Past the chat window: four steps of using AI at work
Most businesses stop at the first step. Here is what the next three look like, and how to tell which one you are on.
Ask the owner of almost any business whether they use AI, and the answer is yes. Ask them to show you, and they open a chat window. They type a question, paste in a document or a price list, read the answer and copy the useful part back out into an email, a quote or a spreadsheet.
That is a real step, and I'm not going to tell you it's wrong. But in my trade we went through the same change over the last few years, and the chat window turned out to be the first of four steps. Each one moves a little more of the carrying off your plate.
Step one: the chat
You ask, it answers, and you are the one who connects it to your business. Everything it knows about your company, you typed or pasted in that morning, and everything it makes, you copy out and put away.
You can tell you are on step one when you find yourself explaining the same things over and over: how you price, which parts go together, how you write to customers. The AI forgets all of it the moment you open a new chat.
Step two: the chat that remembers
Most chat apps now let you keep standing instructions and a few files that come along to every conversation. You write your rules once and upload your price lists once, and every new chat starts from them.
This is a big improvement for very little effort. But the work still happens inside the chat. The quote it writes is still text in a window, and you still download it, rename it and move it to where it belongs.
Step three: the AI works in your folder
This is the step my trade took, and it's the one I'd like you to reach. You give an agent, which is an AI that can open and change files, one folder on your own computer. Your rules sit in a plain file in that folder, your price lists sit beside it, and the quotes it writes land in it.
Two things change. Nothing travels up into a chat and back down again, because the work is done where your files already are. And when you correct it, it can write the correction into the rules file, so the correction is still there tomorrow. My guide on rules in a file shows that in detail.
I use Claude for this. In the Claude desktop app it is called Cowork, and it needs a paid plan. Other tools can do it too; what matters is the move, not the brand.
Step four: the work starts without you
At the last step you are no longer the one who starts each job. A request that arrives is turned into a draft quote, a weekly report is put together from the week's files, and the agent leaves its work in your folder for you to check.
I'd only take this step once step three has run well for a while, because by then your rules are written down and tested. Software teams learned this the hard way: you automate what you already do reliably by hand, never the other way round.
Whichever step you're on
Treat the AI like a junior employee. You tell it things and you teach it things, and it is very good at remembering what is written down. But it is still a junior: it makes mistakes, so you check its work, and every mistake you find becomes the next rule.
You don't need to jump to step four. Moving one step up is worth doing on its own, and each step makes the next one smaller.