AI conversations move fast, and the jargon moves with them. This is a plain-language reference for the terms that come up most often — written for someone running a business, not building a model.
Where it's relevant, each term links to the Keystone page where it matters in practice, so you can see how it plays out in a real Irish business rather than in the abstract.
An AI agent is software that can take actions on its own to reach a goal, not just answer a question when you ask one. Give it a job — chase an overdue invoice, sort the morning's emails, reconcile two systems — and it works through the steps itself. Think of it less as a chatbot and more as a junior team member who never clocks off.
Keystone runs a nine-agent hub for a 30-vehicle Irish waste operation, where separate AI agents handle billing, triage, credit control and reconciliation without anyone pressing go each morning.
An AI hub is one central intelligence layer that all your agents and systems plug into, rather than a scatter of disconnected tools. You build the hard part — the connections to your data and software — once, and everything new bolts onto it afterwards. That's why the second thing you add costs far less than the first.
When we handle the custom build of a hub for an Irish business, the first agent carries the setup cost; every agent after that plugs into the same foundation, so the price comes down each time.
Automation means getting software to do a task so a person doesn't have to. Rule-based automation follows fixed steps and breaks the moment something unexpected turns up; AI automation can handle the variation — a supplier who words their invoice differently, an email that doesn't fit the usual pattern. The difference matters when your real-world work is rarely identical twice.
A rule-based script can move a tidy invoice into your accounts; an AI-driven automation can read fifty invoices that all look different and still pull out the right figures.
Business Process Automation is using software to run a whole business process from start to finish, rather than automating one isolated task. Instead of speeding up a single step, it joins the steps together — an order comes in, gets checked, invoiced and logged without a person shepherding it between systems. The payoff is a process that runs itself, not just a faster keyboard.
For an Irish services firm, that might mean a new job flowing from quote to contract to first invoice automatically, with staff stepping in only for the judgement calls.
A Data Processing Agreement is the contract that sets out how a supplier is allowed to handle data you give them — where it's stored, who can see it, and what they can and can't do with it. If you're a legal, healthcare or financial firm handling sensitive client information, it's the document that keeps you on the right side of GDPR. Any serious AI supplier should be happy to sign one.
Before an Irish solicitor's practice lets an AI tool near client files, the DPA is what confirms that data stays in the EU and is never used to train someone else's model.
The EU AI Act is the European Union's law that sorts AI systems into risk categories — from minimal risk right up to banned — and puts stricter obligations on the higher-risk ones. For most Irish SMEs the everyday tools you'd actually use, like drafting, summarising or reconciling, sit in the low-risk band with light-touch rules. It's worth knowing which category your use falls into, but it's rarely the roadblock people fear.
An Irish business using AI to draft emails or tidy its accounts is in low-risk territory; the heavy obligations land on things like AI used in recruitment or credit scoring.
A hallucination is when an AI states something false while sounding completely confident — a made-up figure, a citation that doesn't exist, a plausible answer that simply isn't true. It happens because the model is predicting likely words, not checking facts against reality. That's exactly why a human should review anything that leaves the building.
An AI might confidently quote the wrong VAT rate in a client letter; the safeguard is a person signing off before anything with your name on it goes out the door.
A Large Language Model is the engine behind most modern AI tools — the technology that reads a request and generates a written response. It's trained on an enormous amount of text so it can predict what words should come next, which is how it drafts, summarises and answers. When people say "the AI," an LLM is usually what's doing the work underneath.
The tool that reads a messy customer email and drafts a sensible reply is an LLM at work; Keystone's job is shaping it to your business so the replies actually sound like you.
Prompt engineering is the craft of writing the instructions that make an AI reliable for a specific job. The same model can give you vague waffle or a precise, usable answer depending on how well it's briefed. It's the difference between a tool that's impressive in a demo and one you'd trust on a Monday morning.
When Keystone calibrates a system to an Irish business context, a large part of the work is writing and testing the instructions so the AI understands your terminology, your rules and your context — not the generic internet's.
RAG is a method where the AI answers from your own documents and data rather than the open internet. It first retrieves the most relevant passages from your files, then uses them to write the answer — so the response is grounded in your reality, not a guess. It's a big part of why an AI built on your material gives accurate, checkable answers.
Ask an AI built with RAG "what's our cancellation policy?" and it pulls the answer straight from your actual contract, not from something vaguely similar it once saw online.
RPA is software that repeats rigid, rule-based steps exactly the same way every time — clicking through screens, copying fields, moving data between systems. It's fast and tireless, but it follows fixed rules and stumbles the moment the input changes. The contrast with AI is simple: RPA follows the rules, AI adapts to the variation.
RPA can copy figures from one Irish accounting system to another all day; the moment a supplier changes their invoice layout, it takes AI's flexibility to keep up.
A system prompt is the standing instruction that sets how an AI behaves for a given task — its role, its tone, its boundaries, the things it should always or never do. The user's question changes each time; the system prompt stays put in the background, keeping the AI on-brief. It's what stops a professional tool from drifting off into generic chatbot mode.
An Irish firm's customer-service AI might carry a system prompt telling it to stay polite, stick to company policy, and hand anything it's unsure about straight to a human.
AI reads and writes in tokens — small chunks of text, roughly a few characters each — and its context window is how many of those it can hold in mind at once. Once a conversation or document runs past that limit, the earliest parts start to fall out of view, which is why a long thread can seem to "forget" what was said at the start. Knowing the limit exists helps explain why AI sometimes loses the thread.
Paste a 90-page contract into a small-window tool and it may lose track of clause two by the time it reaches clause forty; a system built for the job keeps the whole document in view.
Training data is the collection of examples an AI learned from before you ever used it — the text and information that shaped what it knows and how it responds. It explains both what the model is good at and where its blind spots and biases come from. It's the most common follow-up to "what is AI," because everything the model does traces back to what it was trained on.
A general AI trained on the wider internet won't know your pricing or your processes — which is why Keystone grounds systems in your own documents rather than relying on generic training data.
A vector database stores your documents in a way that lets an AI find information by meaning rather than by exact keyword. Ask a question and it retrieves the passage that's actually most relevant, even when it doesn't share the same words. It's the quiet machinery behind an AI that can answer well from your own material.
Search your files for "getting out of a contract early" and a vector database will still surface the clause headed "termination," because it matches on meaning, not just wording.
Twenty minutes on a free Discovery Call and we'll translate it into what it means for your business — no jargon, no pitch deck, no follow-up pressure.
Book Your Discovery Call