Setting up AI in a service business is three jobs in one order. Decide what should stop depending on you, write down how that work is actually done, then give a tool the context to carry it. Most people start at the third and it fails, because a tool given no context automates a bottleneck instead of removing it.
Nearly every owner who asks this question has already tried. They opened Claude or ChatGPT, asked it to do something real, got back something that read like a press release, and quietly decided the technology was not ready yet.
The technology was fine. It had no idea what the business was.
That gap, between what a tool can do and what it knows about you, is the entire job. Everything below is how to close it, in the order that works, with the part most people skip marked clearly.
The reason is not prompting skill. A model with no information about your business produces the average of everything it has ever read, and the average of all business writing sounds like a press release. It is doing the only thing available to it.
Underneath that sits a harder problem, and it is the one that decides whether any of this sticks. The context a tool needs is the part of your business nobody has written down. It was never written down because you have been the one holding it, and holding it has worked. You are the fastest, most accurate place to get an answer, so every question routes to you and nothing gets documented.
So the setup fails, and the conclusion is we tried AI and it did not stick. Which is the same conclusion these businesses reached about the last three systems they bought, for the same reason every time. The systems were not the problem.
AI does not remove a bottleneck. It runs whatever you point it at, faster. Point it at an undocumented process and you get an undocumented process running faster, with less oversight.
Not one thing. Four, and they stack. Miss one and the whole thing stays a toy.
Most people build the second layer and wonder why it does not hold. A skill with no brain underneath it is a clever prompt, and a clever prompt is something you have to be present for.
The fourth layer is the one that gets treated as optional and should not be. Sending, publishing, paying, deleting and changing permissions all get a human in front of them. Approve the action, not the session. Yes, send this one is a control. Yes, you can use email today is a blank cheque. More on what to connect and what to hold back.
Three steps. Each one is useless without the one before it, which is the whole reason setups fail when they start at the end.
Not what could be automated. What should stop reaching you at all. This is a decision about your own role, and until it is made, every system you install gets quietly routed back through you.
Not how you think it is done. The version that comes out of a week of real questions rather than an hour of memory. Three questions to run before you automate anything.
Now it has something to work from, and the output stops being generic because it stops being generated from nothing.
A reader who has not made the first decision will not get value from the third. That is not a philosophy. It is the mechanism by which every previous attempt failed.
Write the context file.
Six sections, in full sentences rather than bullet fragments, because fragments are ambiguous and a model resolves ambiguity by averaging, which is the problem you started with.
If you do only one of those six, do the last one. A description of your voice is a hypothesis. Three examples of it are evidence, and it changes the output more than the other five put together. The full walkthrough, with the file to fill in.
Two tests, and neither of them is whether you liked the output.
Ask it a question only somebody who works at your business could answer.
If the answer is right, the file works. If it is plausible and wrong, the file has a hole and you have just found exactly where it is.
Name one thing that used to reach you and no longer does. One specific thing, by name.
If you cannot, nothing was set up. Something was bought.
The second test is the one that matters, and it is the reason the order above is not negotiable. A setup that produces better drafts while every decision still routes through you has changed your output and not your position.
Both work. They suit different situations, and the honest version is that the choice is about time against urgency, not about capability.
There is a third option nobody offers, and it is the one that sticks. Do the mindset and the system yourself, and have the technology built.
It costs less than a full build, and the person who wrote the process is the person who understands the system afterwards. A build handed to someone who did not do the thinking gets abandoned the first time the business changes shape.
Honest numbers, in time rather than money, because time is the constraint that stops people.
That last line is the one to sit with. The build is the short part. The long part is you changing how you work, because a system only gives you time back if you stop doing the thing it took over. Most of the delay in this work is not technical.
Everything above, run against your business rather than in general. Describe what you have now and this works out which of the four layers you are missing, which step you skipped, and the single thing to do first.
It takes about fifteen minutes and you need to be honest in the first three answers, because the rest is built on them.
No. If you can describe how your business works out loud, you can do the setup. The hard part is not technical, it is deciding what should stop depending on you and writing down how the work is actually done. Nothing in the first two steps involves code.
Pick one and go deep rather than sampling four. The setup work, your context file, your written processes, your decision rules, transfers between tools in an afternoon. The tool is the cheapest part of the decision and the easiest to change later.
Because the previous attempt almost certainly started at the tool. A system that depends on knowledge living in one person’s head will always route back to that person. Write the knowledge down first and the same tool behaves completely differently.
It depends on what it can reach and what it can do unattended, not on the tool. Give each connection the minimum access the job needs, keep anything irreversible behind an approval step, and never let it both read outside text and send things out unsupervised.
It ranges from your own time to a five-figure install. Doing the first two steps yourself costs nothing but afternoons, and it is worth doing regardless, because it is what makes any paid option work. See the programs and services pages for what each option includes.
If you have read this far and nothing has been written down yet, start with the context file this afternoon. It is the smallest step that produces a visible change, and it makes every later step cheaper.
If you already have a setup and it is not giving you time back, the problem is almost certainly step one rather than step three. Run the one-month test above and name the thing that stopped reaching you. If there isn’t one, go back to the beginning.
Every step above has a guide that goes further than this page does. They are all in the guide library, and the books, tools and habits actually in use are on the resources page.
Thirty minutes, no pitch, and we will tell you which of the three steps your business is actually stuck on.
Book a call →Each one picks up where the last leaves off. Start wherever you actually are.
Teach it your business once, in about thirty minutes. The context file, your voice, your boundaries, and a prompt library you actually reuse.
Thirty-two skill and agent files that already do the job, plus everything new we release while you are a member. Load one and it runs the same afternoon.
One complete system, end to end. Take the blueprint and install it yourself, or have us build it around you in about six weeks.