When your user sends an agent instead
The Platform Shift
What happens to platform value creation when the person using it isn’t a person anymore — and what that might mean for the products we build?
April 2026
16 min read
Something interesting is happening that I don’t think we’ve fully processed yet.
People are still ordering food, booking rides, buying things online — probably more than ever. But increasingly, they’re not the ones doing the ordering. An AI assistant is doing it for them. Not in a sci-fi, hypothetical way. In a “this shipped last quarter” way.
In January 2026, Swiggy integrated with Model Context Protocol, which means you can now order groceries, food, or a restaurant table through ChatGPT or Claude without ever opening the Swiggy app. A month later, Uber Eats launched an AI assistant that builds your grocery cart from a photo of a handwritten list. Google announced the Universal Commerce Protocol at NRF 2026, an open standard that lets AI agents handle the full shopping journey, from discovery to checkout to returns with endorsements from Shopify, Walmart, Target, Visa, Mastercard, Flipkart, and two dozen others.
These developments are raising a question I find genuinely hard to think through: what happens to a platform’s value proposition when the primary user of its interface is no longer human?
I’ve been pulling on this thread for a few weeks now, and this is an attempt to lay out what I’ve found, the behavioral signals, the business model implications, and the genuinely open questions to which I don’t think there are any clean answers yet.
First, a useful distinction: the user as actor vs. the user as principal
In every platform interaction today, the user is the actor. They open the app, they browse, they tap, they decide, they pay. The entire product experience — the UX, the recommendation engine, the discovery feed, the promoted listings is designed around a human making decisions on a screen.
When an AI agent enters the picture, the user becomes something different. They’re still the customer, it’s still their money, their preferences, their intent. But they’ve delegated the execution to an agent. The user is now a principal in the economic sense: someone who sets the objective and lets an agent act on their behalf.
This sounds like a minor semantic distinction, but I think it changes almost everything about how platforms create and capture value. Because most of what platforms have built over the last decade the browse experiences, the recommendation carousels, the surge pricing UX, the impulse-purchase nudges, these were designed for actors. They don’t work on agents.
An agent doesn’t scroll. It queries. It doesn’t feel the emotional pull of a “Trending Near You” carousel. It doesn’t impulse-add dessert because the photo looked good. It doesn’t experience FOMO during a flash sale countdown timer.
Which raises an uncomfortable possibility: a lot of what platforms call “value creation” might actually be “friction monetization” and agents are about to remove the friction.
The behavioral signal that caught my attention
I started paying attention to this when I came across the zero-click data, which I think is the clearest early signal of how agent-mediated behavior actually plays out at scale.
Zero-click searches — where someone searches for something and gets their answer without clicking any link — aren’t exactly an “agent” use case. But they’re the closest large-scale behavioral proxy we have for what happens when an AI layer sits between the user and the platform. And the numbers are striking:
56% → 69% Zero-click search share between May 2024 and May 2025, after Google’s AI Overviews rolled out. A 13 percentage point shift in one year.
−61% Drop in organic click-through rate for queries where AI Overviews appear, based on Seer Interactive’s analysis of 25 million impressions across 42 organizations.
−20% (US) Decline in Google searches per user year-over-year. But only −2% to −3% in the EU/UK, where AI features have rolled out more slowly. The gap is drastic.
80% is the share of consumers who rely on zero-click results for at least 40% of their searches, per Bain & Company. This is already mainstream behavior, not an edge case.
What’s happening in search is a preview of what’s going to happen — is probably already beginning to happen — in commerce, travel, food delivery, and every other platform where the user’s primary need is transactional.
The pattern is consistent: when an AI layer resolves intent before the user reaches the platform, engagement metrics decline but the underlying need is still being met. The user got their answer. They just didn’t click through to get it.
Replace “clicked a link” with “opened the app” and you’re looking at the future of every transactional platform.
What I think actually changes (and what doesn’t)
It’s tempting to paint this as “apps are dead, agents are everything.” I don’t think that’s right. What seems more accurate and more interesting is that certain types of platform interactions are vulnerable to agent mediation, while others are basically immune. The split has to do with why the user is there in the first place.
Interactions that are vulnerable: utility tasks
Ordering groceries. Booking a cab. Comparing flight prices. Reordering contact lenses. Renewing insurance. Finding the cheapest wireless earbuds with good bass.
These are tasks people do not because they want to, but because they need to. Given the option to delegate them entirely, most people would. An agent that handles your weekly grocery run by checking prices across platforms, applying the best coupons, and optimizing for delivery time? That’s not a futuristic convenience. That’s a description of what people already wish their apps would do — minus the 15 minutes of scrolling.
For platforms built primarily around utility tasks, the risk is real: as agents get better, the time users spend in-app declines. And with it, the ad impressions, the impulse purchases, the promoted listing views, and the behavioral data that feeds the recommendation engine.
Interactions that are immune: experience and identity
This is where it gets more interesting. Not all platform time is utility time.
70%+ Content consumed on TikTok, YouTube, and Instagram via algorithmic feeds rather than active search. Users aren’t looking for something specific. They’re looking for surprise, connection, cultural participation.
54% More time spent by Gen Z on social platforms and user-generated content discovery than average consumers. The generation most fluent in AI is also the one most deeply engaged in non-delegable social experiences.
26% Of US adults already using AI for product discovery in 2025. But social discovery learning about a product because a creator you follow used it remains entirely human-driven.
Nobody asks their AI agent to scroll TikTok for them. Nobody delegates the experience of watching a cooking reel, or discovering a new artist, or laughing at a meme. These are things people do because doing them is the point.
This suggests a split I keep coming back to, which I’ll call the delegation stack vs. the experience stack. The delegation stack is everything you’d happily hand off to an agent: order groceries, book a ride, find the best deal, pay the bills. The experience stack is everything you do because you want to be doing it: browse TikTok, watch YouTube, message friends, explore a new neighborhood on foot.
Every platform lives somewhere on this spectrum. And where you sit determines how much the agent transition threatens your core business.
The question I keep asking myself: is there a version of food delivery or ride-hailing that moves up the spectrum from delegation to experience?
The discovery question — which I think is genuinely unresolved
There’s a line of thinking — recently articulated in an HBR piece on AI and platform revenue — that AI agents create “zero-click commerce” and thereby undermine the discovery-driven business model. The argument is straightforward: if the agent resolves the user’s intent directly, the user never browses the platform, and all the value that gets created during browsing (ad impressions, impulse purchases, cross-selling) disappears.
I think this is partially right and partially missing something important.
It’s right that intent-driven discovery is migrating to agents. If you know what you want — “best noise-cancelling headphones under $200” an agent is objectively better at resolving that than 45 minutes of browsing review sites. The Bain data showing 80% of consumers relying on zero-click results for informational queries confirms this pattern is already mainstream.
But there’s another kind of discovery that doesn’t start with intent. It starts with curiosity, boredom, or social context. You open Instagram and see a friend wearing a jacket you like. You didn’t know you wanted a jacket. You didn’t ask an agent to find you one. The discovery happened because you were there, doing something else.
This kind of serendipitous, identity-driven discovery is the one I don’t think agents can replicate — and it might actually become more valuable as utilitarian discovery moves to agents. If the only people still browsing your platform are the ones who are there for the experience rather than the errand, they might be a more engaged, higher-value audience than the mixed pool of browsers you have today.
I genuinely don’t know how this plays out. But the early behavioral data suggests something nuanced: people aren’t discovering less overall. They’re discovering differently depending on the category.
A signal worth watching
AI-referred traffic to Shopify merchants grew 7× between January 2025 and early 2026. But here’s what’s interesting: that traffic converts at dramatically higher rates than traditional organic search traffic. One analysis valued AI-referred visitors at 4.4× higher economic value than traditional organic visitors. Fewer visitors, but each one arrives with clear, pre-resolved intent.
This suggests that agent-mediated commerce might actually be better for platforms on a per-transaction basis, even though it looks worse on an engagement-metrics basis. The dashboard says engagement is falling. The P&L might tell a different story.
What happens to the time agents free up?
This is the part I find most interesting, and the part where I think most analysis of the agent era falls short. Once agent starts focussing on platform, where does that freed-up user time goes?
Think about it concretely. If an agent saves you 15 minutes on grocery ordering, 5 minutes on booking a cab, 20 minutes on comparing products — that’s 40 minutes a day of reclaimed time that used to be spent inside platform apps. What does a person do with those 40 minutes?
The data gives us some early clues, and they consistently point upward on what you might loosely call Maslow’s hierarchy of digital needs:
People are spending more time in creative and social spaces. TikTok, YouTube, and Instagram are seeing the most resilient engagement precisely because their value proposition is experiential, not transactional. Over 70% of content consumed on these platforms comes through algorithmic feeds — passive, serendipitous, identity-driven consumption that no one delegates.
People are spending time with AI itself — but as a companion, not a servant. ChatGPT hit 800+ million weekly active users by mid-2025, with a DAU/MAU ratio of 36%. Many of those sessions aren’t task delegation. They’re conversations, brainstorming, learning, creative collaboration. AI isn’t just taking tasks off people’s plates, it’s becoming something people choose to spend time with. That’s a fundamentally different kind of engagement.
There are early signs that community-driven platforms are gaining share. Reddit, Discord servers, niche communities — spaces where the value is belonging, not efficiency, seem resistant to the patterns hitting transactional platforms. When consumer enthusiasm for AI-generated content dropped from 60% in 2023 to 26% in 2025 (audiences increasingly dismiss it as “AI slop”), the counter-signal was clear: people are hungry for authenticity, human voice, and real connection. Platforms that offer that are positioned well.
The implication, if this directional pattern holds, is that the agent era doesn’t reduce total digital engagement. It reshuffles it. Time migrates away from errand-running apps and toward experience, community, and creativity. That’s a big deal for where the next wave of platform value gets created.
How I think about the business model shift
I’m not going to pretend to have a definitive framework for how platform business models adapt. We’re too early. But there are a few directions I find compelling enough to think through.
From monetizing attention to monetizing outcomes
If you’re a marketplace or delivery platform, your current ad model works like this: a restaurant pays for a featured listing, which gets impressions from users who are browsing, and some fraction of those impressions convert to orders. You’re selling probabilistic attention.
In an agent-mediated world, there’s no browsing, so there are no impressions. But the restaurant’s underlying need — more orders doesn’t go away. What changes is the monetization mechanism. Instead of selling eyeballs, you sell completed transactions. A restaurant pays when an agent routes an order to them, not when a human happens to see their listing.
Google’s already testing this with “Direct Offers” in AI Mode exclusive discounts presented to shoppers who are actively ready to buy, inside the AI interface. It’s outcome-based advertising rather than impression-based advertising. The conversion rate is structurally higher because the agent has already matched intent to product before the offer is shown.
I think this shift from attention to outcomes is probably the single most important business model evolution for transactional platforms. The revenue might actually be more predictable and higher-quality but it requires a completely different commercial infrastructure than the ad-sales model most platforms have built.
The platform as API vs. the platform as experience
There’s a strategic fork emerging that I think every platform team will eventually have to navigate. On one path, you lean into being the best API — the most reliable, data-rich, agent-friendly infrastructure for completing transactions. Your “product” is no longer the consumer app; it’s the MCP server, the structured product catalog, the real-time inventory feed. You compete on data quality, fulfillment reliability, and API uptime. You might never interact with the end user directly again, and that’s fine — you’re the plumbing.
On the other path, you lean into being the best experience: the place humans choose to spend time because it’s interesting, social, or identity-affirming. You build the cooking content feed, the community reviews, the neighborhood guides, the creator ecosystem. Agents can’t replace this because the experience is the product.
The most interesting (and difficult) path is probably both: be the API for delegated transactions and the experience for direct engagement. But that requires running two fundamentally different product strategies simultaneously, which is genuinely hard.
The preference graph as a new kind of moat
Here’s something I keep thinking about. When an agent orders food on your behalf, it needs to know what you like. Not just what you’ve ordered before (that’s transaction history, which any platform has), but deeper context: you’re mildly lactose intolerant, you prefer medium spice, you’ve been trying to eat more protein lately, you don’t like cilantro, you tend to order heavier meals on weekends.
Whichever platform has the richest, most nuanced understanding of a user’s preferences becomes indispensable to every agent that tries to act on that user’s behalf. The agent can resolve the logistics — finding the restaurant, comparing prices, applying coupons. But it needs the preference context to resolve well.
This reframes what a platform’s data moat actually is. In the pre-agent era, the moat was behavioral data: what users browse, what they click on, how long they linger. In the agent era, that behavioral data loses value (agents don’t browse). What gains value is preference data — a structured understanding of what the user actually likes, across contexts. That’s harder to build than a recommendation engine, and harder for competitors to replicate.
What I’m watching, and what I still don’t know
I want to be honest about the limits of this analysis. We’re very early in the agent-mediated commerce transition. MCP integrations are live, but the vast majority of orders across food, rides, groceries, e-commerce are still placed by humans inside apps. We’re looking at early signals, not established patterns.
A few things I’m specifically watching to see how this unfolds:
Will users actually trust agents to transact on their behalf? There’s a gap between “I use ChatGPT to research products” and “I let an AI spend my money unsupervised.” The OpenAI Instant Checkout experiment in ChatGPT was quietly shut down by early 2026, partly because brands and users weren’t comfortable with purchases completing entirely inside the AI. Trust may be the rate limiter, not technology.
What happens to impulse purchasing? Quick commerce economics depend heavily on AOV being 40-60% higher than the user’s “intended” basket — the extras you add because you saw them while browsing. If agent-mediated orders reflect only the intended basket, does the unit economics model still work? Or do platforms find new ways to introduce serendipity into agent flows?
Do agents commoditize platforms or consolidate them? There are two plausible outcomes. In one, agents comparison-shop across platforms in real time, making switching costs near zero and turning every marketplace into a commodity API. In the other, the platforms that become most agent-friendly (best APIs, best data, best reliability) attract disproportionate agent traffic and get even stronger. Early search data suggests a “rich get richer” pattern — AI traffic concentrates among established platforms rather than flowing to smaller players. But it’s early.
Will new protocols (UCP, MCP, A2A) become the TCP/IP of commerce? Google’s Universal Commerce Protocol is designed as an open standard, like HTTP for shopping. If it achieves real adoption, it could structurally lower the barrier for agents to transact across any merchant — which is amazing for consumers and agents, but potentially threatening for platforms whose moat was ecosystem lock-in. Or it could become the next OpenSocial — technically sound, adopted by everyone, used by no one.
What do people do with themselves? This is the big, philosophical one. If agents handle the logistical layer of daily life — food, transport, shopping, scheduling and humans spend their freed-up time in creative, social, and experiential spaces, that’s a meaningful shift in what people value and pay for. The platforms of the agent era might not be marketplaces at all. They might be community spaces, creative tools, or shared experiences. I have no idea what those look like yet. But it feels like where the energy is heading.
A few closing thoughts
I started looking into this because of a specific tension I noticed at platform companies: power users are asking for AI-native features that reduce friction, but reducing friction reduces the engagement metrics the entire business is built on. That tension is real, and I don’t think there’s a clean resolution.
But what I’ve come to believe, after pulling on this for a while, is that the tension is actually a symptom of something deeper. Platforms have historically conflated two different kinds of value: the value of getting things done (utility) and the value of being somewhere (experience). They bundled both into a single app and monetized the combined attention. Agents are unbundling that.
The utility value — finding, comparing, transacting will increasingly flow through agents and APIs. The experience value — discovering, belonging, creating, exploring will remain with human-facing interfaces. The platforms that thrive will be the ones that figure out which kind of value they’re actually best at creating, and rebuild around that honestly.
I don’t know exactly how the transition plays out, or how fast. Nobody does. But the behavioral signals are directional, the infrastructure is live, and the user expectations are shifting. The interesting work is in figuring out what to build for the world that’s emerging not defending the one that’s changing.
I’ll keep thinking about this. If you’re working on it too, I’d love to hear what you’re seeing.
Thanks for reading the Working Agents.
If this sparked something, share it with someone who’s thinking about these questions too.