Somebody is selling you an AI agent that finds suppliers, negotiates prices, and manages your orders. The demo is impressive. The pitch deck opens with a statistic about how most companies are failing at AI and you are about to be left behind.
Here is the short version. AI moves the paperwork in sourcing, and it is genuinely good at that. It does not move the risk, because the parts of this job that go wrong are the parts that need a person in a room, a registry lookup, and somebody whose name is on the customs entry.
And the statistic in that deck is almost certainly being misquoted, which is a useful thing to know before you buy anything.
This is written for founders and brands importing a physical product, not for enterprises running procurement platforms. If your immediate problem is a supplier who agrees to everything and ships something else, that is how to communicate with Chinese suppliers, and it is a people problem rather than a tooling one.
The statistic everyone quotes at you
You have seen it. Ninety-five percent of AI pilots fail, usually attributed to MIT.
It comes from a July 2025 report called The GenAI Divide: State of AI in Business 2025, produced by Project NANDA, which stands for Networked Agents and Decentralized AI. NANDA is a spinoff of the MIT Media Lab rather than a formal MIT research programme, and the report presents itself as preliminary findings rather than a peer-reviewed paper.
Three things about it are worth knowing before you repeat the number.
The sample is small and thinly described. The report alludes to 52 interviews, with little detail about who was interviewed or how the data was gathered, and the quotes are anonymous.
The authors are not neutral. NANDA is backed by major AI companies and the report advocates for agent infrastructure, which is the thing its authors build. That does not make the findings wrong, and it is a conflict a reader should be told about.
The number is barely in the report. The 95% figure appears in essentially one sentence with no supporting method. The nearest supporting figure is a 5% rate for custom enterprise AI tools being successfully implemented, in one section, for one category.
Kevin Werbach at Wharton made the point most clearly. “Successfully implemented” is a far narrower bar than “failed,” and a project that did not clear it has not necessarily returned nothing. A 5% success rate inside one subset of one category cannot carry a 95% failure claim across everything.
The honest version
Plenty of AI projects do stall, and that part is real. Teams buy tools, run a pilot, and quietly stop using them, which is the normal fate of most new software rather than something unique to AI.
But the specific number does not measure that, and the study it is attributed to does not say what the deck says it says. Forbes, Fortune, and Inc. all ran the headline, which is how a thinly supported figure becomes common knowledge in about three weeks.
Here is why that matters commercially. If a vendor opens with that statistic, they did not check it, which tells you how they will handle the numbers in your supply chain.
What AI genuinely does well in sourcing today
Quite a lot, and all of it is upstream of the decisions rather than instead of them.
Language and tone. Translating a technical requirement into clear Mandarin, and translating a supplier’s careful reply back into what it probably means, is work AI does well and fast. It will not catch the soft no, but it will stop you sending something ambiguous.
Reading long documents. Supplier catalogues, test reports, certificates, and contracts can be summarized and compared in seconds. Use it to find the clause, then read the clause yourself.
First drafts of specifications. Give it your product and let it produce a specification skeleton with the fields you would otherwise forget, including tolerances, packaging, and marking. You fill in the values.
Narrowing a long list. Turning 200 supplier results into 20 worth looking at is tedious pattern work, which is exactly what this technology is for.
Candidate classifications. Asking for plausible HS codes to then confirm with your broker is a reasonable use. Treating the answer as the classification is not.
Spotting what does not match. Comparing a quote, an invoice, and a packing list for inconsistencies is dull and AI is tireless.
Notice what all six have in common. The output is a draft, and a person checks it before it becomes a decision.
The four things that still need a body
These are the points where sourcing actually fails, and none of them are solved by a better model.
Confirming the company is real. A supplier either appears in the national enterprise registry with an active status and a matching legal representative, or it does not. That is a lookup against a government database, and the certificates behind it have their own registries, which we covered in how to verify a supplier’s certificates.
Seeing the line run. A live, unedited walk through the factory with the machines going tells you things no document will, and somebody has to ask the follow-up question in the moment.
Hearing the soft no. “It might be a little difficult” is usually a refusal. Reading that correctly requires knowing the person, the context, and what they said last time.
Carrying the liability. This is the one that ends the argument, and it deserves its own section.
Why the last one is not a technology problem
Because responsibility is assigned to a legal person, and software is not one.
US law puts the obligation to use reasonable care on the importer of record for the value, classification, and origin declared on an entry. That obligation does not transfer to the tool that suggested the code. When customs has a question, it writes to a party, and that party answers with documents and a deadline, as we covered in customs holds.
The same logic runs through the rest of it. When a vessel declares general average, somebody has to post security before the container moves, and an agent cannot sign a bond. When a lot fails inspection, somebody decides whether to sort, rework, discount, or reject, and that decision costs money either way.
So the useful question is not what AI can do. It is which parts of your sourcing carry consequences you cannot delegate, and who is standing behind those.
What to automate first if you are small
Start where the work is repetitive and the mistakes are cheap.
Put your supplier correspondence through a drafting step before you send it, so your requirements go out specific rather than friendly and vague. Summarize the long documents you currently skim. Build one comparison sheet for quotes so the numbers sit side by side instead of in six email threads.
Then stop. Do not automate the decision to send money, the approval of a sample, or the answer to a customs enquiry. Those three are where the losses live, and they are worth your own attention.
Four questions to ask before you buy an AI sourcing tool
Ask them in writing, and notice which ones produce a straight answer.
“Where does your supplier data come from, and when was it last refreshed?” A tool recommending suppliers is drawing on something. If that something is a scrape of directory listings from two years ago, you are being sold a stale list with a chat interface on it.
“Does it verify against the official registry, or against your own database?” These are completely different claims. Checking a company’s status in a national enterprise registry is verification. Matching it against an internal record of companies someone once listed is not.
“When it gets a classification or a supplier wrong, who is accountable?” Read the answer carefully, because it is usually a terms-of-service paragraph saying nobody. That is not disqualifying, and it does tell you where the output sits: a suggestion, not advice you can rely on.
“Can it show the source for any given answer?” A tool that cites the registry entry, the certificate number, or the tariff line is usable, because you can go and check. One that produces confident prose with no trail is a liability dressed as a shortcut.
A good tool answers all four plainly, because a good tool is not pretending to be more than it is. The useful ones in this category tend to describe themselves as drafting and research aids, and the ones promising an autonomous agent that handles your sourcing are describing a product that does not exist yet.
What this means when you choose a partner
The pitch you should be suspicious of is the one where the agent does everything.
A middleman with an AI layer is still a middleman. The tooling makes the emails faster without changing who checks the registry, who stands on the factory floor, and whose name goes on the entry.
We have launched 18 brands and verified 800+ suppliers, with our own team on the ground in China for 15+ years. We use the tools where they help and we do the verification ourselves, because that is the part that has consequences. Our fee comes from you and we take zero commissions from suppliers, so we have no reason to let a tool wave something through.
What to keep from this
Four things, if you remember nothing else.
The statistic about 95% of AI pilots failing comes from a preliminary, non-peer-reviewed report with a small sample and commercial backing, and the number is not supported by what the report measured. AI is genuinely useful for translation, document reading, first drafts, shortlisting, and spotting inconsistencies, and in every case the output is a draft that a person checks. The things that actually go wrong in sourcing need a registry lookup, a factory visit, a read of what the supplier meant, and somebody who carries the liability. And nothing you can buy changes who answers when customs writes.
Let us handle the part that cannot be automated
We have launched 18 brands and verified 800+ suppliers. We check the registry, walk the line on camera, confirm the specification is understood rather than agreed to, and stand behind the paperwork when something is questioned.
Tell us what you are making, and we will tell you straight which parts of your process a tool can take and which ones it cannot.
Frequently asked questions
Is it true that 95% of AI projects fail?
Not as the claim is usually made. The figure traces to a July 2025 report from Project NANDA, a spinoff of the MIT Media Lab, which described itself as preliminary findings, alluded to around 52 interviews, and was not peer reviewed. The 95% appears in roughly one sentence, and the nearest supporting number is a 5% successful-implementation rate for one category of custom enterprise tools. Academic critics have pointed out that “not successfully implemented” is much narrower than “failed,” so the headline overstates what was measured.
Can AI find and vet suppliers for me?
It can find and shortlist them, and it cannot vet them. Narrowing a long list of candidates is exactly the kind of pattern work AI handles well. Verification is different: it means checking the company in the national enterprise registry, confirming certificates in the registries that issued them, and seeing the production line run. Those are lookups against official records and a physical visit, not something a model can infer.
What should I actually use AI for in importing?
Translating and tightening your supplier correspondence, summarizing long documents like test reports and contracts, drafting a specification skeleton so you do not forget tolerances and packaging, shortlisting suppliers, suggesting candidate HS codes for your broker to confirm, and comparing quotes and invoices for inconsistencies. Treat every output as a draft. Keep the decision to pay, the approval of a sample, and any response to customs in human hands.
