Exam-style
A support team runs three systems. The first reads an incoming customer email and drafts a reply, which a support agent edits and sends. The second answers staff questions from the company's policy library and cites the documents it drew each answer from. The third reads an incoming email, looks up the order in the order system, decides whether a refund is warranted, issues the refund, and writes back to the customer, choosing for itself which of those steps a given email needs. Which of the three is agentic AI, and what makes it so?
Reveal the answer
C. The third, because it decides its own sequence of steps and carries them out against external systems rather than producing output for a person to act on
Agentic AI is a generative model given tools it can call, memory that carries across steps, and the autonomy to choose a multi-step course of action against systems outside itself. Only the third system has all of that. It reads the email, queries the order system, applies a judgement about whether a refund is warranted, moves money, and replies, and it decides per email which of those steps are needed. Nobody signs off between the steps, and the world is different afterwards. The first system generates and stops. A drafted reply sits in an inbox until a support agent reads it, edits it and presses send, so the human is the one taking the action and the model produced text for them to act on. Fluent, useful, and not agentic. The second is retrieval-augmented generation: a search over the policy library retrieves relevant passages, those passages go into the prompt, and the model answers from them and cites where each part came from. That is grounding, which improves how accurate and traceable the answer is, and it changes nothing about what the system does with the answer. The retrieval is a fixed step in a pipeline somebody wrote, not a tool the model chose to call, and the output still lands in front of a person. The count of agents is a separate question. A single system that plans and acts on its own is agentic AI; multi-agent collaboration is one way of building agentic systems, not the definition of them. And building all three on foundation models makes all three generative AI applications, which is a wider label. The line to hold is whether the system acts or only produces output.
Q. One system drafts a reply for a support agent to send. One answers questions from a document library and cites its sources. One reads an email, checks the order system, issues the refund and writes back, deciding for itself which steps are needed. Which is agentic AI?
A. The third. Agentic AI is a generative model given tools, memory and the autonomy to take multi-step action against external systems, instead of producing text for a person to act on.
Why? Artificial intelligence (AI) is the outer label, generative AI (GenAI) is the part that produces new content, and a large language model (LLM) is what sits at the centre of all three of these. That shared machinery is why the first option tempts. Sort them by what happens to the output. The drafter stops at text and a human sends it. The document-library system is retrieval-augmented generation, which grounds the answer in retrieved passages so it can be cited and checked, and retrieval is a step in a pipeline somebody wrote rather than a tool the model elected to use. Only the refund system chooses a sequence and executes it, and one agent acting alone is enough; a second agent is not part of the definition.