Your colleague just said “we need to implement a RAG pipeline with human-in-the-loop checkpoints to reduce hallucinations across our agentic workflows.”
You nodded. You had no idea what that meant.
Here’s what they were actually saying.
Hallucination
AI makes things up. Not maliciously—it just can’t tell you when it doesn’t know something, so it fills the gap with something that sounds right.
You’ve probably already hit this. You ask an AI tool to summarise a document, and it adds a statistic that wasn’t there. You ask it to draft a client email with specific figures from a report, and it invents one. You copy it across without checking, and now that invented number is in front of your client.
The fix is simple: treat AI output the way you’d treat a first draft from a junior colleague. Read it, check the facts, and don’t send anything you haven’t verified.
Agentic AI
Most AI tools answer questions. Agentic AI does things.
Learn more about the difference between an AI tool and an AI agent.
Think of it this way: a standard AI tool is like a very fast researcher who hands you information. An agentic system is more like an assistant who takes the information and acts on it—booking the meeting, sending the follow-up email, updating the CRM entry.
If you’re in sales, this is relevant now. Some CRM platforms already use agentic AI to log calls, draft follow-up emails, and flag deals that have gone quiet.
The upside is hours saved per week. The thing to watch is that if it’s sending emails on your behalf, you need to know what it’s saying. Check the outputs before they go out, at least until you trust the system.
RAG (Retrieval-Augmented Generation)
This sounds complicated, but is quite simple.
Normally, an AI model knows only what it was trained on—which has a cutoff date and doesn’t include your company’s internal documents. RAG connects the model to your own files before it answers, so its responses come from your data rather than general knowledge.
In practice: imagine asking an AI tool “what’s our refund policy?” and getting the actual answer from your internal handbook, not a generic guess. That’s RAG.
If your team uses any AI tool that’s connected to your company documents, SharePoint, or internal knowledge base, it’s probably using some version of this.
Learn more about RAG in my article on Retrieval-Augmented Generation.
Shadow AI
This one’s about what’s already happening in your organisation without anyone officially approving it.
Someone on your team is pasting meeting notes into ChatGPT to write a summary. Someone else is using an AI tool to draft client proposals. Another person is running customer feedback through an AI analyser. None of it went through IT. None of it was risk-assessed.
This matters because some of those tools store the text you paste into them. If that text contains client data, financial information, or anything confidential, it’s now sitting on a third-party server somewhere. Worth knowing what you’re putting in before you hit send.
Learn more about shadow AI and how to guard against the risks.
Human-in-the-Loop (HITL)
Exactly what it sounds like. A human checks what the AI did before anything important happens.
An AI tool flags a contract clause as risky. A lawyer reads it before anyone signs. An AI draft goes out after a human edits it. An AI recommendation gets reviewed before it becomes a decision. That’s human-in-the-loop.
The reason it matters: AI systems can be confidently wrong (see: hallucination). In any situation where a mistake has real consequences—a client commitment, a financial decision, a legal document—you want a human checkpoint in the process.
P.S. A human in the loop doesn’t magically make things better. Here’s why.
Model Card
Before you buy a piece of software, you check what it does. A model card is the equivalent for AI—a document that tells you what a model was trained on, what it’s good at, and where it falls short.
In practice, most people never see these. But they’re worth knowing about if you’re evaluating an AI tool for your team.
A model card might tell you the tool wasn’t trained on medical data (relevant if you work in healthcare), or that it performs poorly on languages other than English (relevant if you have an international team). It’s the spec sheet most vendors don’t advertise.
Learn more about how to scrutinise a vendor’s model card properly.
AI Maturity
A way of measuring how ready a team or organisation actually is to use AI—not whether they’ve bought a tool, but whether they can actually use it well.
Low maturity looks like this: the company buys a licence, runs one demo, and the tool gets forgotten because nobody knows what to do with it.
High maturity looks like this: clear use cases, staff who know how to prompt well, clean data the AI can work with, and someone responsible for checking outputs.
If your team is somewhere in the middle—using AI occasionally but not consistently—that’s normal. Most teams are. The question worth asking is: what’s the one workflow where this would actually save us time?
Prompt Engineering
The way you talk to AI directly shapes what you get back.
“Write me a summary” produces a generic paragraph. “Write me a three-bullet summary of this document for a non-technical finance director, focusing on budget implications, in under 100 words” produces something you can actually use.
Prompt engineering is just the practice of being more specific with your instructions. You don’t need a course. You need to treat the AI like a new hire on their first day—give it context, tell it who the audience is, specify the format you want, and tell it what to leave out.
| Term | Plain English | Watch out for |
| Hallucination | AI confidently making things up | Always verify facts before sending |
| Agentic AI | AI that acts, not just answers | Check outputs before they reach anyone |
| RAG | AI connected to your own documents | Know which files it can access |
| Shadow AI | Unsanctioned AI tools your team already uses | Check what data you’re pasting in |
| Human-in-the-Loop | A human reviews AI output before it matters | Don’t skip this step under time pressure |
| Model Card | The spec sheet for an AI model | Ask vendors for it before you commit |
| AI Maturity | How ready your team is to use AI well | Buying a tool isn’t the same as using it |
| Prompt Engineering | Better instructions = better outputs | Be specific: context, audience, format |
Take the next step in your journey by booking a personalised Claude training session today.