A company decides to use AI in procurement. A vendor gets selected. A pilot is announced. Procurement people attend a kickoff. Six months later, the pilot is quietly discontinued, or limps along with three users and no clear outcome.
This is not a rare story. It is the default outcome for most organisations that try to introduce AI without understanding what actually makes it work in a procurement context.
The technology is usually fine. ChatGPT, Claude, Copilot — they can all read documents, categorise spend, and draft communications. The problem is consistently elsewhere.
1. They started with the tool, not the process
The most common failure mode: someone gets excited about a tool, buys a licence, and starts looking for problems to fit it to. The order of operations is wrong from the start.
In a well-run procurement function, you already know which processes hurt. The qualification questionnaire that takes three people two days every time a new supplier is onboarded. The monthly spend report that requires someone to spend four hours slicing Excel. The quote comparison that sits in someone's inbox for a week because they are the only one who knows how to read the supplier's pricing format.
AI on a broken process does not fix the process. It just makes the brokenness happen faster and at scale.
The right starting question is not "what can AI do?" It is "which of our processes is slow, error-prone, and done more than 10 times a year?" That answer usually writes the project brief by itself.
2. Procurement was not in the room when it was designed
Procurement AI projects are frequently initiated by IT, digital transformation teams, or management consultants. These people design something technically correct. Then they hand it to procurement to use.
The tool does not match how procurement actually works. The categories do not reflect the company's spend taxonomy. The email templates do not match the tone buyers use with long-term suppliers. The approval flow assumes every purchase has a single approver — but in practice, direct materials go through three departments and a supply chain lead.
The people who actually do the work were not consulted. So the tool gets used reluctantly, then not at all.
3. The data was dirty
Supplier master data is, in most companies, a disaster. Duplicate entries. Incomplete addresses. One supplier registered under three different names because three buyers at three different sites set them up separately ten years ago. GL account coding that is consistent within one plant but means something different in the next one.
AI models do not fix dirty data. They work with what they get. If you ask an AI to categorise your spend and your supplier names are a mess, you will get a categorisation that is 80% plausible but quietly wrong in all the edge cases that matter most.
Before you run any AI on procurement data, you need a week of unglamorous data cleaning. Most organisations skip this step because it does not feel like progress. It is the most important step.
4. Nobody defined what success looked like
A pilot runs for three months. At the end, someone asks: "Is it working?" Nobody knows. There was no baseline. Nobody measured how long the process took before the pilot. Nobody tracked whether accuracy improved. The only evidence of progress is that the vendor sent a report showing the tool was used 47 times.
This is a standard problem in procurement improvement projects, not just AI ones. But AI projects are especially vulnerable because the technology feels impressive enough to obscure the absence of a result.
Define the metrics before the pilot starts. Literally write them down: this process currently takes X hours. We aim to reduce it to Y hours. We will measure this by tracking Z for 90 days. Then actually track Z.
5. They picked the hardest problem first
Contract review is interesting. End-to-end supplier risk assessment sounds impressive at a board presentation. Fully automated sourcing decisions make for good vendor marketing material.
They are also the most technically complex, legally sensitive, and organisationally resistant projects you can run. When they do not go smoothly — and they usually do not — the whole AI programme gets labelled as "not ready."
Start somewhere smaller and more boring. Automate the supplier qualification questionnaire processing. Automate the spend coding for one category. Automate the reminder emails to suppliers who have not confirmed order receipt.
These are not impressive demos. They save 200 hours a year per person. They build credibility. They give procurement the confidence to take on harder problems next year.
The pattern underneath
All five reasons come back to the same thing: AI was treated as a technology decision, not a process improvement decision.
The companies that get this right do not have better AI tools. They have a clearer understanding of which processes need improving, cleaner data going into the model, and someone in procurement who is willing to spend time defining what "better" actually means.
That last one is the scarcest resource of all.