If your team promises to deliver (or buy!) "Agentic AI", then everyone needs to have a shared understanding of what that means; you don't want to be the one left trying to explain the mismatch to stakeholders six months later. There's no current (2025-03-30) widely accepted definition, so if you're using the term, be clear on what you mean, and if someone else is using the term, it's worth figuring out which one they mean.

So here are four definitions of “Agentic AI”, varying in utility and accuracy, presented in order from broadest to most specific:

• Marketing: “the next generation of AI that’s right around the corner”
• Overly simplistic: “an LLM with tool use capacity”
• Broad Agency: “an LLM with agency; it can make things happen outside the chat window”
• Narrow Agency: “an LLM with agency; it can plan and execute tasks that happen outside the chat window”

The Marketing Definition

If you’ve spent any time reading marketing material, listening to pitches, or attending AI expos recently, you’ll have come to appreciate that “Agentic AI” is a buzzword that means “software that will magically solve your problems” + a heavy dose of science fiction. It is by far the most common among non-technical people.

Your “HR AI” will talk to your “payroll AI” who has a team of “recruiter AIs” and so on. Your “Agentic AI” will tell you what to wear in the morning before telling your cupboard to steam and press appropriate clothes. “Agentic AI” will find and fix security holes in your codebase before a single line of code has been written.

The marketing definition of “Agentic AI” is simply ‘advanced LLMs’, and it is everywhere. This framing may be strategically useful if you’re selling to audiences with limited technical background or seeking investment from those more captivated by trending terminology than technical implementation.

Definition rating: technically meaningless, but excellent for fundraising decks

Overly Simplistic Definition

If you start to push vendors telling you about their already-deployed or sellable “Agentic AI”, this is what you’re usually going to find under the surface: they’ve wired up an LLM — potentially or actually — to systems that don’t simply generate tokens. Maybe this is a RAG system (“it goes off and searches all your documents”), maybe this is making real-time queries on a database using MCP, maybe this is a chat bot that can pass-off a conversation to a human.

This is … not terrible? But it’s way too broad to be useful, and if you’re the user or receiver of the term, it’s very useful to be clear that this is overly broad, and that technical or well-read people are going to look at you askance if this is the definition being used…

Definition rating: far too broad to be useful, but just specific enough to avoid outright dismissal

Moving from Marketing to Meaningful Definitions

The previous definitions reflect how “Agentic AI” is often discussed in non-technical contexts. However, to develop more useful definitions, we should consider what “agency” actually means.

Agency is a noun that means “the ability to take action or to choose what action to take”. This gives us two more precise ways to understand Agentic AI:

1. “Broad Agency”: AI systems with “the ability to take action” beyond generating text
2. “Narrow Agency”: AI systems with “the ability … to choose which action to take”

Let’s examine each of these more substantial definitions…

Broad Agency

This is the first reasonable definition of “Agentic AI”, although it’s very broad: any LLM system that can cause effects outside its chat window. My work involves building NPCs for an online game that can send users items based on chatting with them. Under this definition, that’s “Agentic AI”, as is a system that can raise a technical support ticket, make a booking, or change a value in a database based on chatting to it.

Often these systems rely on the LLM’s generally excellent ability to classify or identify a user’s intent, before passing it off to Plain Old Code to make decisions about that and execute on it. In this definition, the LLM provides a chat interface to software that otherwise looks pretty traditional: if/then/else, state-machines, decisions trees, API calls, whatever else under the hood, but the LLM is providing an interface to it.

Definition rating: practical and defensible interpretation that covers many real-world systems without making excessive claims

Narrow Agency

The “Broad Agency” definition used the definition of “agency” that meant “the ability to take action”, and the “Narrow Agency” definition uses the second part: “the ability to … choose what action to take”. This overlaps with Anthropic’s definition in the seminal article Building Effective Agents, which describes agents as tools that “plan and operate independently” — - making their own decisions about when and how to use available tools to accomplish tasks. “Agentic AI” in the “Narrow Agency” sense is AI LLMs that independently select from available tools, plan solutions, and evaluate their own progress—with decision-making occurring within the LLM itself.

Working with LLM outputs daily has taught me to be cautious about their probabilistic nature. These systems can perform consistently for hundreds of iterations, then unexpectedly produce different results with identical inputs. For production environments, this unpredictability creates meaningful risks when LLMs directly control business logic. I generally recommend leveraging LLMs for their strengths—understanding context, classifying intent, and generating human-like text—while keeping core decision-making in more deterministic systems where reasoning about behavior is more straightforward. That said, with appropriate monitoring and fallbacks, some organizations may find this trade-off acceptable for specific applications where the benefits outweigh the potential downsides.

Definition rating: conceptually distinct and appealing to theorists, but perhaps too restrictive for the messy reality of current implementations

Choosing the Right Definition for Your Context

Each of these definitions of “Agentic AI” serves different purposes, from marketing to technical implementation. Understanding which one is being used — or should be being used — in a particular context is crucial.

If you’re discussing AI systems with colleagues, clients, or stakeholders, take a moment to clarify which definition you’re using. Are you talking about LLMs that can take actions (Broad Agency), LLMs that can choose which actions to take (Narrow Agency), or something else entirely?

The terminology around AI continues to evolve rapidly, and “Agentic AI” is just one example of how the same term can carry significantly different meanings depending on who’s using it. Being precise about these distinctions isn’t just semantic nitpicking — it’s essential for building systems that actually deliver on their promises and for having productive conversations about where this technology is headed.

Whatever definition you choose, make it explicit. Your future self — and your stakeholders — will thank you.