AI Jargon, Translated: 15 Terms Every UK Business Owner Should Actually Understand

You’ve heard these words a hundred times. AI. LLM. Agentic. RAG. They come at you in vendor pitches, on LinkedIn, at trade shows, in that newsletter you keep meaning to unsubscribe from.

Everyone uses them. Nobody explains them.

So you nod along, don’t ask, and make buying decisions based on language you don’t fully understand, which is exactly what the people using that language are counting on.

That stops now. Fifteen AI terms in plain English. One real example each. One line on why it matters. Grouped by how you’ll actually encounter them.

---

Part 1: The Basics

Understand these four and everything else clicks into place.

AI (Artificial Intelligence)

Software that performs tasks normally requiring human thinking: understanding language, spotting patterns, making predictions. An umbrella term covering everything from your spam filter to ChatGPT. When a vendor says “we use AI,” that’s as specific as saying “we use vehicles.” A bicycle and a Tesla are both vehicles.

Example: Xero uses AI to auto-categorise bank transactions and predict cash flow for over a million UK businesses.

Why it matters: “AI-powered” tells you nothing. Ask: what does the AI specifically do, what data does it use, and how often is it wrong?

LLM (Large Language Model)

A type of AI trained on enormous amounts of text that can read, write, summarise, and hold conversations. ChatGPT, Claude, and Gemini are all LLMs. Think of an extremely well-read new hire who’s consumed every business book ever written but has never worked at your company. Brilliant first drafts, but gets your specific pricing and processes wrong unless you feed it your own information.

Example: Microsoft embedded GPT-4 into Copilot across Microsoft 365. If your business uses it, you’re already being offered an LLM.

Why it matters: LLMs work from general training, not your data, unless specifically set up. Out of the box: clever but generic. Connected to your business data: genuinely useful.

Prompt

The instruction you give an AI tool. “Prompt engineering” is the pretentious term for writing better instructions. It’s not engineering. It’s writing a decent brief. “Make me a logo” gets clip art. “Logo for a Kent-based engineering consultancy, clean, modern, navy and silver, works at small sizes” gets something usable. Same freelancer; the brief made the difference.

Example: Anthropic’s Claude lets businesses set system prompts: standing instructions defining the AI’s role, tone, and boundaries for every conversation.

Why it matters: The businesses getting the most from AI aren’t the most technical. They’re the ones with the clearest communicators.

Foundation Model

A large, general-purpose AI model that serves as the base layer for more specific tools. GPT-4, Claude, Gemini, and Llama are all foundation models. Think of a fully kitted commercial kitchen. It can make anything, but it’s not yet a restaurant. You still need your menu, chefs, and service.

Example: Amazon Bedrock lets businesses access multiple foundation models (Claude, Llama, Mistral) through one platform.

Why it matters: “Powered by GPT-4” tells you the engine, not the car. What matters is what the vendor built on top for your use case.

---

Part 2: The Doing

Where AI went from something you chat with to something that works for you. The shift happened in 2025, and it’s where the language gets thickest.

Agent / AI Agent

An AI that takes actions (searching the web, sending emails, updating spreadsheets, booking meetings) rather than just answering questions. You give it a goal; it figures out the steps. A chatbot is a colleague who answers questions but never leaves the desk. An agent is a PA who books the train, emails the client, updates the CRM, and sends you a summary.

Example: Intercom’s Fin operates as a customer service agent: resolving queries, processing requests, only escalating to humans when it hits limits. UK bank Monzo uses it.

Why it matters: Real productivity gains live here. But risk increases too: the AI is doing things. Ask: what actions can it take, and what approval is needed?

Agentic AI

The adjective form. Describes AI that plans, decides, uses tools, and completes multi-step tasks with minimal hand-holding. Non-agentic AI is a satnav giving one direction at a time. Agentic AI reroutes around traffic, suggests a fuel stop, and texts your contact that you’ll be late, all unprompted.

Example: Anthropic’s Claude in agentic coding mode: give it “build a dashboard for our sales data” and it writes code, runs it, finds errors, fixes them, and iterates until it works.

Why it matters: You’ll hear “agentic” in every pitch. Ask: how much autonomy is appropriate here? Booking rooms: low risk. Sending quotes without approval: high risk. It’s a spectrum, not a switch.

Tool Use

The ability of an AI to use external systems (databases, browsers, your CRM) rather than answering from memory. Ask a colleague the GBP-to-EUR rate: without tool use they guess; with it they check XE.com and give you the live number.

Example: ChatGPT browses the web, runs code, and analyses files via tool use. Claude supports custom tools so a business can connect it to inventory for answers based on real stock levels.

Why it matters: Tool use is the line between AI toy and AI tool. Without it, answers come from frozen training data. With it, the AI checks your live systems.

Workflow Automation

Software that automates multi-step business processes. AI adds the ability to handle tasks needing judgement (reading unstructured emails, classifying requests) which traditional automation couldn’t touch. Traditional automation (Zapier, Power Automate) is a conveyor belt that jams on the unexpected. AI adds a supervisor for the odd-shaped packages.

Example: Make (formerly Integromat) is hugely popular with UK SMEs. Power Automate, included in many Microsoft 365 plans, now has AI Builder for adding intelligence to flows.

Why it matters: Probably the most practical AI application for most SMEs right now. If your team spends hours chasing invoices, sorting enquiries, or compiling reports: start here.

Autonomous AI

AI that operates with minimal human oversight. True full autonomy is still aspirational. Today’s systems work within boundaries with human checkpoints at critical decisions.

Example: Intercom’s Fin handles customer service autonomously: resolving issues, looking up orders, escalating only when it hits limits. Monzo runs it this way.

Why it matters: For most SMEs, semi-autonomous is the sweet spot: AI handles the routine 80%, humans approve anything consequential. Be sceptical of “just works” promises. Ask: what’s the error rate? What’s the escalation path?

---

Part 3: The Architecture

How AI systems are built and connected. You won’t use these daily, but they help you see through vendor smoke.

Harness

A framework wrapping an AI model with structure, rules, tools, and guardrails. Controls what the AI can access, what it’s allowed to do, and what happens when things go wrong. The AI is a talented new hire; the harness is their job description, access permissions, and line manager in one.

Example: LangChain is the most widely used open-source AI application framework. Amazon Bedrock provides a managed harness with governance controls.

Why it matters: When a vendor claims “enterprise-ready,” ask about guardrails. What stops the AI accessing things it shouldn’t? How is it logged? No good answer? You’re the crash test dummy.

Orchestrator

A system coordinating multiple AI agents: deciding which handles what, passing information between them, keeping things on track. A construction site project manager: the plumber, electrician, and plasterer are all skilled, but someone decides the order and handles problems.

Example: Microsoft’s AutoGen is an orchestration framework. In practice, an orchestrator coordinates one AI reading emails, another checking CRM records, a third drafting responses.

Why it matters: AI from different vendors works in silos unless something coordinates it. Picking tools that integrate well matters more than picking the “best” individual tool.

Multi-Agent

Multiple AI agents working together, each specialised. Tasks split among specialists rather than one AI doing everything poorly. One person handling accounts, marketing, HR, and IT versus a small team where each focuses on their strength.

Example: Salesforce’s Einstein uses specialised agents for lead scoring, email drafting, and forecasting, coordinated within the Salesforce ecosystem.

Why it matters: When a vendor says “end-to-end AI,” ask: one model doing everything, or specialists working together? Best multi-agent systems behave like managed teams. The worst create chaos.

Context Window

How much information an AI holds in mind at once: your conversation, documents, instructions, its responses. Measured in tokens (roughly ¾ of a word). Small context window: tiny desk, you lose track of earlier discussions. Large one: massive desk with 20 documents spread out.

Example: Claude offers up to 200,000 tokens (~150,000 words). Gemini 1.5 Pro: 1 million. The original ChatGPT had 4,096, about 3,000 words.

Why it matters: Determines what you can practically do. Employee handbook queries, contract analysis: you need a big window. If a tool “forgets” what you said, the window is too small.

RAG (Retrieval-Augmented Generation)

The AI searches your documents first, retrieves the relevant bits, then generates its answer, rather than relying on general training alone. Ask a colleague about your returns policy: without RAG they answer from memory (possibly wrong); with RAG they check the wiki first.

Example: Microsoft Copilot for 365 uses RAG, searching SharePoint, OneDrive, Teams, and email to ground responses in your real data.

Why it matters: The difference between generic answers and answers from your actual data. When a vendor says “works with your documents,” ask: where is my data stored? How current is the index? Who controls what the AI sees?

MCP (Model Context Protocol)

An open standard for how AI connects to external tools and data. One universal protocol instead of bespoke integrations for every system. Before USB, every device had its own cable. MCP is the USB of AI connections.

Example: Claude Desktop supports MCP natively. Block (Square payments), Microsoft, Google DeepMind, and Amazon have all engaged with it.

Why it matters: Won’t affect you today; within 12–18 months it likely will. When choosing software, ask about open integration standards. USB-C versus proprietary charger. One future-proofs you.

---

What Now?

You don’t need to memorise any of this. You need to recognise the words when they’re thrown at you and ask better questions because of it.

Next time a vendor says “agentic AI with RAG-powered orchestration,” you’ll know that means: AI that takes actions, grounded in your documents, with something coordinating the steps. And you’ll know to ask: what actions? What data? What happens when it’s wrong?

Not to become an AI expert. Just to stop being at a disadvantage in conversations about tools that affect your business.

If any of this raised questions, get in touch. No pitch. Just a conversation.