AI glossary
An AI glossary, no hype.
The words you'll hear when people talk about automation and AI, each explained in one line. You don't need to know them all. That's my job.
Behind every automation there are lots of pieces. You don't need to know what each one is: that's my job. But if you're curious, tap any of them.
Generative AI / LLM
The “brain” that reads and writes text, like ChatGPT or Claude.
Large and small models
Large ones reason better; small ones are fast and cheap. Efficient means using each where it fits.
Prompt
The instructions you give the AI. Written well, they change everything.
API
The door two programs use to talk to each other.
Webhook
An automatic “something just happened” ping.
JSON
The format data travels in between apps.
MCP
A standard plug that connects AI to your tools.
RAG
Letting AI answer from your documents, not from memory.
AI agent
AI that doesn't just answer, but takes steps and uses tools.
Human-in-the-loop
The point where a person checks before things move on.
Evals
Tests that check the AI answers well, before and after every change.
Prompt injection
Hidden instructions in an email or web page that try to trick the AI.
Least privilege
Each piece gets only the access it needs, nothing more.
The AI
- Generative AI / LLM
- The “brain” that reads and writes text, like ChatGPT or Claude.
- Prompt
- The instructions you give the AI. Written well, they change everything.
- Context window
- How much the AI can keep in mind at once.
- Context engineering
- Choosing what information the AI gets to see, not just how you ask.
- Tokens
- The little chunks of text AI reads, and pays by.
- Hallucination
- When AI makes something up, confidently.
- AI agent
- AI that doesn't just answer, but takes steps and uses tools.
- RAG
- Letting AI answer from your documents, not from memory.
- Embeddings and vector database
- How your information is organised so AI finds what's relevant.
- AI model
- Each company (OpenAI, Anthropic, Google…) has several, and each shines at something: text, code, images, video or research.
- Large and small models
- Large ones reason better; small ones are fast and cheap. Efficient means using each where it fits.
- Jev
- An AI that doesn't write: it gives the probability of each option.
- Pay per use
- Most models charge for every request. With the wrong model at scale, the bill takes off.
- Multimodal
- AI that understands or creates more than text: images, audio or video.
- Deepfake
- A fake video, voice or image made with AI, and harder to spot every month.
- Human-in-the-loop
- The point where a person checks before things move on.
- Evals
- Tests that check the AI answers well, before and after every change.
- Fine-tuning
- Extra training so a model learns your style or your task.
- Reasoning models
- Models that “think” step by step before answering. Slower, but more reliable on hard problems.
- Temperature
- The dial that sets whether the AI answers creatively or predictably.
- System prompt
- The background instructions an assistant always follows, even if you can’t see them.
- Bias
- When the AI repeats the prejudices in the data it learned from.
- Grounding
- Tying the AI’s answers to real sources you can check.
- Few-shot
- Giving the AI a few examples so it picks up the pattern.
- Open models (open source)
- Models you can download and run yourself, like Llama or Mistral.
- Local AI
- AI that runs on your own computer or server: your data never leaves the building.
Connecting tools
- API
- The door two programs use to talk to each other.
- Webhook
- An automatic “something just happened” ping.
- JSON
- The format data travels in between apps.
- MCP
- A standard plug that connects AI to your tools.
- OAuth
- The “sign in with Google” that grants access without sharing passwords.
- Trigger
- What starts a flow: an email, a form, a time of day.
- OCR
- Reading text from photos and scanned PDFs.
- Low-code / no-code
- Tools like n8n, Make or Zapier for building flows without coding everything.
- Python
- The programming language most used for AI and custom automations.
- Self-hosting
- Running a tool on your own server instead of using its cloud version.
- CRM
- Where your customers and contacts live.
- Vendor lock-in
- Depending so much on one tool that switching costs a fortune.
Data and automation
- Workflow
- The chain of steps an automation follows, from start to finish.
- RPA
- Software “robots” that click through screens the way a person would.
- Structured and unstructured data
- Tables on one side; emails, PDFs and audio on the other. AI can now read the second kind too.
- Data extraction
- Pulling the useful bits out of a document: amounts, dates, names.
- Classification
- The AI sorting things into boxes: urgent or not, invoice or quote.
- Transcription
- Turning audio and meetings into text.
- Chatbot
- An assistant that answers customers or the team at any hour.
- Single source of truth
- One place where the right version of each piece of data lives.
- Integration
- Two tools sharing data without copy and paste.
- Scraping
- Collecting data from websites automatically.
- Database
- Where organised information lives: Airtable, Notion or a classic one.
- Text-to-speech
- Turning text into a voice that sounds natural.
- ETL
- Extract data from one place, clean it and load it into another.
- Parsing
- Breaking a text or file into parts a program can understand.
Security
- API keys
- The “key” that unlocks a tool. Kept locked away, never in an email.
- Secrets manager
- The safe where passwords and keys live.
- Least privilege
- Each piece gets only the access it needs, nothing more.
- 2FA / MFA
- A second step when signing in, like a code on your phone.
- SSO
- One company login for every tool.
- Encryption
- Data travels and is stored unreadable to anyone who shouldn't see it.
- Prompt injection
- Hidden instructions in an email or web page that try to trick the AI.
- Data leak
- Information ending up where it shouldn't.
- Shadow AI
- AI tools people use without anyone approving them.
- Audit log
- The record of who did what, and when.
- Anonymisation
- Removing names and personal details before AI sees them.
- EU data residency
- Keeping your data on European servers.
- DPA (data processing agreement)
- The contract that makes a provider look after your data.
- Backup
- So one mistake doesn't wipe out your work.
- AI phishing
- Fake emails and calls that look far more real with AI.
- Jailbreak
- Tricks to make an AI break its own rules.
- Sandbox
- A safe copy for testing changes without touching the real data.
- Watermark
- A hidden signal that shows a text, image or video was made with AI.
Rules
- GDPR
- The personal-data rules every automation must respect.
- EU AI Act
- The EU law on using AI responsibly.
- AI literacy
- Knowing enough about AI to use it well at work. The AI Act asks companies to support it.
- High-risk AI
- Uses the AI Act watches closely, like hiring, credit or health.
- Transparency
- Telling people when they’re talking to an AI or seeing AI-made content.
- AI use policy
- The company’s rules on which AI tools to use and how.
The hard part isn't any one piece. It's knowing how they fit.
An automation that really works brings many of these pieces together. Choosing which ones you need, which you don't and how to connect them so it doesn't break the first time is exactly my job. And if you're curious, I'll explain it along the way.
Let's talk
Got a workflow worth taming?
Book a free 20‑minute call. Tell me what's holding you back and I'll tell you honestly whether pleko is a fit.