AI Concepts Map

Get a working vocabulary for AI that holds up in a meeting or a vendor pitch: what the terms mean, which of your tasks today's tools help with, and three questions that place any claim you hear.

Concept definitions follow the NIST AI Risk Management Framework and its generative AI profile (standard, not research: NIST AI 100-1, 2023; NIST AI 600-1, 2024). The capability evidence is mixed research, strong within each study's setting and narrow outside it (Noy & Zhang, 2023; Brynjolfsson, Li & Raymond, 2025; Dell'Acqua et al., 2026). The checklists are practice, not research.

Use it when:

• You are new to AI at work and want to understand the words before you form opinions.
• You are going into a meeting or a vendor pitch where AI will be discussed and want questions that cut through.
• A colleague, your manager or the news makes a sweeping claim about AI and you want to judge it calmly.

Skip it when:

• You already know the terms and want to decide which of your tasks to give AI: use 10.7 AI Task Map.
• Your question is what you are allowed to do: use 10.3 Rules That Apply to My Role.

Most confusion about AI comes from treating it as one thing. The NIST framework (NIST, 2023) defines AI broadly, as machine systems that infer from input how to produce outputs such as predictions, content or decisions, and its generative AI profile (NIST, 2024) handles content-producing systems separately because they bring their own risks, such as fluent statements that are false. A language model is built to produce likely text, not to look up facts, so a wrong answer arrives with the same confidence as a right one.

The evidence on usefulness is real but narrow. Experiments found gains on bounded writing tasks (Noy & Zhang, 2023) and in customer support, mostly for newer agents (Brynjolfsson, Li & Raymond, 2025). A preregistered study with consultants found that AI help improved work on tasks within the model's reach and hurt accuracy on one task just beyond it (Dell'Acqua et al., 2026). Each study covers one setting and one model generation. So this page sorts the terms first, then your tasks, and ends with three questions for any claim.

Each rung is a subset of the one above. For each, name a tool or feature you have actually met and one thing people wrongly expect from it.

Strong answer: LLM: "The chat assistant in our office suite. Wrongly expected: that it knows our current price list. It has never seen it, so any price it gives is a guess."

Why language models make things up: the core operation is producing likely text, not retrieving checked facts. When the likely text is wrong, it reads exactly as fluently as when it is right. That is a design property, not a rare bug, which is why checking output (10.9) is a standing habit.

Illustrative example: the roles, the organization and the vendor are made up.

The procurement lead at a municipal water utility in Denmark hears a vendor say its assistant "automates supplier risk assessment".

Step 1: she places it on the ladder as an LLM-based generative tool. Wrongly expected: that it has checked financial data on each supplier. It reads public news only.

Step 2: strong side for her: summarizing what she gives it. Weak side: judging whether a supplier can deliver, which needs facts the tool cannot see.

Step 3: claim: "automates risk assessment", heard in a sales meeting. Which AI and task: news summaries, not credit judgment. Evidence: a vendor demo, ticked as such. Who checks: her team, and the utility stays accountable for the choice.

Step 4: verdict: use only with checking. First test: three summaries of suppliers using public news, compared with the original articles. Decide by the end of the month.

The purchase changes shape: from an "assessment engine" to a monitoring aid with a human in the loop.