Every conversation I have about AI agents ends up at the same place. We end up talking about how AI agents work. Usually linking them back to the tools or software someone is already using. And about how much time they can save.
The business owner nods along.
The examples land, the quote chasing, the invoice nudges, the Monday report. They can see how it can help them in their own week in it. And then they say some version of the same sentence:
“Sounds brilliant. But I’m not technical. This isn’t something I could actually have, is it?”
Yes, it is. And this post exists to prove it.
I’ve been building and coding automation since 2007.
Long before any of it had “AI” in the name.
And I can tell you the truth about AI agents: the concepts are simple.
There are only three moving parts. Once you understand them, the mystery evaporates, and so does most of the fear.
Let’s “look under the hood”.
How AI Agents Work: The Three Moving Parts
Every agent ever built
From a simple invoice chaser to something more sophisticated. There are 3 parts to any AI agent.
Part 1: The Trigger (What Wakes It Up)
An agent doesn’t sit there thinking.
It sleeps until something specific happens. That something is the trigger.
Triggers come in two flavours:
Something happens.
A new enquiry lands in your inbox.
A quote gets marked as sent in your CRM.
An invoice hits seven days overdue in your accounts package.
A client fills in a form on your website.
It’s an action that happens somewhere inside your process that kicks the AI agent into life.
A time arrives.
Monday, 7 am, time for the weekly report.
Every evening at 6 pm, check for quotes that have gone three days, and you haven’t had a reply yet.
That’s it.
That’s a trigger.
You already think in triggers, but you call them something else. “When a quote’s been out a few days, I chase it.” The first half of that sentence is the trigger.
You’ve been running these rules in your head for years.
All we are doing now is writing them down.
Part 2: The Brain (What Decides)
This is the AI bit; an LLM does some thinking or processing.
The same kind of intelligence that powers ChatGPT or Claude.
When the trigger fires, the brain looks at the situation and works out what to do.
Not from a rigid script but from understanding the needs of a task and what the current situation is.
Take the enquiry that lands at 7 pm.
The brain reads it. It works out what the situation is.
Is this a genuine enquiry or spam? Are they asking about a service we actually offer? Is this a small job or a big one? Have we dealt with this person before? What are they actually asking, underneath how they’ve phrased it?
Then it makes a decision.
Reply now with times for a call, or ask a qualifying question first. Or, if something’s odd about it, flag it to get a human in the loop and do nothing.
This is the part that separates agents from old-fashioned automation.
Automation from ten years ago followed scripts. If email contains “quote”, send template B.
One unexpected phrasing and it fell over, or worse, sent something stupid.
The brain reads context the way a decent member of staff would.
That’s the leap that’s happened in the last couple of years, and it’s why this conversation is worth having now. This level of intelligence is now available to us all. And it plugs the gaps that “traditional” coding solutions have never been able to.
The game has changed, for the better.
Part 3: The Hands (What Acts)
A brain with no hands is a chatbot.
It can tell you what it would do, or what you should do.
The hands are what let the agent actually do it.
The hands are connections into the tools you already use.
Your email and calendar. The CRM your business runs on. Your accounts package, Xero, QuickBooks, Sage, whatever you run. The project management system.
These connections are how the agent sends the follow-up from your email address. Logs the chase against the right client in your CRM. Reads the overdue invoice list from your accounts, and books the call into your real calendar.
And here’s the thing most owners don’t realise: these connections already exist.
The tools you use every day have standard, secure ways of talking to other software. The same plumbing that lets your card machine talk to your accounts package.
Nobody’s hacking anything together.
We’re plugging into doors the software companies built on purpose.
And even over the last few years. A mass adoption in a standard framework called MCP has happened. Which allows these companies to make it even easier for your AI agents to talk to their systems.
The barrier to entry is almost non-existent.
So that’s how an AI agent works. 3 distinct parts.
Trigger. Brain. Hands.
These 3 three things make up how AI agents work.
Something wakes it, something decides, something acts.
Every agent, no matter how impressive. Is those three things arranged around a job, process or task.
One Agent, End to End
Let’s walk the quote follow-up agent through a real week.
So you can see the three parts working together.
Monday, 2 pm.
You send a quote to a client.
£4,800 for a job. You do nothing else; you don’t tell the agent. It’s connected to your email and your CRM, so it saw the quote go out.
A clock starts ticking.
Thursday, 6 pm.
The trigger fires.
This quote has gone three days with no reply.
The brain wakes up and checks the situation. Has the client emailed about anything else since?
No.
Is the job marked as won or lost in the CRM?
No.
Right — this needs a chase.
The brain writes the follow-up. Not a template. It references the actual quote, the job, the figure. The thing the client said they cared about when they enquired.
It’s written in your tone, because it’s learned from the emails you actually send. Short, polite and to the point.
The hands take over.
The email goes out from your email address. The AI agent logs the chase in the CRM against that client, with a timestamp.
A new clock starts for the second follow-up.
Friday, 9:15 am.
The client replies.
“Thanks for the nudge, been a mad week. Can we book a call?”
The brain reads the reply, recognises it as a positive response, and does two things.
1) Stops all future chasing on this quote, and flags the conversation to you as warm.
2) If you’ve allowed it, it suggests 3 call times to your client using your calendar.
Friday, 9:20 am.
You see only one thing.
A warm reply and a call scheduled in your calendar.
The noticing, the deciding, the writing, the sending, the logging, the remembering. You don’t need to do any of it. There’s been ZERO interruption to your day.
Now multiply that by every quote you send in a year.
The Myths — Let’s Kill Them.
“I’d need a developer on staff”
No. You build these AI Agents on workflow platforms.
The one I use is a tool named n8n. But there are several established ones out there. And new ones are coming online all the time.
Think of them as digital Lego.
The connections to your email, calendar and accounts are pre-built blocks. And the skill is in arranging them around your processes and rules.
That’s a job for someone who does it all day (that’s me), not a reason for you to hire an IT department.
Once it’s built, nobody technical needs to be in your business.
Your team sees the results: emails sent, calls booked, reports arriving.
“It’ll go rogue and email something embarrassing”
This is the fear underneath most hesitation with using AI agents.
So let’s take it seriously.
An AI agent built by a professional has guardrails.
Hard limits it cannot cross. It only does the specific job it was built for.
The invoice chaser cannot email your client list about anything except invoices. It has approval steps wherever you want them.
For the first few weeks, every email it wants to send sits in your drafts folder waiting for your sign-off.
You read them, you send them, you learn to trust it, then you take the stabilisers off, at whatever pace suits you.
And everything gets logged. Every action, timestamped. You can see exactly what it did and when, any time you like.
Compare that to your current setup.
Where follow-ups depend on whoever’s least busy. Whoever remembers. In whatever mood they’re in that day.
The AI agent is the more consistent, more auditable option, not the riskier one.
“It won’t sound like me”
It learns from your real emails.
How you open, how you sign off, whether you’re a “Cheers” or a “Kind regards”, how direct you are.
Clients can’t tell. And during that approval phase, every draft you tweak teaches it.
By the time it’s running solo, it’s had the same induction a new office hire would get.
I did this myself the other day. I gave Claude 3 emails that I sent to clients that week, and it learned my writing style.
It can be that simple.
“It’ll break, and I’ll be stuck”
Now you’re justified with your worries about this one.
And the honest answer is that AI agents. Like anything connected to other software, need looking after.
The tools they plug into update and change.
That’s why every AI agent we build for our clients gets added to our monitoring system. So we can maintain it for a client. I won’t work with someone if they don’t want this level of service.
Then something needs attention. It’s known about and fixed before it costs you anything.
Usually before you’d have noticed.
I covered how that works, and what it costs, back in the running costs part of how I work.
The point for now: “who looks after it” has a proper answer, and the answer isn’t “you.”
“What about my data?”
Depends on your business.
For most service businesses. AI agents running on mainstream infrastructure with sensible security is absolutely fine.
It’s the same standard your accounts package already meets.
But if you handle sensitive client data. And the idea of any of it touching a big provider’s servers makes you uncomfortable.
There’s a simple solution.
Run the whole thing, the brain included, on your own infrastructure.
What This Costs to Run
I’ll keep this short, because Part 4 breaks down the costs in more detail.
A working agent, built, connected to your tools, monitored and maintained. Costs less per month than one day of staff time per week.
That’s the frame worth holding onto.
You’re not buying software. You’re buying 2–3 hours a week of reliable, never-sick, never-on-holiday capacity. For less than the cost of one human day.
Against what it costs to recruit right now. If you can even find someone.
I’ll let you do that maths yourself.
What You Need To Know About How AI Agents Work
You don’t need to understand servers, and you never will.
What I want you to take from this post is this:
Everything an agent does is a rule you already follow.
Chase at day three. Nudge invoices at day seven. Reply to enquiries fast. Send over the numbers on Monday.
You know these rules; you’ve been carrying them around in your head for years. Executing them by hand at 9 pm on a Friday, having a beer at the kitchen table.
Building an AI agent is taking one of those rules out of your head and giving it a trigger, a brain and hands.
Which raises the question: How do you actually get one?
How do you pick what process to give to an AI agent first? What happens during the build, how long does it take, and what does it cost?
That’s Part 4.
And if you want a head start. My Time Freedom Audit takes five minutes. And maps which of your weekly tasks are the best candidates that you can hand off to AI agents.
Do it before Part 4 lands, and you’ll read it with your own first AI agent already in mind.
Download the Time Freedom Audit here.
Next week: Part 4 — Getting Your First AI Agent Live: What It Actually Looks Like.
