Plain-English terms for the ideas that otherwise slow people down.

An agent is an AI-assisted workflow that can do more than answer a single question. It may call tools, read approved sources, create drafts, update artifacts, or continue through multiple steps.

The useful question is not whether something is called an agent. The useful question is what it may read, what it may change, when it must stop, and who reviews the result.

The context window is the working space available to a model in one exchange. When it fills up, important instructions or reference material may fall out of view.

Grounding means giving the model the material it should rely on, instead of asking it to guess from general training patterns alone.

A guardrail is a limit placed around AI use. It may restrict what data can be used, which tools can be called, what output shape is allowed, or when a person must review the result.

Guardrails should be visible in the workflow, not hidden as good intentions. They work best when they are tied to stop conditions, review gates, and evidence that can be checked later.

A hallucination is not a rare bug. It is a predictable outcome when a model lacks the right context, tries to satisfy a prompt too aggressively, or fills gaps with plausible-looking text.

Hallucinations are especially risky because they often sound specific, helpful, and complete. Good AI use does not depend on never seeing them. It depends on grounding, verification depth, confidence labels, and escalation before unsupported details become trusted work.

Human in the loop means a person has a defined role in the AI workflow. The person may approve an output, check sources, resolve uncertainty, or decide whether the work should stop or continue.

This is different from vague oversight. A useful human checkpoint names what the person reviews, what evidence they need, and which actions still require approval.

A prompt is the working request you give a model. It can include the task, audience, source material, constraints, examples, and the shape you want the answer to take.

A stronger prompt does not make review optional. It makes the work easier to check because the model’s instructions, source limits, and expected output are visible.

Retrieval is the process of bringing relevant source material into the model’s context when a request is made. It is often used when the model needs current documents, internal knowledge, or approved reference material.

Retrieval helps only when the fetched material is appropriate for the task. The system still needs permission checks, source-quality checks, and a way to handle missing or conflicting material.

A source trail is the evidence path behind an AI-assisted answer. It shows which source supports a claim, which claims are inferred, and which claims still need checking.

The source trail does not need to be elaborate. A short table with claim, source, confidence, and review owner is often enough to keep fast work inspectable.

Verification is the habit of checking important AI output before it moves forward. The check may be light for low-risk drafts or strict when the answer affects decisions, risk, money, policy, or people.

Good verification names what was checked, which source or owner was used, and what remains uncertain. It turns review from a vague feeling into a visible decision.