If you run the day in an accounting practice, an AI employee learns your business from what your office hands it: the job written out the way it's done now, examples of finished work, and one person who decides. After that, it learns from your corrections. Nobody on staff has to become a trainer.
An AI employee learns your practice from the pages your website is missing
Your website still lists something you stopped offering. The last news post is from a tax season nobody remembers. And the questions people ask before they call get answered on the phone, one at a time, by whoever picks up. Whether you take on bookkeeping. What a rental property needs at tax time.
An AI employee can take that job. Picture one whose only work is the pages on your website, the ones that answer what people search for before they call you. Call her Nadia. She writes those pages from the notes and examples you give her, and in your practice she'd report to the person who runs the day, which is usually you. In my own business the same job is split across two AI employees: one works out what my pages should say and publishes them, and the other does the writing. Both report to me.
Nadia is judged on a number you already watch: how many enquiries come in through the website.
She isn't there to take anyone's work. Those pages are the ones nobody has had time to write, and they're the reason you're answering that question for the fifth time this month.
An AI employee learns from three things your office hands over before the build
Before I build anything, I ask the practice for three things, in this order. The job written out the way it's done now. Examples of the finished work. And one named person who decides what the AI employee does on its own.
None of the three is a course, and that's where a search for AI training goes wrong. Type it into Google and everything on the first page is about teaching people on staff to use AI tools. Here it runs the other way. The one doing the learning is the AI employee, and what it learns from is the way your office already works.
For an accounting practice, the job written out is the practice described the way you'd describe it to a client. That means the work you take on, the work you send elsewhere, and who your clients are. Examples of finished work could be a page you're proud of, or the email you send a new client the week they sign up. And one name, the person Nadia reports to, who says yes or no when she can't tell whether something is right.
There is one more piece to a build, and it isn't one of the three, because it's my job and not yours. Getting a way in to the website and the inbox is mine to set up. You shouldn't have to know what a connection is.
An AI employee needs the rules your office has never written down
Most of what your office knows has never been written down anywhere. You know which clients ring instead of emailing. You know the question you get every January that the website has never answered. And you know the work you turn down.
I find most business owners already run rules like that, out of their heads and out of their sent items. The person who runs the day carries most of them. Nadia can learn from both, but only once they're handed over.
So don't start from a blank page. Start with the answers you've already written: the email you send when a new client asks what you'll need from them, the paragraph you keep retyping. The rules no email shows go into the written job as plain sentences. Things like "We don't take on bookkeeping" or "Ask about the rental property before you quote a fee."
Say it the way it's done, not the way it ought to be done. If your office asks about the rental details before it quotes a fee, that goes in, even if the binder says otherwise. A rule nobody hands over is a rule Nadia can't follow, and the person who ends up explaining it to a client is you.
An AI employee learns from corrections, and one person decides how far it goes
Once Nadia is writing, she learns from you reading her work. She sends a page. You read it, change a sentence, or send it back. The correction goes into the instructions she works from, so the next page doesn't make the same mistake twice.
After the build those instructions are written in plain words, so you can read them and tell me what's wrong. I write the instructions for my own setup the same way. Giving an AI agent a soul file is weirder than it sounds is the account of how that record gets written.
What Nadia can't do is a short list. She can't do physical work, so she can't go through the paper files on your shelf. She can't work inside a program she can't connect to, so if your practice software won't let her in, that page waits. And she can't use the phone.
Everything else is a choice, and it belongs to the business owner. Can Nadia publish a page without you reading it first? Can she answer a question a client sends through the site? Those choices start tight and loosen as she earns it, and the person she reports to is the one who decides.
Four tests that show an AI employee has learned the job
You can tell an AI employee has learned the work when it passes four tests, and you can check each one yourself.
It starts the job. Nobody has to tell Nadia what to write next. The questions you keep answering turn up as pages, written and waiting for your read.
It finishes the job. A page doesn't sit half done. It goes up, or it comes to you with a note saying what's missing and why.
It owns one number. Enquiries from the website is Nadia's number, and you can watch it climb on the same report you already check.
It reports and escalates. When a page needs something she doesn't have, like a fee you haven't given her or a service question, she comes to you instead of guessing. The report sums up work you could have read any day, because every page Nadia writes sits where you can open it.
Book a 30-minute Let's Chat AI call, and bring the job an AI employee would learn
If you'd like an AI employee writing the pages your practice is missing, book Let's Chat AI. It's 30 minutes. Bring a page you're proud of and one of the questions that comes up before anyone books, and we'll go through what it would take to learn your practice.