Agentic AI · 4 September 2026
Is ChatGPT an AI agent?
No. ChatGPT answers your question and then waits for you. An AI agent goes and does the job, changes something real on your computer or in your accounts, and then tells you what it changed.
An email will land in your inbox this month offering you an AI agent. Maybe one already has.
Before you answer it, work out whether they are selling you an agent at all. Very often they are not, and when they are, there are 4 questions worth putting in the reply. And almost everybody I talk to about this describes the same evening. They nodded along in the meeting, then looked the word up on their own afterwards.
I use these every day. One of them builds and runs my website. So instead of defining the word, I will show you a job I gave it this morning. By the end you will tell the two apart in about 10 seconds.
AI assistant vs AI agent, in one line
the version people actually usePeople who use these every day settle on one short line. A chatbot responds. An agent acts.
That holds up until somebody asks "acts how?" So here it is one level further down, in the version that survives the follow-up question.
You paste, it reads, it waits
You keep three supplier invoices in a folder. You open ChatGPT and ask which one is unpaid. It cannot see your folder, so you upload the files.
It reads them and answers well. Then it waits. You copy the answer out and save it yourself.
It opens the folder itself
You point an agent at the same folder and ask the same question. It opens the files, works out the answer, writes it into a new file, and tells you what it did.
When it stops, the folder holds something that was not there before.
What one of these actually did this morning
one run, my machineI did not want to describe this from memory, so I ran it before writing this.
I dropped three files into an empty folder. Two came from a timber supplier and looked nothing alike, because I saved them on different days. The third came from an ironmongery, which sends a monthly statement rather than an invoice.
One filename carried a space and a lowercase month, the way a real folder looks rather than the way a demo folder looks.
Then I typed one sentence in plain English.
> Look at the files in this folder. Tell me what I spent with each supplier in August, and whether anything is still unpaid. Write the answer into a file called august-summary.txt as well. What I put in it
- ▸invoice_northgate_0812.txt
- ▸BRIGHTWAY - aug statement.txt
- ▸receipt-scan-0823.txt
question
What it left behind
- ▸invoice_northgate_0812.txt
- ▸BRIGHTWAY - aug statement.txt
- ▸receipt-scan-0823.txt
- ▸august-summary.txt
26 seconds. One sentence typed. A 4th file that nobody created by hand, holding both supplier totals and £233.28 named as still unpaid.
The new file listed both suppliers, gave their August totals, and named the one invoice that had not been paid.
Then it did the thing that made me keep the run instead of deleting it. Underneath the numbers it had added a note of its own.
The two Northgate totals include VAT at 20%. The Brightway statement lists no VAT line at all, so its 741.04 may be net of VAT and not directly comparable.
Nobody asked it to check that. It read the two documents and saw that the suppliers format them differently. It added the numbers anyway, because I asked for a total, and then told me where its own answer might be wrong.
That is the whole difference. A chat window answers the question just as well. It does not go and find the files, and it does not leave the summary where the invoices live.
Three different things share the name
answer · workflow · agentA third thing sits between the two, and skipping it is why these conversations go in circles. Inside companies I see the same pattern almost every time. The team asks for an agent, then describes a workflow.
Which one does your job actually need?
| What you are choosing | An answer | A workflow | An agent |
|---|---|---|---|
| What it does | Tells you, you do it | Same steps, same order, every time | Decides what to do next, then does it |
| Does it decide anything? | No | No | Yes |
| Cost to set up | Nothing | Write the steps down once | Highest of the three |
| How it fails | You notice at once | It stops, and you notice | It carries on quietly |
| Good for | Most questions | The job you repeat weekly | Messy input, real decisions |
Most of the jobs people describe to me are workflows wearing an agent's name. That matters for a practical reason.
An agent costs more to build and more to run, and it fails in a way that stays quiet, because something that decides can decide wrongly and keep going. A workflow that breaks just stops, and you notice.
You can do better than picking one, and this is what I do.
Where a job runs the same way every time, with no real exceptions, I hand that part to a plain program. It runs faster, costs almost nothing, and gets it right more often, because a program cannot take a shortcut.
Then I bring the AI in only where the job needs judgement. That combination costs less and holds up better than either one on its own.
Why everything gets called an agent
both sides of itThe word signals independence, and buyers want to hear independence. So sellers put it on things that only answer questions, which is why what the word means when a small business uses it is worth pinning down before you walk into the meeting.
Do not treat the person emailing you as a liar, though. Two very different people send that email.
Building this properly is hard
Making a real agent work inside somebody else's business takes weeks. They have to learn your information, find your exceptions, and test them one at a time.
Anyone charging real money for that charges correctly. If they show you what it does on your own numbers, talk to them.
A new word to sell an old thing
Every new technology attracts people who sell the word rather than the thing. AI is not special here. The same crowd arrived with apps, with cloud, with blockchain.
They give themselves away the same way every time. They cannot tell you what it does when you are not watching.
The question to take into the meeting
one sentenceWhat does it do when nobody is watching?
They say somebody types a question and it comes back with a good reply. Useful, often the right thing, and it should carry an assistant's price.
They say it runs the same steps on a schedule. Cheaper than an agent, and far easier for you to check.
They say it looks at what it finds and decides what to do next. Now ask the second question: what happens when it decides wrongly at 2 in the morning?
Use reversibility. Sending money, emailing a customer, deleting anything: those wait for your approval. Everything else runs on its own.
You need no idea how any of it works to ask that. That is the point of it.
What I would actually do
my honest recommendationLearn to do it yourself. That is my first recommendation and I mean it plainly.
Nobody understands your work the way you do. Somebody else will build you a very good first 80%, far faster than you could. The last 20% is every exception your business actually runs on.
That last stretch is where the money goes, and the two do not line up the way people expect.
You will also learn it far more easily than you expect. Nobody is asking you to take a 4 year engineering degree.
It works more like getting used to a new colleague. You learn how they think, what they do well, and where they cut corners. After that you know how to work with them.
Start with Anthropic's own free course, Claude Code 101. It costs nothing, and the people who built the thing wrote it. Whether you ever buy anything from me or not, start there.
After that, if you want the whole vocabulary rather than the three words on this page, there is my AI Agents course. Two hours, no code, and no assumption that you have ever opened a terminal.
The whole thing on one card
if you read nothing elseChatGPT answers and waits. An agent goes and does the job, then tells you what it changed.
Shape 1
An answerIt tells you, you do it. Most jobs stop here, and that is fine.
Shape 2
A workflowSame steps, same order, every time. Decides nothing. Never forgets.
Shape 3
An agentDecides what to do next, then does it. Costs most. Fails quietly.
What does it do when nobody is watching? If it only answers, they are selling you an assistant at an agent's price.
Reversibility. Sending money, emailing a customer, deleting anything waits for your approval. Everything else runs on its own.
Give the repeatable part to a plain program and bring the AI in only where the job needs judgement. Less money, fewer mistakes.
Learn it yourself. Start free with Claude Code 101, and give one real job of your own an afternoon.
Where do you stand?
15 short situations about handing a job to an AI tool. In each one you pick what you would do next.
- ▸Whether you sized the job right
- ▸What you handed over
- ▸How you checked what came back
- ▸When you stopped
About 5 minutes. You see your score, your weakest area and the reasoning behind every answer before anybody asks you for an address.
Questions people ask next
answered in one line eachIs ChatGPT an AI agent?
No. ChatGPT answers your question and then waits for you. It cannot open a folder on your computer, change a file, or send anything for you unless you connect it to something that can. An AI agent does the job itself and leaves a change behind.
Is an AI agent the same as a bot?
A bot follows a fixed script. Somebody wrote the steps and it repeats them. An agent looks at what it finds and decides what to do next. That is why an agent handles a messy job that would break a script, and also why you have to check it.
Is Copilot agentic AI?
Some of it. When Copilot suggests a sentence or answers in a chat panel, it acts as an assistant. When you hand it a job that runs across your files and it makes the changes itself, that part is agentic. The name on the product does not tell you which one you get.
Is Claude agentic AI?
Claude in a chat window is an assistant. Claude Code, which runs on your own computer and can open and change files in a folder you point it at, is agentic. Same company, same model underneath, two different things.
What is the difference between an AI agent and a chatbot?
A chatbot responds. An agent acts. One level down: a chatbot gives you an answer that you then act on yourself, and an agent does the acting and reports back on what it changed.
Do I need an AI agent for my business?
Often not. If you need an answer, an assistant is enough. If you need the same steps done in the same order every time, that is a workflow, which is cheaper to build and easier to check. Pick an agent only when the job needs somebody to decide something while it runs.