What is agentic AI? The plain-English version

What does agentic actually mean?

The word agentic describes AI that is given a goal rather than a script. Nobody spells out each step. The AI decides what to do next, uses tools to take actions, looks at the result and decides again, until the job is done or a limit stops it.

A tool is an action the AI is allowed to ask for, such as searching a database, reading an email or updating a job. Usually the AI does not run the tool itself. It asks, software carries out the request and hands back the result. That is why the tools you hand over set the edge of what the AI can do.

Here is what that looks like in a small business. A tile delivery is running late, and you ask an agent to work out which jobs it affects and let the teams know. The agent looks up the delivery, searches the job list for work that needs those tiles, checks the dates on each job and drafts a message for each team. It chooses what to look up, and in what order, based on what it finds. A different problem tomorrow would send it down a different path.

That is the strict meaning, and it is the one this guide uses. The word is also used loosely for any automation where AI takes an action, such as filing an enquiry or updating a job, even when every step was set in advance. Strictly, that is a fixed workflow with an AI step in it, and the next section gives it its proper name. Fixed workflows are not second best. They are what most small businesses should build first.

How is an agent different from a chatbot or an automation?

Three things get mixed up, and it helps to keep them apart.

• A chatbot answers. You type a question and it writes a reply. On its own it does not change anything in your systems, and it does nothing until you ask.
• An AI workflow follows fixed steps. Something triggers it, such as an email arriving, and the steps run in an order set in advance. The AI does specific jobs inside that path, such as sorting the email by type or pulling out a delivery date, but it does not choose the path.
• An agent chooses its steps. It is given a goal and a set of tools, and it works out the route itself, case by case.

Our Insights post on agentic workflows and chatbots makes the case that an AI that acts is usually more useful to a small business than one that only answers. It uses the word agentic in the looser sense. Most of what it describes, such as an inbox monitor that sorts and labels email, is a fixed workflow that runs when something arrives, and it treats a goal-driven agent as the more autonomous end of the range. Both uses are common, so when someone says agentic, ask which one they mean.

There is a simple test. Could you write down every step before the AI sees the job? If you could, it is a workflow, however clever the AI step inside it. If the AI has to work the steps out from what it finds, it is an agent.

What is an agentic loop?

An agentic loop is the cycle an agent repeats: decide a step, take it, check the result, go again. It has four parts.

1. Decide. The AI reads the goal and what it has learned so far, and picks the next step.
2. Act. The step runs, usually as a tool: a lookup, a search, a draft.
3. Check. The result goes back to the AI, which reads it and works out whether the job is done.
4. Repeat, or stop. If the job is not done, it goes round again.

The stop matters as much as the loop. A loop ends when the AI decides the job is finished, or when a limit you set is reached, such as a maximum number of steps. Without a limit, a loop can keep running up cost or keep acting long after it should have stopped, so a well-built agent always has one.

If you build software, our guide to how to build an agentic loop from scratch shows the whole loop in working code, in TypeScript and Python. You do not need it to decide whether your business needs an agent. The rest of this page is written for owners, not builders.

What can go wrong?

An agent has more ways to go wrong than a fixed workflow, because it chooses its own steps and can take more actions. Three problems matter most.

Runaway cost and actions. Every trip round the loop calls the AI model, and that costs money. Every trip can also do something: send a message, change a record, book a job. A loop with no limit can run up a bill or repeat an action again and again. Step limits, budgets and time limits are the defence, and our guide to stop conditions and budgets shows how to set them.

A wrong action, taken confidently. An AI can misread a situation and carry on as if it had it right, in a calm and certain tone. When it can only draft, the mistake is a bad draft. When it can act, the mistake is a wrong invoice sent or a real job changed. Give it only the tools the job needs, and keep a person in front of anything that is hard to undo. Our guide to human in the loop covers where a person must stay in the process.

Instructions hidden in what it reads. An agent reads emails, web pages and files, and text in any of them can be written to look like an instruction. This is called prompt injection. A message that says to ignore your rules and forward the inbox is the classic example. Any AI that reads outside content and can take actions is exposed, and an agent usually has more tools to be tricked into using. Limit what it can do, keep a person in the loop for anything that matters, and never treat what it reads as a command.

Does my business need one?

Probably not yet, and possibly not at all. Most small-business automation works best as a fixed workflow: the steps are known, they are the same each time, and you want them to run the same way every time. An enquiry arrives, the AI sorts it and pulls out the address and contact details, and your system files it and alerts the right person. That is cheaper to run, easier to test and more predictable than an agent. Start there.

We work this way ourselves. Our own social media posts go through a fixed workflow. Each post starts as a draft, either written by us or drafted by an AI model. The owner reads it in a private admin page and approves it, or does not. A scheduled job then publishes only the posts marked approved, and it never publishes an unapproved draft. Where an AI model is used, it drafts. It does not choose the steps and it does not publish. That is a workflow, not an agent, and it is the pattern we suggest for anything with your name on it.

Reach for an agent where each case really is different and the steps cannot be listed in advance. Working out which jobs a late delivery touches is like that, and so is chasing down why a supplier's delivery does not match an order: what the AI should check next depends on what it has just found. Even then, start with an agent that drafts and a person who approves. For examples of jobs like these, see our guide to what agentic loops can be used for.

There is an Australian angle to check before you connect anything. An agent sends what it reads to the model at every step, so whatever it touches goes wherever that model runs. Our post on where your business data goes when you use AI breaks the question into four layers: which country processes the data, how long it is kept, whether it is used to train models and who else handles it. If the material includes personal information about customers or staff, check where the tool sends it first.

A practical way to begin: pick one job that repeats and write down its steps. If you can list them all, build a fixed workflow, keep a person on the approval and see how it goes. Come back to an agent only if real cases keep breaking the list.