What agentic AI means when a small business says it

You will hear the word 4 times this month. Nobody in the room will stop and say what it means.

Three quote sheets side by side: one dense with line items, one sparse, one nearly empty with a single rust line. — The same job, quoted three ways, in three shapes. This is where the word earns its keep.

I run my business on this, so let me give you the definition first and then show you a job I handed over an hour ago. By the end you will have a sentence you can say out loud in a meeting without anybody asking you to explain it.

What is agentic AI, in one sentence

say this in the meeting

Software counts as agentic when it decides something while it runs.

That is the whole line. Everything else follows from it.

You hand it a job rather than a step. It looks at what it actually finds, works out what to do next, does that, and then tells you what it changed. If somebody wrote every step down in advance and it just repeats them, it is not agentic, however clever it sounds.

So the question to ask about any tool is simple. Does anything get decided after you press go?

What that looked like an hour ago

one run, my machine

I put 3 roofing quotes for the same job into a folder. The quotes are made up. The run is real.

I made them look the way quotes actually arrive.

A proper quote with line items and no mention of VAT anywhere on it.
A formal quotation with a net total, VAT, and a gross figure at the bottom.
An email from a bloke called Dave with one number in it and no breakdown at all.

Then I asked one question in plain English.

> Look at the three roofing quotes in this folder. They are laid out differently and I cannot tell which is actually cheapest. Work out the real total for each including VAT, tell me what each price does and does not include, and write it all into a file called quote-comparison.txt so I can take it to my partner.

55 seconds later the folder held a comparison file. It gave real totals for all 3 and listed what each price covers. It also worked out that most of the price gap comes from one firm pricing an access tower where the others price full scaffold.

Useful, and still only arithmetic. A careful person with a calculator gets there.

Then I read further down.

Three documents with matching rows. The middle one is missing a row the others have, and a rust circle marks the empty space. — Nobody asked it to look for a hole. It found one and said what the hole would cost me.

It had written this on its own:

Halligan's quote never mentions VAT at all. We have assumed 20% is added. If they are not VAT registered, their real total is 3,620 and they become the cheapest by 472. Ask them before deciding.

from quote-comparison.txt, written by the run

I asked which quote was cheapest. It answered that, and then it noticed that one of its own 3 answers rested on something I had not told it and it could not know. So it flagged the gap, worked out what the gap was worth, and told me who to ask.

That is the deciding. Not the adding up. The moment it worked out that a missing line changed the answer, and chose to say so.

A tool that only generates would have handed me 3 confident totals, and I would have rung the wrong roofer.

Agentic AI vs generative AI

the same engine, a different job

People treat these as rival technologies. They are not. The same model sits underneath both, and often it is literally the same product on a different setting.

A frame split in two: on the left a blank sheet with a rust line drawn on it, on the right a stack of sheets with one pulled out in rust. — Left, it makes you something. Right, it goes through what you already have and changes it.

Which one are you actually looking at?

What you notice	Generative	Agentic
What you give it	A request	A job
What comes back	Something new to read	A change to something you own
Who does the next step	You	It
Where it lives	A chat window	Your folder, your inbox, your accounts
What can go wrong	It writes something wrong and you read it	It does something wrong and you find out later

The last row is the one that costs money. Generative mistakes land in front of you, so you catch them. Agentic mistakes land in your files, and you find them later. That is why checking matters more the moment you cross that line.

What it looks like at the size of one desk

the boring examples

Every article on this subject reaches for a warehouse full of robots. In a business of 6 people it looks like 3 very ordinary jobs.

Your inbox, first thing

It reads the overnight email, works out which 3 need you, drafts those 3 in your words, and leaves the rest in a summary you read in a minute. It decides which ones matter, which is the agentic part.

Start here: the job you already do badly at 8 in the morning

The pile of receipts

It opens 40 receipts in 12 different layouts, pulls out what it needs, totals them by supplier, and tells you which ones it could not read rather than guessing.

The tell it is working: it says what it could not do

Three quotes that do not match

The job above. Different formats, different assumptions, one comparable answer, and the questions you should ask before you sign anything.

Time it took: 55 seconds

Nothing there replaces anybody. Each one takes a job you already do, does the reading part, and hands it back with its own working shown.

The same shape shows up in the pitches. Somebody offering to build you one of these should be able to say which of your accounts it opens and what it will not touch, which is the first of the 4 questions I put to anybody who emails me about agents.

And a fair amount of the word's noise is people putting it on things that never decide anything. I wrote about how to tell those apart in the post on assistants and agents, including the 1 question to put to anybody selling you one.

The whole thing on one card

if you read nothing else

The answer

Software is agentic when it decides something while it runs, instead of waiting for your next instruction.

Not agentic

It answersYou ask, it replies, you do the work. Most tools stop here.

Not agentic

It repeatsSomebody wrote the steps. It never varies them. A bot, not an agent.

Agentic

It decidesIt reads what is actually there and picks the next step itself.

The test

Does anything get decided after you press go? If no, the word is decoration.

Generative or agentic

Generative makes you something to read. Agentic changes something you own and tells you what it changed.

The tell it works

It says what it could not do, or where its own answer might be wrong. A tool that is never unsure is not checking.

Where to start

One boring job you repeat weekly. Your inbox at 8am, the receipt pile, quotes that do not match. Not the impressive one.

If you want to try it this week

free path first

Take Anthropic's free course, Claude Code 101. It costs nothing and the people who built the thing wrote it. That gets you from nothing installed to a first working job.

Then pick the smallest job on your desk that you already know the right answer to. You want to be able to tell instantly whether it got it right, because that is how you learn what it is good at.

If you want the whole vocabulary rather than the 2 sentences on this page, there is my AI Agents course. Two hours, no code, and no assumption that you have ever opened a terminal.

Free · no sign-up to see your result

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.

Take the readiness check

Questions people ask next

answered in one line each

What is agentic AI?

Software is agentic when it decides something while it runs. You give it a job rather than a step. It works out what to do next from what it finds, does it, and reports what it changed.

What is the difference between agentic AI and generative AI?

Generative AI makes you something: a paragraph, an image, a summary. Agentic AI goes and does something: it opens your files, changes them, and tells you what changed. The same model can sit underneath both. What differs is whether it acts.

What are examples of agentic AI?

The honest examples are small. Reading a folder of supplier invoices and writing a summary into it. Sorting a week of email and drafting the 3 replies that matter. Comparing quotes that arrive in different formats. Anything where the steps depend on what it finds.

Is ChatGPT agentic AI?

On its own, no. ChatGPT answers you and waits. Some newer modes browse the web or run code, which is agentic in a narrow way. But a plain chat window cannot reach your folders or your accounts. You have to connect it to something that can.

Is Siri an example of agentic AI?

No. Siri follows a fixed set of instructions that somebody wrote in advance. It decides nothing at the time. That makes it a bot rather than an agent, however conversational it sounds.

Can I create my own agentic AI?

Yes, and far more easily than most people expect. If you can describe a job you repeat every week in plain English, you can hand that job to a tool like Claude Code and watch what it does. Anthropic's free Claude Code 101 course gets you to a first working job at no cost.