"Agent" is the word of the year. Every software company has one now. Your CRM has an agent. Your accounting software has an agent. Someone on LinkedIn has probably tried to sell you one this week.
If you're quietly not sure what an agent actually is, you're in good company. Most of the people using the word aren't sure either.
So let's clear it up. No jargon, no hype.
Two very different things
There are two types of AI you'll come across in a business.
The first is the assistant. It's the one you talk to. You open ChatGPT, Claude or Gemini, type a question and it answers. You go back and forth. It's brilliant at explaining things, drafting things and thinking a problem through with you.
The second is the agent. It's the one that does the work. You give it a job, and it goes off and gets it done using your actual tools, then comes back and tells you how it went.
An assistant gives you an answer. An agent gives you a result.
Same job, done both ways
Take a job every business owner hates: chasing unpaid invoices.
With an assistant, you ask it to write a polite chaser and it gives you a good one in seconds. Then you open your accounts, find who's overdue, copy the email, change the name, change the amount, paste it into your inbox and hit send. Twelve times. The AI saved you the writing. You still did the job.
With an agent, you say "chase anyone more than 14 days overdue". It checks your accounting software, finds the twelve customers and writes each one a message that fits: a gentle nudge for a good customer, something firmer for a third reminder. It sends them, notes it in your CRM and comes back with "Sent 12. Two people are disputing their invoice, so I've passed those to you."
Same technology. Completely different amount of your time.
What actually makes something an agent
Plenty of products calling themselves agents are chatbots with a new label. Gartner even has a name for it, "agent washing", and reckons only around 130 of the thousands of companies selling agentic AI are offering the real thing.
Here's a quick test. A real agent has five things:
- A goal, not just a question. "Get these invoices paid", not "write me an email".
- Tools. Access to the places work actually happens: email, calendar, CRM, accounts, your website.
- Permission to act. It can send, book, update and buy, not just suggest.
- A loop. It plans, does a step, checks the result and adjusts. If an email bounces, it looks for another address instead of giving up.
- Knowing when to stop. It knows what it's allowed to do on its own and when to come back to you.
If a product can't do numbers two and three, it's an assistant. Maybe a very good one, but still an assistant.
Think of it as a dial
This is the bit most explanations miss. You don't jump from "chatbot" to "fully automated" overnight. It's a dial for how much you let the AI do on its own, and you can set it differently for every job. Try it.
Most businesses should start in the middle. The agent does the legwork, and you approve before anything goes out. As you start to trust it, you turn the dial up for the boring, low-risk jobs and keep it down for anything involving money, unhappy customers or your reputation.
What agents are good at today
Agents work best right now on jobs that are:
- Repetitive, with steps you could write down for a new starter
- Spread across several tools, where most of the time goes on copying from one to another
- Easy to check, so a mistake gets spotted before it costs you
That covers things like qualifying new enquiries and booking calls, chasing invoices, keeping your CRM up to date, pulling a weekly report together from five different places, answering routine customer questions and handing the tricky ones to a person, or researching a prospect before a meeting.
They're still shaky on anything that needs real judgement, taste or a relationship. They can also be confidently wrong, which is exactly why the dial matters.
Plenty of projects fail, too. Gartner expects more than 40% of agentic AI projects to be scrapped by the end of 2027, mostly because costs climbed and nobody could show the value. In my experience that usually happens when someone starts with the technology instead of a job that needed doing.
Where this is heading
Agents that work for you
One agent, one job, inside your own tools.
Agents that work together
A small team of agents passing work between them like a department.
Agents dealing with agents
Your customer's agent shortlists suppliers. Your agent answers it.
Now: agents that work for you. One agent doing one job inside your own systems. This is where most businesses should be playing today.
Next: agents that work together. One researches, one writes, one checks, and they pass work between them like a small department. Gartner expects at least 15% of day-to-day work decisions to be made by agents by 2028, up from zero in 2024.
After that: agents dealing with other agents. Your customer's agent finds suppliers, compares prices and books. Your agent answers it. A lot of buying will start with a machine doing the shortlist.
That last one matters even if you never build an agent yourself. If an agent is doing your customer's research, your website has to make sense to it. I wrote about that in Selling to Machines.
Where to start
Don't start with "we need an agent". Start with the job.
Pick one task that eats hours every week, has clear steps and is easy to check. Write the steps down as if you were training a new starter. If you can do that, an agent can probably do it too, with you approving everything at first.
If you can't write the steps down, the problem isn't AI. It's the process.
Assistants make you faster. Agents give you your time back.




