatbot’s invented refund policy is a warning that confidence without grounding is risky in commerce. As assistants become a more common interface to purchasing, we cannot rely on ad-hoc scraping and probabilistic synthesis to carry the full weight of commercial decisions.
Forecasts reflect this tension. Analysts expect some portion of traditional search to shift to AI interfaces over the next few years. That change will not reach its full potential if the underlying data and protocols remain human-centric and adversarial. To move from “faster browsing” to deterministic, accountable transactions, we need a different foundation.
Rebuilding the Foundation: From a Scraped Web to an Intent-Driven Market
For AI to evolve from helpful research assistants into true, autonomous agents that can act on our behalf, they need a different kind of internet. They cannot rely on interpreting the chaotic, human-centric web. They require a foundational layer built for machines — a new set of rails for commerce.
This view is echoed by leading industry analysts. As the venture capital firm Andreessen Horowitz (a16z) notes, AI agents can’t fulfill their potential in commerce without major shifts in the substrate they operate on. They argue that AI’s potential is first and foremost bottlenecked by "content, not compute." Until agents have access to better data and unified APIs, they will remain "clever summarizers — not true commercial agents."
This new infrastructure is not an incremental upgrade; it's a fundamental necessity. It must provide three core things that the current web lacks.
1. A Trusted, Real-Time Product Catalog
Digital Product Passports (DPPs) are a step forward, but commerce also requires live pricing, stock, and standardized specs—a single source of truth. Scraping the web isn’t sustainable, especially as adversarial AI can game descriptions to mislead buyer agents. A canonical, trusted catalog is essential for reliable decision-making
2. A Domain-Specific Language for Commerce
Agents need more than a generic protocol. Standards like Google’s A2A establish the connection, but commerce requires a shared language with primitives like bids, comparisons, and negotiation terms. A specialized Commerce Protocol makes transactions predictable, auditable, and agent-ready.
3. Secure Identity and Authorization
For users to delegate safely, agents need verifiable identity and scoped permissions. Using DIDs and Verifiable Credentials, a buyer could authorize an agent to spend up to $500 on electronics from verified sellers — creating a secure, auditable chain of trust.
4. An Open, Interoperable Foundation
Above all, these rails must be open. Closed ecosystems tilt the rules in their favor. An open foundation ensures no single gatekeeper controls the market and keeps data and choices aligned with the user.
Building this new commerce layer is precisely why we created the Intents Protocol. It is an open infrastructure designed from the ground up to provide these missing rails. Among other foundational components, The Intents Protocol combines a structured, verifiable product graph with a real-time market and a commerce-specific language that allows sellers to respond directly to user needs. It is the foundational layer that finally allows AI agents to move beyond simply browsing the old web and begin transacting reliably and deterministically in a new, machine-native marketplace.
The Great Inversion: A New Era of Commerce
This new foundation doesn't just make the old process of shopping more efficient. It makes an entirely new, and radically better, process possible. It allows us to completely recreate the flow of commerce itself.
In the old model, you, the buyer with an intent, had to go out and hunt for a seller. You were the one doing the work, searching, and chasing down options. The new infrastructure enables a complete inversion of this dynamic.
Now, you simply express your intent. You tell your agent what you want — "I need a waterproof, breathable jacket for hiking in the Pacific Northwest, my budget is around $300, and I prefer sustainable materials" — and the agent broadcasts this intent to the market.
Instead of you hunting for sellers, the sellers come to you.
This is the great inversion of commerce. That is why we built the Intents Protocol. It is an open foundation for an intent-driven market, now equipped to understand your specific, machine-readable intent, and respond with its best offers. Sellers will compete for the right to fulfill your need, bidding with their most relevant products, best prices, and unique value propositions.
This is the true future of commerce. The power dynamic is completely flipped. Your role shifts from being an active shopper to being a director of your own personal marketplace. You set the vision, and the market responds. This isn't a dream of a slightly better search bar. It's a new era of effortless, intent-driven commerce, and it's being made possible not by layering AI onto the old web, but by building a new foundation for the agent-driven world to come.
References
[1] Okendo. (n.d.). Consumer Decision-Making Process: The 5 Key Stages. Retrieved August 28, 2025, from Okendo.
[2] Think with Google. (n.d.). Shopping research statistics. Retrieved August 28, 2025, from Think with Google.
[3] 1WorldSync. (2024). 73% of Consumers are Moving Their Shopping More Online in 2024. Retrieved August 28, 2025, from 1WorldSync.
[4] BusinessDasher. (2024). 37+ Online Review Statistics: The Ultimate List in 2024. Retrieved August 28, 2025, from BusinessDasher.
[5] SellersCommerce. (2025). Shopping Cart Abandonment Statistics (2025). Retrieved August 28, 2025, from SellersCommerce.
[6] Gartner. (2024, February). Gartner Predicts a 25% Drop in Google Searches by 2026 Due to AI. As cited in SEOCOM.
[7] Allouah, A., Besbes, O., Figueroa, J. D., Kanoria, Y., & Kumar, A. (2025). What Is Your AI Agent Buying? Evaluation, Implications, and Emerging Questions for Agentic E-Commerce. arXiv preprint arXiv:2508.02630.
[8] Frost Brown Todd. (2025). AI Chatbots, Hallucinations, and Legal Risks. Retrieved August 28, 2025, from Frost Brown Todd.
[9] Andreessen Horowitz (a16z). (2025). AI x Commerce. Retrieved August 28, 2025, from https://a16z.com/ai-x-commerce/