.png){kind=logo}
For most of the internet era, the hard part of global sourcing was finding anyone at all. A buyer in Ohio who needed a specific component from a factory in Shenzhen had no practical way to locate it, vet it, or trust it. Digital marketplaces largely solved that. The global B2B e-commerce market is now worth roughly $36 trillion a year, and about 80% of B2B sales interactions already happen through digital channels. Supplier visibility, the problem that defined a generation of commerce platforms, is essentially over. Which raises a more uncomfortable question: if buyers can now see millions of suppliers, why is sourcing still so hard?
Jia (Chris) Lu, who leads AI strategy for North America at one of the world’s largest B2B commerce platforms, has spent years inside that question. She examines it in her book, The Sourcing Intelligence Advantage: How AI Is Reshaping the Way Businesses Find, Evaluate, and Trust Global Suppliers, where she argues that access to suppliers is no longer the primary constraint. The harder problem is understanding what a buyer actually wants and translating that intent into a sourcing decision that holds up. Her view cuts against the prevailing industry instinct, which is to answer every commerce problem by reaching for a bigger AI model.
We spoke with Chris about why more search results have stopped meaning better sourcing, what she calls the Demand-Signal Loop, and why the winners in AI commerce will be decided by how well they understand buyers, not by whose model is largest.
The momentum is real, and it is not hype. AI agents are projected to manage 90% of B2B procurement by 2028, with more than $15 trillion of business spend expected to flow through agent-driven exchanges. That is not a future scenario, it is a roadmap most large platforms are already building against. So the industry is right that AI is about to reshape how buying happens. Where it goes wrong is in what it believes the hard part is.
Almost everyone is competing on the model. Whose AI is smarter, whose has more parameters, whose can reason better. That is the wrong contest. In commerce, the model stopped being the bottleneck a while ago. The bottleneck is the signal you feed it. A brilliant model working from a vague, half-expressed idea of what the buyer wants will confidently produce the wrong answer. A simpler model working from a clear, well-understood intent will beat it every time. We are optimizing the part that is already good enough and ignoring the part that is actually broken.
Because I have watched it play out. The value in AI commerce does not come from the sophistication of the model. It comes from how well the system understands what the buyer is trying to do and converts that into a useful demand signal. Think about what a buyer gives you at the start: often a few words, a rough spec, a budget they are unsure about, and a set of requirements they have not fully articulated even to themselves. The entire game is turning that mess into something precise enough to match against real supplier capability.
Get that translation right and an ordinary model produces excellent results, because it is working from truth. Get it wrong and the best model in the world just hands you a more articulate mistake. This is why I am skeptical when a company tells me their advantage is their AI. That is table stakes now. The durable advantage is in the signal, and the signal comes from understanding the buyer, which is slow, domain-specific work that no model gives you for free.
It is a way of treating commerce as a system that learns, and it has four parts. First, you clarify buyer intent, actively helping the buyer express what they need instead of taking their first vague query at face value. Second, you match that clarified intent against what a supplier can genuinely deliver, which is a different thing from matching the keywords on a listing. Third, and this is the part people skip, you validate the match through the transaction itself: did the deal happen, did it go well, did the supplier deliver what was promised. Fourth, you feed that outcome back in, so the system gets better at the next match.
The loop matters because each stage corrects the others. Transactions are the truth serum. A buyer can click on a supplier and a model can predict a perfect fit, but only the completed transaction tells you whether the match was real. Most systems stop at matching and never close the loop, which is exactly why they plateau. And the pressure to get this right is enormous, because buyers are ready to act on their own. 67% of B2B buyers will now spend $50,000 or more online without ever speaking to a salesperson. Once the human intermediary is gone, the system’s grasp of intent is the only thing standing between the buyer and a bad decision.
Incomplete and ambiguous requirements, which is to say, almost every real buyer. People do not arrive with a clean specification. They arrive with a problem and a rough idea, and they expect the system to help them work out the rest. Most platforms are built for the opposite buyer, the one who already knows exactly what they want and just needs to find it. That buyer is rare. So the system dumps 10,000 results on someone who cannot tell which of them is right, and calls that a successful search.
More results have stopped meaning better sourcing. Often they mean worse, because they shove the entire burden of interpretation back onto a buyer who came to you precisely because they could not do that interpretation themselves. The real work of AI in commerce is the opposite of returning more. It is asking the right clarifying question, catching the requirement the buyer never stated, and narrowing 10,000 options down to the three that actually fit. That is a much harder product and engineering problem than training a bigger model, and it is worth far more.
Constantly. Judging a Cowboy Ventures startup competition, I reviewed pitch after pitch from founders building AI for commerce and adjacent spaces, and the tell was almost always the same. The strong teams could explain, in specific terms, how they capture and sharpen a user’s intent, and how they know when they got it right. The weaker ones led with their model architecture and their benchmark scores, then went quiet the moment you asked how they actually understand what the customer wants.
It is a useful filter, because model capability is a commodity everyone can access now, so it tells you almost nothing about who will win. What separates companies is a proprietary understanding of their users and a system that keeps learning from real outcomes. When I meet a founder who is obsessed with signal quality rather than model size, I pay attention, because they are working on the part of the problem that is actually defensible. The others are polishing something their competitors can buy off the shelf.
It is close to universal. Any system that connects what a person wants with what is available runs on the same logic. A consumer marketplace, a recommendation engine, a search product, a shopping assistant, all of them live or die on how well they understand intent and how honestly they learn from what happens next. B2B sourcing is an unusually demanding version, because the stakes are high, the requirements are complex, and a bad match costs a business real money. But the underlying pattern holds everywhere in commerce.
Which is why I think the industry’s fixation on model supremacy is a strategic mistake across commerce broadly, in sourcing and well beyond it. Everyone now has access to strong models, and nobody can differentiate on that for long. The lasting moats are being built in the parts nobody brags about: the quality of the intent signal, the tightness of the loop between prediction and outcome, and the depth of a system’s understanding of its specific users. Those are earned slowly, and they cannot be downloaded.
Spending their energy on the signal, not the model. In practice that means investing in the unglamorous work of understanding buyers deeply, building the mechanisms that turn a vague request into a precise one, and closing the loop so every transaction makes the next match better. It means resisting the urge to treat a more advanced model as the answer to a problem that is really about interpretation. The model will keep improving on its own, from vendors everyone shares. The understanding will not. You have to build that yourself.
The companies that win the next decade of commerce will be the ones that understood their buyers well enough to ask the right question at the right moment, and had the discipline to keep learning from what actually happened, not the ones with the most impressive AI. Better AI is table stakes. The missing ingredient was never the intelligence of the model. It was the intelligence of the signal we give it, and building that signal has always been our job, not the model’s.
