5 AI Concepts Every Procurement Professional Should Know

Why does a procurement professional need to understand how AI works?

There are two types of conversation happening about AI in procurement right now.
The first is full of acronyms and technical jargon that few of the people using them can clearly explain. The second treats AI as magic: type a prompt, get an answer, hope for the best.

Neither is much use when you are across the table from a vendor asking for three years of licence fees, or when the CFO wants to know why the contract review tool missed a termination clause.

As procurement professionals we do not need to become data scientists or machine learning engineers. We do need to understand enough about how these systems work to evaluate solutions, challenge vendors, manage risk and tell real business value from a marketing narrative.

Five concepts get you most of the way. Understand them and every conversation you have about AI-driven sourcing, contract management, spend analytics, supplier risk and procurement automation gets sharper.

1. What is a token, and why does it decide what AI costs?

Large language models do not read the way you do. They process text in tokens: a whole word, part of a word, a number, a piece of punctuation. As a rule of thumb, 1,000 tokens is roughly 750 words of English. A 60-page master services agreement is around 40,000 tokens before anyone has asked a question about it.

This matters because every AI interaction is metered in tokens, in and out. Whether the product is a procurement copilot, a contract review assistant or a spend analytics platform, somebody is paying per token, either to a model provider or in compute to run the model themselves, and that cost reaches you one way or another.

For procurement the count climbs fast, because the documents we work with are long: framework agreements, supplier proposals, tender responses, policy documents, category strategies, due diligence reports. Comparing five bids against a 40-page specification is not one question. It is a few hundred thousand tokens.

Understanding tokens explains three things vendors tend to gloss over: why there are limits on document size, why the pricing has a consumption element, and why the "unlimited" plan has an asterisk. Tokens are the raw material the AI consumes to produce anything. Price the raw material before you sign. (The contracting side of this is covered in The LLM Token Trap.)

Ask the vendor: "How is usage metered, and what does it cost to process our typical document set?"

2. What is a context window, and why does AI forget things?

A common belief is that AI "remembers everything you give it". It does not.

Every model has a context window: the amount of information it can hold in view at one time, measured in tokens. Picture a whiteboard. Everything on the board is available for analysis. Once the board is full, something has to be wiped to make room, and whatever was wiped is no longer part of the answer.

For procurement this bites when you are reviewing a lengthy contract, comparing several supplier bids side by side, analysing large RFP responses, checking policy compliance across a set of documents or working through a complex supplier risk report. If the context available is small relative to the material, details get dropped, and the model will not tell you which ones.

It is also one reason output quality varies so much between platforms that claim to do the same thing. Two tools can run on the same underlying model and give different answers, because one fits the whole contract in and the other is quietly working from the first thirty pages.

Ask the vendor: "How much of our document does the model actually see at once, and what happens to the rest?" If the answer is a model name rather than a number, ask again.

3. What is temperature, and when do you actually want creativity?

Not every AI task should be creative. Most procurement tasks should not be.

Temperature is a setting that controls how predictable the output is. Low temperature gives consistent, factual, repeatable answers. High temperature gives varied, more imaginative ones. Same model, same prompt, different behaviour.

Low temperature is what you want for contract clause extraction, spend classification, supplier assessment summaries, policy compliance checks and procurement reporting. If you run the same classification twice and get two different answers, the setting is wrong for the job.

Higher temperature earns its place in market research, category strategy ideation, risk brainstorming, supplier innovation workshops and facilitation content, where you want options rather than the single most probable answer.

The takeaway is simple: the right setting depends on the business objective. You would not want a contract risk assessment to be creative. You would not want a category strategy workshop to be entirely predictable. A product that claims to do both should be able to tell you how it handles the difference.

Ask the vendor: "How do you control consistency for extraction and classification tasks versus ideation tasks?"

4. What is a hallucination, and why is it the risk that matters most?

A hallucination is when an AI system generates information that reads as credible and is wrong.

The problem is not that AI sometimes makes mistakes. Analysts make mistakes. The problem is that the wrong answer arrives with exactly the same confidence and polish as the right one, so nothing in the output tells you which is which.

In procurement the failure modes are specific. An invented supplier certification. A misquoted contract term. A regulation that does not exist. A spend figure that was never in the data. A reference to a source nobody can find. Each one is a sourcing, contracting or compliance decision made on a fabrication.

Two things reduce the risk. The first is grounding: a system that answers from your documents, shows where the answer came from, and says "this is not in the data" when it is not, instead of filling the gap. The second is process: AI output supports a procurement decision, it does not replace procurement judgement. The teams getting the most out of AI use it to accelerate analysis while keeping review, validation and governance exactly where they were.

Trust, but verify. And prefer a tool that makes verifying easy by showing its sources.

Ask the vendor: "What happens when the answer is not in our data? Show me." Then ask it something it cannot know and watch what it does.

5. What is RAG, and how does AI understand your documents?

Many people assume the AI "learns" their company data when they upload it. In most enterprise products it does not, and you should be relieved about that.

Most enterprise AI solutions use a technique called Retrieval-Augmented Generation, or RAG. The idea is simpler than the name. Rather than retraining a model every time you upload a contract, supplier report or policy document, the system stores those documents in a searchable index. When you ask a question, it first retrieves the passages most relevant to that question, then hands them to the model as context and asks it to answer from those.

The model has not memorised your documents. It has become very good at finding the right passage at the right time.

That is what makes it possible to ask questions across a contract repository, search procurement policy conversationally, analyse supplier documentation, review tender responses, query a category knowledge base or run a procurement help desk. It is how Kato, the Procurement Union companion, answers from Theca, the procurement library behind it.

Two implications follow, and both are more important than the acronym.

First, data quality, document structure, taxonomy and knowledge management still matter, arguably more than before. RAG can only retrieve what is there and findable. A contract repository with a third of its files misnamed produces an assistant that is confidently wrong a third of the time.

Second, the retrieval index is where your data lives. Ask where it is hosted and who can query it. If supplier terms, contract values and sourcing strategies are being sent to a public API with every question, your data governance policy has probably already answered the vendor's proposal for you.

Ask the vendor: "Where does our data go, where is it stored, and what is retrieved for each answer?"

Why does this matter for procurement?

AI is becoming embedded across sourcing, contract lifecycle management, supplier risk management, spend analytics and procurement operations. You do not need to know how to build a neural network. With these five concepts you can:

• Ask better questions of technology providers

• Assess AI-enabled procurement solutions on substance rather than demo polish

• Recognise the risks and limitations before they show up in a signed contract

• Set realistic expectations with your stakeholders

• Identify where AI can create real value, and where it cannot yet