Why commerce isn’t ready for AI yet
What happens when platforms don’t agree on how AI commerce should work
AI based commerce will work only after a painful, uneven second digitisation of commerce, and a clear platform direction emerges.
In his response to my piece on Agentic Commerce, Razorpay co-founder and CEO Harshil Mathur actually identified a key fault line that is currently widening in AI and ecommerce, saying:
“Insightful piece and raises the right questions.
But long term agentic commerce can make both sides more efficient, much like how online price discovery beat over-the-counter haggling for consumers.
While I agree with the conclusion, many of the concerns Nikhil flags (boundaries, trust, transparency) are solvable through standards and user control. Because irrespective of the medium, consumers ultimately decide winners based on efficiency and trust.”
This is how founders who have witnessed shifts in the market, whether from offline to online, or desktop to mobile, see transitions: messy but inevitable.
Siddharth Puri, founder of Tyroo messaged me after reading the essay (I’m sharing his views with consent), saying that he went through Google’s Universal Commerce Protocol, and
“They are asking for richer FAQ data for each SKU [Stock Keeping Unit] as part of catalog for better discover ability with bots.”
When an agent is buying, it has to know what it’s buying. The only way that happens is if there’s enough standardisation and products. But, as Sidharth adds about UCP, there are practical constraints. New standards mean new friction to overcome:
“…my challenge is they are still trying to set input standard for brands to bring in data - largely such initiatives fail - other than electronics/consumer durables type categories. Or maybe shoes is an outlier. It’s tough to achieve in fashion, beauty, food.”
This is a reminder that standards only work where products themselves can be cleanly described.
When you operate with scarce resources (including time), what do you optimise for? What do you ship first? What do you ship this quarter? How do you justify the spend towards enabling commerce for a particular protocol?
History is also full of battles over standards, and we are at the beginning of one with AI and agentic commerce.
Welcome to all the new subscribers at Reasoned. If you’re a subscriber, I’d appreciate your responses to a few short questions.
Why this is the Second Digitisation of commerce
There is a great deal of optimism in how Agents will aid shopping. Vogue writes, in How AI Agent Shopping Could Change Fashion Advertising:
“The UCP means US brands on any platform can now use Shopify’s Agentic Storefronts infrastructure via its ‘Agentic plan’ to sell on AI channels, without needing to have a Shopify-hosted online store. This new “open standard” approach is geared towards a future where AI agents from all the different AI chat providers, like ChatGPT, Google AI and Perplexity, can connect with each other and transact with any merchant online. The UCP allows brands to offer customers discount codes, loyalty plans and different billing options.”
Daniel Danker, Head of AI, Walmart, says:
“We’re essentially having their AI agent, Gemini, partner with our AI agent to create a unified shopping journey.” “Imagine it like a window inside of Gemini where our shopping agent kicks in and helps you complete that purchase.”
The implicit assumption behind this optimism is also that agents understand us better: that you wear size XL, prefer peach over navy blue, and prefer a rib knit over a jersey knit. They can make choices on our behalf.
Agents can be smart, but as I discussed in When AI buys or sells for you, personalisation only works when constraints on the demand side can be matched cleanly with constraints on the supply side.
The first digitisation made products visible: listings, photos, descriptions, reviews. This was all designed for human interpretation and decision making. The shift to AI and agentic commerce exposes a need for a second wave of digitisation, because it exposes information gaps that lead to agents making errors. As the Information writes:
“ChatGPT has to interpret information like pricing and in-stock availability that is often ambiguous and spread out across multiple systems… If the agent gathers information incorrectly, it might charge the wrong price or place orders for something that’s out of stock.”
The perfect decision requires perfect information.
Anyone who’s worked on interoperability knows these aren’t reasoning problems, but alignment issues between systems that were never designed to speak to one another. Databases are just structured differently. Merchants just describe their products differently.
What AI agents want from merchants
Scale alone doesn’t solve this. In May 2025, Google wrote about the “Shopping Graph”, saying:
“The Shopping Graph now has more than 50 billion product listings… each with details like reviews, prices, color options, and availability. Every hour more than 2 billion of those product listings are refreshed.”
The second digitisation demands that products should be unambiguous: legible not just to people, but to machines that must reason, compare, substitute, and act, or even invoke in chat. When interpretation and execution are a part of the same system, ambiguity becomes a systems problem. Autonomous agents with agency only increase the scale of errors that already exist.
AI and Agentic commerce depends on the agent doing many “jobs-to-be-done”:
Identifying the product unambiguously: is it a football or a football key chain?
Reading and reconciling attributes referenced differently across systems: How is “Large” in the US different from “Large” in Hong Kong?
Classifying which attributes are mandatory: Which attributes must match exactly (size, compatibility, certification), and which are optional (color preference, finish, brand)?
Understanding substitutability: if a royal blue sweater doesn’t work, would navy blue be acceptable?
Interpreting availability semantics: does “in stock” or availability confidently mean its available with the marketplace, or is there a risk that a third party seller may have not updated their inventory?
Confirming price finality: Are shipping charges or taxes included in the listing price, or are they added at different stages? Is the price displayed for a monthly subscription versus a one time purchase?
Transaction state: when is the transaction deemed to have been completed? Are there intermediate states where the order can still fail?
Understanding post-purchase constraints: Can the product be returned, exchanged, modified, or cancelled?
Determining responsibility for failure: If something goes wrong, who is expected to resolve it by default: the merchant, the platform, or the agent?
Deciding whether the purchase can or should be repeated automatically: Is this a one-off decision, or can it be safely automated in the future without human review? Does the platform allow it?
There are probably other factors, but human based commerce was never built for this level of specificity.
Agents don’t simplify commerce: they force it to be explicit. Understanding and implementing what makes commerce explicit is where the second digitisation of commerce lies, and what makes it difficult.
When platforms move differently, markets are unsettled
Digitisation determines what AI can invoke or autonomously transact with, and how accurate that experience is. Standards become unavoidable infrastructure for market participants, and bring in interoperability and trust. I wrote earlier that the fact that MCPs are not owned by anyone means it is used by everyone. When standards are clear and universally adopted, merchants have one less decision to make.
When they’re not, like with AI and commerce, there’s chaos because each major market participant behaves differently.
Moves from OpenAI and Google to integrate commerce into AI can be seen as a direct attack on Amazon, to add a distribution cost to Amazon and bring it on level with others, and the largest ecommerce company on the planet isn’t taking this threat lightly: it is protecting its turf, and at the same time, leveraging its own agents for customers. From “Amazon, OpenAI and Google Face Off in AI Shopping Wars” in New York Magazine / Intelligencer:
“Customers using Rufus were being shown items from outside stores, sometimes with a button labeled ‘Buy for me,’ which would trigger an Amazon-powered bot to browse the outside merchant’s website, check the item’s price and availability, place the order, and handle the payment process.”
In November last year, Amazon sued Perplexity for unauthorised access and trespass, for agentic commerce. This text from the lawsuit reveals the emerging tension between ecommerce platforms and AI agents:
Since November 19, 2024, Amazon has told Perplexity’s executives on at least five separate occasions that its AI agents may not covertly access the Amazon Store. First Perplexity agreed, then went back on its word.
Next, after Amazon detected the Comet AI agent covertly accessing private customer accounts and told Perplexity to stop, Perplexity claimed that Comet AI was not agentic when its own marketing materials admit otherwise.
Amazon then set up a technological barrier to restrict the Comet AI agent from covertly accessing private customer accounts. In response, Perplexity released a Comet software update specifically designed so that the Comet AI agent could evade that technological barrier.
And when Amazon again addressed Perplexity’s unauthorized conduct with Perplexity on two separate occasions, Perplexity refused to stop. Perplexity’s CEO understood that Perplexity was deliberately flouting Amazon’s rules, but had no legitimate justification for why Perplexity would not act honestly and transparently.
If agentic commerce were merely incremental, Amazon wouldn’t be litigating and improvising at the same time.
*
Google’s Universal Commerce Protocol it has 60 partners: including Shopify, Etsy, Target, Walmart, Best Buy, Flipkart, Macy’s, American Express, Mastercard, Visa, Stripe, among others. AI Mode in Search and the Gemini app will allow shoppers to check out from eligible retailers.
A business agent will allow shoppers to chat with brands, and offer “Direct offers”, as discounts for shoppers who are ready to buy. While the ad units are paid, it isn’t clear whether Google is taking a fee for enabling checkout.
Walmart’s head of AI tells Business Insider:
“We’re essentially having their AI agent, Gemini, partner with our AI agent to create a unified shopping journey… Imagine it like a window inside of Gemini where our shopping agent kicks in and helps you complete that purchase.”
Sounds like the checkout is with Walmart here.
My guess is that Google may not charge for external checkout or limit checkout to its own platform just yet, because it knows the platform game (remember what happened with the Play Store?), and it can play the long game while OpenAI can’t.
*
The Information points out about OpenAI:
“OpenAI has told investors it wants to generate around $110 billion in revenue from nonpaying users by 2030.”
That target creates a clear constraint: new surfaces that can be monetised at scale need to come online quickly. Under revenue pressure, it is rushing monetization, like with advertising. Shopify, which is opening up its merchants to AI and agentic commerce, following OpenAI’s Agentic Commerce Protocol, is apparently saying that OpenAI will charge sellers as much as 4% fee for completed purchases via Instant Checkout. For sellers, this means about 7% if you include card fees (MDR) and taxes. Why will merchants sign up for an expensive sale when free alternatives exist?
This is may not work for OpenAI, because leverage comes from dominance, and it doesn’t have that in ecommerce. It is skipping that part of the platform playbook that says that platforms must wait for user adoption and network effects before triggering monetization.
OpenAI is compressing the onboarding timeline for merchants: it’s almost as if it’s pulling them into an evolving environment instead of inviting them into an ecosystem with settled user behaviour led by discovery, with standards that have stabilised. Standards succeed only when economic incentives precede compliance.
What should a merchant do?
If agentic commerce were settled, the biggest platforms wouldn’t be experimenting in public.
If you’re a merchant, you’re probably confused because there are three options:
First, to take the Amazon approach and protect your independence, even as users shift behaviour to AI.
Second, you know that the switch to AI Mode will hurt your discovery, and Google is forcefully pulling you into AI mode to eventually monetize your presence inside AI mode. You need to be there lest your competitors get there first.
Third, you’re tempted by a potential leverage from exclusive invocation inside ChatGPT, but the fees are just too high.
To compound this issue, the act of digitising your inventory for two separate platforms is both complex, and an additional expense for you, with no clear path to monetization. The data is hard to do, and the standards are fragmented.
AI and Agentic commerce is asking you, as a merchant to invest up-front even though there’s no clarity about scale, and no clear winners.
For larger merchants and aggregators, the choice is clear: they need to do all three. Shopify and Walmart are smart: they’re working with both Google and OpenAI. For them this is hedging. It’s risk management.
For the smaller sellers, I think history tells us that the platforms will come to you. They will incentivise and lure you, instead of frightening or bullying you into coming on board.
OpenAI and Google will eventually invest in teams that will take care of merchant onboarding: they will help you with improving your inventory data, they will offer you information and intelligence on user behaviour, they will explain what kind of user understanding you might get from the platform that helps you curate your products better. They will sell you success stories from the early adopters.
The choice is yours: early adopters often gain exclusivity in their segment and make more money. We’ve seen this with almost all early adopters, whether businesses who went online, drivers who joined Uber, and creators who joined YouTube.
A few other things are predictable:
Success will be patchy before it becomes better.
What begins as an exercise to acquire new customers eventually becomes your only storefront because user behaviour shifts.
Platform fees will emerge: OpenAI’s approach already tells you that.
Nothing about this transition will feel clean while it’s happening. Merchants won’t get a clear signal for when to commit, only stronger signals that delay has a cost. The shift to AI and agentic commerce will be uneven: category by category, platform by platform, until opting out is no longer a real option.
Harshil Mathur is probably right about where this ends. Siddharth Puri is right about what it feels like to operate before it does. Like every transition before it, clarity won’t precede commitment. It will follow it.