Terms like LLM, tokens, RAG and MCP now pop up everywhere — in supplier presentations, board papers, news items, even in a category brief for something as ordinary as pallets. You meet them whether or not you work with the IT category. This guide explains fourteen of them in plain language, each with what it means for procurement — plus a one-page cheat sheet.
14 concepts
2–4 min each
Jump straight to any term
Written for procurement & supply chain
Product names & figures verified Aug 2026
The concepts on this page are stable. The product names and market figures — mainly in sections 04, 05, 11, 12 and 14 — change every few months, and are marked so they can be refreshed quickly. If one looks out of date, the shape it describes almost certainly still holds.
Jump to any concept
Search in the sidebar, or pick a card. Nothing depends on anything else — each one stands alone.
Large Language Model. The engine under Copilot, ChatGPT, Claude, Gemini — all of it.
The 30-second version
An LLM is a very large statistical model that has read an enormous amount of text and learned one skill extremely well: given everything so far, predict what comes next. Everything else — summarising a contract, writing a business case, reasoning through a spend file — is that one skill applied at scale.
It is not a database and it does not look answers up. Nothing is “stored” in a way you could go and read. What it has is a set of learned patterns — weights — that let it produce text that fits.
That single fact explains almost every strength and every weakness you will meet:
It is brilliant at shape. It knows what a business case, a contract summary or a negotiation email looks like, so it produces one instantly.
It is unreliable about fact unless you give it the facts. Left alone it will produce something that looks right. That is hallucination, and the fix is RAG.
It has no memory between sessions unless the product adds one.
DiagramOne word at a time — how generation actually works
The supplier’s payment terms are▍
At every step the model produces a probability for every possible next token and picks one. Repeat a few thousand times and you have a business case. Nothing is retrieved; it is all generated.
Procurement lens
When a vendor says “our AI knows your contracts”, ask how. A base model knows nothing about your contracts. Either they feed the contracts in at question time (RAG — good, normal, auditable) or they trained on them (expensive, slow, and a data question you need to answer).
02Tokens and the context window
The two units that explain cost, speed and why it sometimes “forgets”.
The 30-second version
A token is roughly ¾ of a word — the unit the model reads and writes in, and the unit you are billed in. The context window is how many tokens it can hold in front of it at once: its desk, not its long-term memory. Anything off the desk is gone.
ChartWhat fits on the desk — documents vs context window
A 3-page contract~4k
A 40-page MSA~25k
Ten supplier contracts~150k
Typical frontier window~1M
A year of spend data (CSV)~5M+
Illustrative orders of magnitude, not vendor specifications. Frontier windows reached roughly 1M tokens during 2026 and will keep growing — which is exactly why the point here is structural rather than numerical. Your spend file is still bigger than the desk, and accuracy still degrades when the key sentence is buried in a very long input. Retrieval beats pasting at every window size.
Why this matters more than it sounds
Cost tracks tokens, but not one-for-one. Pasting a 200-page document into every question is a real invoice line, repeated. Two things bend the curve: prompt caching (repeated context billed at roughly a tenth of the rate on some providers) and batch processing (around half price for non-urgent work). Ask a vendor whether they use either — if not, you are paying full price for the same tokens over and over. Note the other direction too: agents make many calls per task, so consumption can climb even while per-token prices fall.
“It forgot what I said” is almost always a context problem, not a quality problem.
Bigger windows are not a strategy. Accuracy sags when the important sentence is buried in the middle of a very long input. This has been re-tested on every generation of long-context model since 2023 — including today’s million-token models — and the pattern has not gone away. Retrieving the right 3 pages beats dumping 300.
Procurement lens
Token-based pricing means your spend is driven by usage volume, not headcount — a very different commercial model from the per-seat licences you are used to. Ask for consumption forecasts, per-user caps and an overage mechanism before signing.
03Hallucination — and grounding
The failure mode you must be able to explain to a stakeholder.
The 30-second version
A hallucination is a fluent, confident, plausible answer that is simply not true. It happens because the model is optimising for what fits, not for what is on record. The industrial fix is not “a better model” — it is grounding: forcing every answer to be built from documents you supplied, with citations back to them.
InteractiveSame question, ungrounded vs grounded
Question: “What is the termination notice period with Nordkap Logistics?”
Both answers are equally confident. Only one can be checked. “Can I click through to the source sentence?” is the single best test of whether an AI deployment is production-grade.
04The supplier landscape
Four layers, and most vendors you meet are only in one of them.
The 30-second version
Very few companies actually build frontier models — realistically OpenAI, Anthropic and Google, with a second tier of xAI and Mistral, and a fast-moving open-weight group led by DeepSeek, Alibaba’s Qwen and Moonshot. Everyone else resells or wraps them. Knowing which layer your counterparty sits in tells you what you are really buying, and where the margin is.
Layer 1
Model builders
OpenAI, Anthropic, Google DeepMind, plus xAI (now within SpaceX), Mistral, Meta, and the open-weight labs DeepSeek, Alibaba (Qwen) and Moonshot. They train the models. Capability, safety posture and pricing originate here.
Layer 2
Clouds & distribution
Microsoft Azure, AWS Bedrock, Google Cloud. They host models under enterprise contracts, data residency and existing MSAs. Often the easiest commercial route.
Layer 3
Work products
Microsoft 365 Copilot, Claude Cowork, ChatGPT Work/Enterprise, Gemini in Workspace. What your employees actually touch.
Layer 4
Vertical & point solutions
Contract-lifecycle, sourcing, spend-analytics and CLM vendors with “AI” added. Nearly all of these are Layer 1 models plus a domain interface.
ChartEnterprise LLM API spend share
Anthropic40%
OpenAI27%
Google21%
Everyone else12%
Enterprise LLM API spend share — Menlo Ventures, 2025: The State of Generative AI in the Enterprise, published December 2025. This measures spend on API usage: not seats, not revenue, not users. Note how fast it moves: OpenAI held ~50% in 2023 and 27% by end-2025, while Anthropic went 12% → 24% → 40% and Google 7% → 21%. Total enterprise generative-AI spend tripled to $37bn over the same period. Treat any market-share slide — including this one — as a snapshot with a short shelf life.
Procurement lens
A couple of things to think about. Most AI vendors you meet are building on someone else’s model, so it is worth asking which one sits underneath — and whether you could switch if a better one comes along. Leadership between the model builders has shifted several times in just a few years, which is one reason long exclusivity on a single model deserves some thought. It is also worth knowing that open-weight models have closed much of the gap with the closed ones, and for many procurement tasks they can be a realistic alternative.
05How AI is actually priced
Three commercial models, and they behave very differently on your budget.
Model
How you pay
Cost driver
What to watch
Per seat
Fixed monthly fee per named user
Headcount
Shelfware. Licences bought for a whole department, used by 20%. Insist on adoption reporting and true-down rights.
Per token / consumption
Metered on input + output volume
Usage intensity
Unbounded growth. One badly designed automation can multiply spend overnight. Require caps, alerts and a sandbox budget.
Per outcome / agent
Per task, per document, per resolved ticket
Work volume
Definition risk. What counts as a “task”? Get the counting rule in writing with worked examples.
ChartWhere the money actually goes in a first-year deployment
Licences / tokens~30%
Integration & data work~35%
Change & enablement~25%
Governance & assurance~10%
Indicative shape, not a benchmark — proportions vary widely. The durable lesson: the licence is the small number. If a business case only counts the licence, it is not a business case.
06What an API is
The least glamorous concept here, and the one that unlocks all the others.
The 30-second version
An API is a defined way for one piece of software to ask another piece of software for something. It is the socket. When a vendor says “we integrate with your ERP”, they mean their software calls your ERP’s API. When they say “we’re built on OpenAI” or “we use Claude”, they mean their software calls OpenAI’s or Anthropic’s API.
DiagramA request and a response — that is the whole idea
Everything else — RAG, agents, orchestration — is this same request/response loop, run many times, with extra steps inserted between.
Procurement lens
“Does it have an open API?” is now a standard commercial question, not a technical one. An API is your exit route: it is what lets you extract your data, connect a different tool, or switch vendor without a re-implementation. A closed product is a lock-in you paid for.
07RAG — and why deployments live or die on it
Retrieval-Augmented Generation. If you only take one term away, take this one.
The 30-second version
RAG means: before the model answers, go and fetch the handful of relevant passages from your documents, and put them in front of it. The model then answers from those passages instead of from memory. It is the difference between “an AI that sounds knowledgeable” and “an AI that knows your contracts”.
DiagramThe RAG pipeline, step by step
Two things never leave your control: which documents go in the index, and who is allowed to retrieve from it. That is why RAG is also the answer to “how do we do this without exposing everything to everyone”.
Why it is critical for deployments
Reason 1
It makes answers checkable
Every claim traces back to a source document. Without that you have opinion, and no internal auditor will accept it.
Reason 2
It is current
Update a document and the next answer reflects it. Retraining a model to learn a new contract is absurd; re-indexing takes seconds.
Reason 3
It respects permissions
Retrieval can honour who is allowed to see what, so a user only gets answers from documents they could already open.
Reason 4
It is far cheaper than training
You pay for a short retrieved passage, not for a fine-tuning project — and you can change models without redoing the work.
Where RAG deployments fail — and it is never the model. Bad source documents (five conflicting versions of the same policy), poor chunking (the answer is split across two fragments), retrieval that returns plausible-but-wrong passages, and no permission model. Every one of those is a data-governance problem wearing an AI costume. Ask about them specifically.
Procurement lens
In an RFP, “does it support RAG?” is a weak question — everyone says yes. The strong questions are: where does the index live, who can retrieve from it, does it inherit our existing permissions, can a user click through to the source, and what happens when a document is superseded?
08Prompt vs RAG vs fine-tuning
Three ways to make a general model useful in your world. Vendors conflate them; the costs are wildly different.
What it changes
Effort
Use it when
Cost shape
Prompting
The instructions, each time
Minutes
Almost always — start here
Near zero
RAG
The facts available at answer time
Weeks
The model needs to know your documents, data or policies
Low–medium, mostly data work
Fine-tuning
The model’s own behaviour and style
Months
You need a consistent specialised behaviour at very high volume
High, and repeats with every model upgrade
09How agents actually work
The difference between a model that answers and a system that gets something done.
The 30-second version
An agent is an LLM put inside a loop and given tools. Instead of producing one answer, it works out a plan, takes an action, looks at the result, and decides what to do next — repeating until the job is done or it needs you. A supplier market watch that briefs you every Monday is a typical first one.
DiagramThe agent loop
The loop is the whole idea. A chatbot runs it once. An agent runs it until the goal is met — which is why an agent can produce a finished spreadsheet, and a chatbot can only tell you how to make one.
The three dials that define an agent commercially
Dial 1
Autonomy
Does it ask before each action, or act and report? This is a risk decision, not a technical one — and it should be written into the deployment, not left to the vendor.
Dial 2
Reach
Which systems can it touch, and can it write as well as read? Read-only agents are a very different risk profile from ones that can send an email or change a PO.
Dial 3
Accountability
Who signs off the output, and is there a log of every action it took? “The agent did it” is not an audit answer.
10Tools, connectors and MCP
How an agent reaches out of the chat window and into your systems.
The 30-second version
A tool is any capability you hand the model: search the web, read a folder, query a database, send an email, call your ERP. MCP (Model Context Protocol) is an open standard for describing those tools so any model can use any tool — the USB-C of AI integrations. Since December 2025 it has been governed by the Agentic AI Foundation under the Linux Foundation, with OpenAI, AWS, Google and Microsoft all participating — so it is no longer one vendor’s protocol.
Before tools, an LLM could only produce text. With tools, the same model can read your spend file, look up a supplier’s filings, write a spreadsheet, and put a draft in your outbox. Nothing about the model changed — its reach did.
ToolA single capability the model can invoke, with defined inputs and outputs.
Function callingThe mechanism by which the model requests a tool rather than guessing an answer.
MCPAn open protocol so tools work across vendors instead of being locked to one product.
ConnectorA packaged MCP integration for a specific system — Drive, SharePoint, SAP, a CLM.
Procurement lens
MCP support is a genuine anti-lock-in clause. If a vendor’s integrations are built on an open protocol, the work you paid for survives a change of AI supplier. If they are proprietary, you rebuild. Worth asking explicitly, and worth a price.
The neutral governance is what makes this contractible. A protocol controlled by one supplier is that supplier’s roadmap; a protocol held by a foundation their competitors also sit on is an industry commitment. That is the difference between “must support MCP” as a preference and as a defensible requirement.
11Agentic orchestration tools
The layer that turns “an agent I ran once” into “a process that runs”.
The 30-second version
Orchestration is everything around the agent: what triggers it, which agent handles which step, what happens when one fails, who approves, where the logs go, and how it is monitored. One agent needs none of this. Twenty agents running production processes need all of it.
DiagramThe stack — and where the money and risk sit
Most organisations over-invest in choosing the model at the bottom and under-invest in the orchestration layer, which is where reliability, auditability and actual business value live.
What you will hear in the market
Vendor-native orchestration — Microsoft Copilot Studio (with Agent 365 as the governance control plane), Salesforce Agentforce, ServiceNow, Amazon Bedrock AgentCore, Google Gemini Enterprise. Easiest if you are already deep in that ecosystem; strongest lock-in.
Independent workflow platforms — n8n, Zapier, Make, and developer frameworks like LangGraph. Flexible, model-agnostic, more of your own engineering.
Work-agent products — Claude Cowork, ChatGPT Work. Orchestration is built in and largely invisible; you describe the task and the product plans and executes it. Compared here →
Procurement lens
Ask to see the run history, not the demo. A demo shows the happy path. The run history shows failure rate, retries, human interventions and cost per run — which is the only evidence that the thing works. If a vendor cannot show you that, they do not have production customers.
12Claude Cowork, ChatGPT Work & Copilot
The category is called “work agents”. Here is what distinguishes them, in plain terms.
The 30-second version
All three take the same underlying idea — an LLM plus tools plus a loop — and aim it at your actual work rather than at a chat window. The differences that matter commercially are: where the work happens, what it can reach, and how much it does before checking with you.
Anthropic
Claude Cowork
You point Claude at a folder or a set of connected tools and hand it a multi-step task. It plans, executes and comes back with finished files — a deck, a document, a spreadsheet — that you steer along the way.
OpenAI
ChatGPT Work
An agent inside ChatGPT that gathers context across your connected apps, breaks a goal into steps, and returns finished sheets, slides, docs and small web apps — staying on long projects across many steps rather than answering in one shot.
Microsoft
365 Copilot
AI embedded inside Word, Excel, Outlook and Teams, grounded on your Microsoft 365 tenant. Strongest where the work already lives in Microsoft documents and where IT wants one vendor and one tenant boundary. Governance now has a named answer too: Microsoft Agent 365 is the control plane for agent inventory, permissions and activity across Microsoft and partner agents.
Claude Cowork
ChatGPT Work
Microsoft 365 Copilot
Core shape
Hand over a task, get finished files back
Agent inside ChatGPT that ships finished deliverables
Assistance inside the Office apps you already use
Where work happens
Folders you choose, plus connected tools; execution in an isolated environment
Across your connected apps, web, desktop and mobile
Inside the Microsoft 365 tenant
Best fit
Multi-step file and document work — spend files, contracts, reports
Long-running project work spanning many apps
Everyday assistance at scale for a Microsoft-standard workforce
Main procurement question
Which folders and connectors, and what may it do unsupervised?
Which apps are connected, and what is the consumption model?
Adoption — is anyone actually using the seats you bought?
Why the choice of product matters less than it seems. The skill that pays off — describe the outcome, look at the result, ask for what you actually wanted, iterate — works identically in all three. The products will keep changing. That habit will not.
13Five shapes a real deployment takes
Recognising the pattern tells you what the risks and the costs will be.
Pattern 1
Assistant on the desk
Per-seat licences, people use it as they see fit. Cheapest to start, hardest to measure. Value depends entirely on enablement.
Pattern 2
Grounded Q&A over documents
Classic RAG. “Ask our policies / contracts / specs.” Well-understood, measurable, and stands or falls on document hygiene.
Pattern 3
Document-to-deliverable pipeline
A file lands, something structured comes out. Spend analysis, contract abstraction, invoice exceptions. High, provable ROI. Where most procurement teams should start.
Pattern 4
Monitoring agent
Runs on a cadence, watches something, briefs you. Low risk because it is read-only. A supplier market watch is the classic example.
Pattern 5
Agent inside a process
The agent takes actions in a live business process. Highest value, highest risk, and the only pattern that genuinely needs a control framework before it goes live.
14Risk, data and governance
The five questions that decide whether a deployment is defensible.
01
Is our data trained on?
For enterprise offerings from the major providers the answer is normally no — but it must be contractual, not a web-page promise. Get it in the agreement.
02
Where does it run?
Data residency and sub-processor lists. Matters for GDPR, for sector regulation, and for the security review you will otherwise fail late.
03
Who can retrieve what?
If the AI can read everything, it will eventually surface something to someone who should not see it. Permissions must be inherited, not reinvented.
04
Is there a log?
What was asked, what was retrieved, what the agent did, who approved. Without this you cannot investigate an incident or satisfy an auditor.
05
Who owns the output — and who indemnifies?
IP ownership of generated material, and whether the vendor indemnifies you against third-party IP claims. The major providers all publish one — OpenAI’s Copyright Shield, Microsoft’s Customer Copyright Commitment, Anthropic’s enterprise agreement, Google’s Vertex AI output indemnity — but they are heavily conditional, typically requiring that you left safety filters enabled and did not knowingly infringe. So the question is not “do you indemnify?” but under what conditions, and what is the cap? Smaller wrappers often offer nothing at all.
EU
The EU AI Act
Risk-tiered obligations, with the timetable reset by the Digital Omnibus (in force 27 July 2026): high-risk duties for standalone Annex III systems now apply from 2 December 2027, and for AI embedded in regulated products from 2 August 2028. Transparency duties (Art. 50) and the AI-literacy obligation (Art. 4) were not deferred and apply now. Most procurement uses are limited-risk, but anything touching worker evaluation, recruitment or automated decisions climbs the tiers fast. Ask vendors which tier they place the use case in, why, and what they are doing before December 2027 — a deferral is not a cancellation.
15One-page cheat sheet
Every term on this page, one sentence each. Print it, or keep the tab open in a meeting.
LLMA model that predicts the next word extremely well; the engine under every AI product.
TokenRoughly ¾ of a word — the unit the model reads, writes and bills in.
Context windowHow much it can hold in front of it at once. Its desk, not its memory.
HallucinationA confident answer that is not true. Fixed by grounding, not by hoping.
GroundingForcing every answer to be built from documents you supplied, with citations.
APIThe defined socket one system uses to ask another for something.
RAGFetch the relevant passages from your documents first, then answer from them.
EmbeddingA numeric fingerprint of meaning that lets retrieval find passages by sense, not keyword.
Fine-tuningRetraining a model to change its behaviour. Expensive, rarely the right first answer.
PromptThe instruction you give. The cheapest and most underrated lever you have.
AgentAn LLM in a loop with tools: plan, act, observe, repeat until done.
Tool / function callingA capability handed to the model so it can act instead of only describing.
MCPOpen standard for connecting tools to any model. Your anti-lock-in clause.
OrchestrationTriggers, routing, approvals, retries, logging — everything around the agent.
Multi-agentSeveral specialised agents coordinating. Powerful; usually premature.
Human in the loopA required approval point before an action takes effect.
GuardrailsRules constraining what the system may say or do, enforced outside the model.
InferenceThe act of running the model to produce an answer. What you are billed for.
Foundation modelThe general-purpose base model, before anyone adapts it to a use case.
Data residencyWhich country your data is processed and stored in. A contract term, not a setting.
