Chat AI vs Agentic AI: What AI Agents Really Do

It’s Tuesday morning. A customer emails asking if you can fit them in next week. You open ChatGPT, paste in their email and ask for a friendly reply. Thirty seconds later you have a good draft.

Then you copy it into your inbox, check your calendar for free times, fix the date it got wrong, send it, add a note to your CRM and set yourself a reminder to follow up.

The AI took 30 seconds. Everything around it took ten minutes. That gap is the whole difference between chatting with AI and agentic AI.

Chat AI: a clever adviser that can’t touch anything

When you use ChatGPT, Claude or Copilot as a chat, you start the conversation, you give it the context by pasting things in, it answers, and you do the doing.

That’s genuinely useful. It’s great for drafting, rewriting, summarising a long document or thinking a problem through. For plenty of jobs, it’s all you need.

The limits show up when the same job happens every day:

You’re the glue. Every task starts with you opening a chat and pasting something in.
It only knows what you give it. Your price list, your calendar and your customer history aren’t there unless you paste them.
Copy and paste is where mistakes creep in. The wrong date, the wrong name, the old price.
Personal information ends up in places it shouldn’t. The OAIC recommends, as best practice, not entering personal information into publicly available generative AI tools.

Some chat tools can now connect to your email or files. The question is still the same: who does the work, you or the AI?

Agentic AI: give it a goal, and it takes the steps

An AI agent is AI that’s connected to your tools and can decide what to do next. Anthropic, the company behind Claude, describes agents as systems where the AI directs its own process and the tools it uses. Australia’s cyber security agency describes them as relying on an AI model to understand the situation, then make decisions and take actions.

Here’s that same Tuesday enquiry, handled by an agent:

The email lands. The agent reads it and works out it’s a booking request.
It checks the job is something you do, in an area you cover.
It looks at your calendar and finds two free times next week.
It drafts a reply in your tone and sends it to your phone.
You tap approve. It sends the reply, logs the customer in your CRM and sets a follow-up for day 2 if they don’t answer.

You spent ten seconds instead of ten minutes. And it happened at 2pm, not when you finally got to your inbox.

The middle option most businesses actually need

There’s a third option between chatting and a fully independent agent, and it’s the one we build most often.

Anthropic draws the line clearly. Workflows follow fixed steps that are set in advance. Agents decide their own steps as they go. Their advice to developers is to find the simplest solution possible, and only add complexity when it’s needed.

In practice, a workflow with an AI step looks like this: the steps are always the same and you can predict them, but AI does the hard part in the middle, like reading a messy email or drafting a reply. It’s cheaper to run, easier to test and much less likely to surprise you.

And some jobs need no AI at all. A reminder that goes out three days before an appointment is a simple rule, and we’ll call it a quick win, not an AI project.

Side by side

	Chat AI	Workflow with an AI step	AI agent
Who starts it	You, every time	Something happening, like a new email or a due date	Something happening, or a goal you set
What it can see	Only what you paste in	The systems it’s connected to	The systems it’s connected to
Who decides the steps	You	You, in advance	The AI, as it goes
Best for	One-off drafting and thinking	Jobs that happen the same way every time	Jobs that vary too much for fixed steps
Main risk	Copy and paste mistakes, data in the wrong place	Breaks quietly when something changes	Does something you didn’t expect

One job, three ways: chasing an overdue invoice

With chat AI, you notice the invoice is overdue, ask ChatGPT for a polite reminder, then copy it into an email and send it. It works, as long as you remember.

With a workflow, every morning it checks Xero for overdue invoices. On day 7 it sends a friendly reminder, on day 14 a firmer one, with AI adjusting the wording for each customer. You approve before anything goes out.

With an agent, you set the goal: get overdue invoices paid without damaging the relationship. It reads the customer’s history, replies when they say “we paid last week” by checking the bank feed, and decides whether to nudge, offer a payment plan or flag it to you.

The agent is more capable. It’s also handling money conversations with your customers. For most businesses, the workflow covers what matters with far less to go wrong.

When an AI agent is worth it

An agent earns its place when:

The job varies a lot. A shared inbox where every email needs something different.
It needs judgement across several systems. Checking stock, the customer’s account and the delivery schedule before replying.
There’s enough volume that the set-up pays for itself.
Mistakes are cheap, or a person approves first.

Stick with a workflow, or keep a person doing it, when the steps are always the same, when money moves, when the information is sensitive, or when one mistake would be expensive.

Watch out for agent washing

“AI agent” is the label everyone wants right now. Gartner calls the problem agent washing: vendors rebranding existing chatbots and automation tools as agents. Of the thousands of vendors claiming agentic AI, Gartner estimates only about 130 are the real thing. It also predicts over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls.

Before you pay for an AI agent, ask:

What exactly can it do without me?
What can it access, and can that be limited?
What does it do when it isn’t sure?
Can I see a log of everything it did?
What will it cost each month at my volume?

If the answers are vague, it’s probably a chatbot with a new name.

How to use agents safely

In May 2026, Australia’s cyber security agency, ASD’s ACSC, published joint guidance with its US, UK, Canadian and New Zealand counterparts on adopting agentic AI. The main points translate well to a small business:

Start with low-risk, non-sensitive tasks, and give it more as it proves itself.
Give it only the access it needs. No broad access to sensitive data or critical systems.
A person approves high-impact actions before they happen.
Keep a log of everything it does, and have it stop and ask when it’s unsure.

That’s how we build. Everything runs in your own accounts, each job gets only the access it needs, and a person approves anything customers see.

What you can do this week

List five things you paste into ChatGPT every week. The ones you repeat are the ones worth automating.
Sort them. Same steps every time? That’s a workflow. Different every time? That might suit an agent.
Check what’s going into public chat tools. If staff paste in customer names, addresses or phone numbers, set a rule and give them a business tool instead.
Use the five questions before you buy anything labelled “AI agent”.

Sources

Building effective agents, Anthropic, December 2024
Careful adoption of agentic AI services, ASD’s ACSC and international partners, May 2026
Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, Gartner, June 2025
Gartner says agent washing is taking place, Staffing Industry Analysts (summary of the Gartner release)
Guidance on privacy and the use of commercially available AI products, OAIC

ChatGPT gives you an answer. An AI agent does the job. What's the difference?