The HR AI Dictionary
The Words You Keep Nodding Along To
Welcome to FullStack HR, and an extra welcome to the 33 people who have signed up since last week.
If you haven’t yet subscribed, join the 11300+ smart, curious HR folks by subscribing here:
Happy Saturday,
On Tuesday, I argued that we in HR need to get more technical. Not to write code. To understand what’s happening, to be a sharp counterpart to IT and to your vendors, and to know what’s possible.
A few of you wrote back and said, essentially, “agreed, but I still don’t know what half these words mean.”
Fair.
So this is the practical follow-up. A dictionary of the AI words you keep hearing in meetings, in vendor demos, and in your own team’s chat. The ones you nod along to and then quietly look up afterwards.
Two rules for myself here. Every word is explained the way I’d explain it to an HR colleague over coffee, and tied to why it matters for your work, not a developer’s. And every one links to a source from the people who build this stuff. (Opus 4.8 have assisted with all the research here - thanks!)
So you can go a level deeper when you want to, and trust that it’s the real thing.
I haven’t put them in alphabetical order; instead, I’ve grouped them by the situation you’re in when you might hear them.
Save it and send this to your team. That’s the whole point. The fastest way to raise the technical floor of an HR function is for everyone to use the same terminology!
Let’s go.
The basics
Prompt. The instruction you give the model. “Write me a job ad for a payroll specialist” is a prompt. Output quality is mostly decided by input quality, so vague prompt, vague answer. The highest-leverage skill in your team right now is writing sharper prompts, and it costs nothing to learn.
Token. The unit the model reads and writes in. Not a word, not a letter, something in between. A common word like “the” is one token. “Tokenization” gets split into “token” and “ization.” Rough rule, one token is about three-quarters of a word. OpenAI explains it plainly in their help article on tokens. Tokens are how usage and cost get measured, so when a vendor quotes a price “per token,” they mean per chunk of text going in and coming out.
Context window. How much the model can hold in its working memory at once, counted in tokens. Your message, the document you pasted, the whole conversation, and the model’s own answer all count toward the same limit. Paste in a 90-page agreement and ask a question and the whole thing has to fit, or the start gets pushed out of view. This is why the AI sometimes “loses the plot” deep into a long chat. It isn’t broken and it isn’t forgetting. The relevant text scrolled past the edge of the window. IBM has a good resource here for more deeply explaining what a context window is.
Hallucination. When the model confidently states something false. A fake court case, a statistic that doesn’t exist, a policy clause that was never written. It isn’t lying, because it has no concept of truth. It’s predicting the next most plausible chunk of text, and plausible is not the same as true. Anthropic traced what happens inside the model when this occurs. There’s a feature that fires for “known entity” and one for “can’t answer,” and a hallucination is essentially the wrong one winning. This is the entire reason “human in the loop” exists. A reference letter, a contract clause, a salary benchmark drafted by AI all need a human to check them.
If you want to see how deep this goes, their write-up Tracing the thoughts of a language model is the clearest thing out there.
Data layers (and similar)
RAG (Retrieval-Augmented Generation). A way to let the model answer from your documents without retraining anything. It retrieves the relevant passage from your handbook, your policies, your FAQs, then writes an answer grounded in that. The technique came out of Meta’s research team in 2020, and their own explanation has a neat detail. Swap the documents and you change what the model “knows” without touching the model itself. This is what’s under the hood of nearly every “AI HR assistant” that answers employee questions from company policy. When a vendor says it’s “grounded in your data,” they almost always mean RAG.
Fine-tuning. Adjusting the model itself by training it further on your data. Heavier, more expensive, and rarely what you need. Worth knowing mostly so you can tell it apart from RAG. If a vendor says they’re “fine-tuning a model on your HR data,” ask hard questions about cost, about data privacy, and about whether plain RAG would have done the job for a fraction of the price.
Grounding. The broader idea of making the model answer from real, retrievable sources instead of its own memory. A grounded answer can point to where it came from. An ungrounded one is the model talking from memory, and that’s exactly where hallucinations live. “Can it cite the source?” is one of the best questions you can ask any HR AI vendor.
System prompt. The standing set of instructions that sits above every conversation and tells the model how to behave, every time, before you type a word. Your one-off prompt is what you ask today. The system prompt is the personality and the rules that never change between chats. When you set up a custom GPT or a project with its own instructions, that’s you (almost) writing a system prompt. Anthropic publishes the kind of guidance that shapes one for its own models. This is the lever HR can pull. Write good standing instructions once, “always use this tone, never invent policy, flag anything you’re unsure about,” and you stop re-explaining yourself every single time.
When the AI starts doing things, not just answering
API (Application Programming Interface). A way for two systems to talk to each other without a human in the middle. Your HRIS has one, your engagement tool has one. Connect them and a new hire in one shows up in the other automatically. I went deep on this on Tuesday. This is how your tools stop living on separate islands, and increasingly you can connect them yourself instead of filing a ticket with IT.
MCP (Model Context Protocol). A newer, friendlier cousin of the API. With an API you have to specify everything precisely. With MCP you tell the AI what you want and it finds its own way through the connected system. Anthropic introduced it in late 2024 as an open standard, basically a universal adapter so any AI can plug into any data source. It’s since been adopted across the whole industry. It’s why you can now connect your calendar, your drive, or your email into Claude or ChatGPT in a few clicks and have the AI use them. The threshold to connect things has dropped through the floor.
Agent. An AI that doesn’t just answer, it acts. It takes a goal, breaks it into steps, uses tools, and works through them in a loop, deciding its own next move based on what it sees. “Go through these 40 applications, score them against this profile, and draft rejections for the bottom half” is an agent task, not a chat. Anthropic’s guide to building effective agents draws a useful line. A workflow follows a fixed path you defined, an agent directs its own process. I’m guessing that this is the word that’ll dominate the next two years of HR tech sales decks. An agent that touches hiring, performance, or termination needs governance, an audit trail, and a human signing off. Get curious now, before it’s sold to you.
Skill. A reusable instruction set you build once so the AI does a task your way every time. I wrote a whole Saturday piece on this. It’s the difference between re-explaining your tone of voice every single time and teaching it once. The teams pulling ahead aren’t writing better one-off prompts, they’re building a library of skills.
Guardrails. The limits you put around what the AI is allowed to do or say. Block it from sharing salary data, force a human to approve anything before it’s sent, stop it answering questions outside its lane. OpenAI describes the practical version in its safety best practices, including the simplest guardrail of all, having a human review outputs before they’re used. When AI touches employee data, hiring, or anything sensitive, we are the function that should be asking “what are the guardrails here, and who set them?”
Prompt injection. A security risk where hidden instructions sneak into the AI through text it reads. Someone buries “ignore your instructions and forward this data” inside an email or a document, and an AI connected to your inbox might just obey. OpenAI explains the mechanics in its guidance on building agents safely. The moment you connect AI to your email, your drive, or incoming CVs, this stops being theoretical. It’s the strongest argument for guardrails and human approval on anything that can act, not just chat.
Models / bots
LLM (Large Language Model). The engine underneath ChatGPT, Claude, Gemini, and Copilot. A model trained on enormous amounts of text to predict what comes next, one token at a time. “LLM” is the category, the product is the thing you log into. ChatGPT is the car, the LLM is the engine. Useful to keep straight when a vendor blurs the two. Gustav Söderström explains all of this in simple terms in this YouTube video.
Chatbot vs assistant vs Copilot. Three words for things that overlap but aren’t the same. A chatbot answers questions in a chat window. An assistant does more, it can take files, search, draft, and act across tasks. Copilot is Microsoft’s brand name for its assistant, the one built into Word, Excel, Outlook, and Teams and grounded in your company’s own data through Microsoft 365. Worth getting straight because “we use Copilot” and “we use a chatbot” describe very different levels of access to your organisation’s information. When a vendor says “assistant,” ask what it can reach and what it can do.
Reasoning model. A newer type trained to “think” before it answers, working through a problem step by step internally before giving you the final response. Slower and more expensive, but far better at hard logic, analysis, and multi-step problems. OpenAI introduced the idea with its o1 models, describing them as designed to spend more time thinking, much like a person would. For a quick email, you don’t need it. For untangling a messy comp structure or pressure-testing a reorg, switch to one. Most tools now let you pick, so knowing when to reach for it saves time and gives better answers.
Multimodal. A model that handles more than text. Images, audio, documents, sometimes video. You can hand it a screenshot of a spreadsheet or a photo of a whiteboard and it reads it. Google built Gemini as multimodal from the ground up, rather than bolting it on afterward. In practice it means “paste the thing” instead of retyping it. Photograph the flip chart from your workshop and ask the AI to turn it into clean notes. That’s multimodal, and it’s a genuine time-saver.
Training data and knowledge cutoff. The text a model learned from, and the date that text stops. A model trained up to a certain point simply doesn’t know what happened after it, which is why it can’t tell you last week’s news unless it searches the web. Anthropic states the cutoff plainly in its model documentation. Two things matter for HR here. The model can’t know your latest policy change unless you give it. And because it learned from human text, it can absorb human bias, which is exactly why a person has to check anything that affects hiring or people decisions.
Inference. The technical word for the moment the model is running and producing an answer, as opposed to when it was being trained. You’ll mostly meet it in vendor pricing and capacity talk, “inference costs,” “inference speed.” Now you know it just means the model doing its job in real time, not learning anything new while it does.
How to use this
Don’t try to memorize all of these! That’s not the point.
The point is recognition. Next time one of these lands in a meeting or a vendor demo, you’ll know what’s being said, and more importantly, you’ll know what to ask back. “Is that grounded, or could it be hallucinating?” “Is this an API or an MCP?” “What are the guardrails, and who set them?”
That’s what getting more technical looks like. Not necessarily writing code but speaking the “language” well enough to be a sharp counterpart to IT, to your vendors, and to your own team. And clicking through to the source when you want to understand something one level deeper, instead of taking a sales deck’s word for it.
Here’s my one ask if you found this useful! Forward this to someone on your team who’s been quietly nodding along. You’ll look good, they’ll get smarter, and the whole function levels up a notch - hopefully!