TLDR: AI can reach almost every task in an expertise business now, which is why so many people automate the wrong ones. This piece sorts your work onto three rungs using one test, then covers the part most advice skips: the judgment you protect still fades, because you automated the work that kept it sharp. The fix is a log. To see where your line sits, take the free AI Readiness Diagnostic.
A few years ago, AI took my income.
I was a freelance writer. Clients who used to pay me for articles stopped, nearly all of them, inside a few months. I spent weeks wondering if I had built a career on a skill with no future.
So I learned the tool that replaced me. I rebuilt my whole process around it, got faster, and started earning more than before.
And in the middle of all that, I nearly handed over the one part that made any of it worth paying for.
New here? I’m James. I help founders and creators build AI systems that make them impossible to replace.
This piece is the sorting method I use: which work to hand over completely, which to let AI draft while you keep the call, and which to never touch. Plus why your judgment fades even when you protect it, and the one-minute habit that stops it.
For the foundation underneath this, read The Only Moat Left In The AI Age. For how other operators are running this in practice, read how 20 AI-powered entrepreneurs actually work.
A few years ago this question answered itself. “What to automate with AI? Well, everything that it can automate!”
And that usually meant “the boring work.” But, you automated them because the boring work was all the machine could reach.
Now, it can reach far more into you. It reaches your judgment. It drafts your proposals. It scores your leads. It writes in something close to your voice and suggests the things you would have suggested.
The ceiling is gone. And while most people are treating that as good news, a few are cautious. I'm one of them. While I use AI extensively, I automate conservatively. And I do that by sorting every task in my business onto a ladder with three rungs.
Automation starts from the bottom.
The limit is up to you. Your red line is the highest rung you let it reach. Here are mine.
The bottom rung is everything a client never sees and never credits you for.
Medicine has the cleanest example. AI scribes listen to the appointment and write up the note. A study across six health systems in JAMA Network Open found real drops in documentation time, mental load, and burnout.
AI took over the transcript work. It never took the diagnosis.
Mine look like this:
Client calls get recorded and turned into notes and a task list I never type.
Contract drafts, timelines, and standard client emails come out of templates I built once.
One finished article becomes about fifteen pieces of platform content, and I never touch the reformatting.
Nobody hired you because your invoices arrived in a beautiful font. If a client could not tell whether you or a machine did it, hand it over and stop touching it.
The mistake down here is doing too little. If you are still doing this work by hand because you never got around to it, you are paying yourself minimum wage out of your own margin.
Most consultants underestimate how much of this is already buildable:
A system that reads your calendar, preps you before each client meeting, and nudges the client to actually show up
Meeting recordings turned into notes, next steps, and a draft follow-up email waiting in your drafts folder
An intake form that creates the client folder, the CRM record, and the kickoff checklist the moment someone signs
A searchable archive of every call you have ever had, so you can find the thing a client said in March
None of these touch a decision. That is why they are safe to build for your own business.
The middle rung holds most of the money … and most of the danger. The steps stay the same. The right answer does not.
Mine look like this.
Every client we work with has a project holding their public business information, their marketing, and their goals. I can ask questions of that pile.
This client keeps hitting the same wall on LinkedIn, so give me three approaches nobody has tried. I pick the one that fits the person. Sometimes none of them fit, and that tells me something too. For research, I might feed in hundreds of comments and transcripts and ask what patterns are there. The machine finds them. I decide what they mean.
Other versions an expert could build:
A proposal drafter trained on your last ten won proposals, that stops before scope and price
A prep brief before every session: last session’s notes, open threads, a suggested opening question
Lead scoring that sorts and flags, and never replies to anyone without you
A content engine that pulls ideas from the questions clients actually ask you
An assessment system that reads, scores, and communicates with your people about a specific problem you’ve already solved 100x before
Let the machine lift. You keep the choice, the judgement.
The trap on this rung is comfort. Harvard Business School and Boston Consulting Group ran the biggest study we have on this, with 758 consultants.
On tasks the model was good at, the ones using AI did more work, faster, and better. On the one task the model was bad at, the ones using AI did about 19% worse than the people working alone. Fabrizio Dell’Acqua, who led that part of the research, calls it “falling asleep at the wheel.”
The answer looked confident, so people stopped checking. Even when there was a human in the loop. The draft is usually good enough. Good enough is the sound the trap makes when it closes.
Another study: a 2026 Harvard Business School working paper read the chat logs of more than 70 consultants who worked with LLMs for business problems. As they worked on validating the model’s initial findings with fact, their models did not back down. It apologized, corrected itself, then restated the same wrong answer with more detail and cleaner reasoning. The researchers describe professionals getting talked into it by the thing they were supervising. (Interestingly, one of the writers, Prof. Karim Lakhani, has a Substack, also about AI! Find it here.)
Granted, Gen AI has moved fast, and this paper was published only last year (2025). But I think the larger point still stands: even if you were paying attention, even if there was a human in the loop, you can still make the wrong, AI-influenced decision without proper judgement.
Slowing down does not fix this. Neither does watching harder. What works is putting your judgment inside the build, so the system cannot finish the job without you.
The systems I build does that three ways.
It stops at fixed points. The system runs at full speed up to the stop and then sits there. No nudge, no timeout, no default that fires if I ignore it.
My standards also live in a file the system reads on every run, instead of a prompt I retype when I remember. On a bad day, the file still holds the bar.
And it hands me options with the reasoning attached, so I am choosing between things instead of approving one thing. Approving takes a second. Choosing makes me look.
None of this makes the system slower. Everything below the stop still runs while I sleep. The only thing I added was a place where the work has to pass through me.
That is also what makes it hard to copy. Someone can buy the same tools and clone the same flow. They cannot clone where you put the stops, because you placed them using the judgment they do not have. Your decisions become part of the infrastructure.
The tell that you have slipped is boredom. When the review stops feeling like a decision and starts feeling like a keystroke, that task already moved down a rung without asking you.
Everything on this rung comes down to one line. Never automate the point where your work meets a person.
Everything before that point is machinery. The point of contact is the product.
For a coach, this point might be the session, and the silence right after someone finally says the true thing. For a consultant, it is the moment you tell a client what you actually think, and then stay in the room while they deal with whatever good or bad news you got. For an advisor, it is the call where someone decides whether to trust the recommendation.
Automate everything around those. Never automate it.
For most expert businesses the list looks like this:
The session itself, and the read on what a client is avoiding
Delivering bad news, and the work of fixing it with them
The call on whether a prospect is a fit, and the decision to turn one away
Every reply, comment, and DM you send to your own audience
Asking for a testimonial, or pricing one specific job
None of that is efficient, but that’s kind of the whole point. Human contact is inefficient, but it’s the most valuable.
These are also the moments that get more of you every time you automate a rung below well. The hours you take back do not disappear. They go here.
Mine covers every point where a client meets me.
The first call. The progress updates. The call when something breaks and we are working out the fix. I write my own emails and I try to make them warm, because that email is the relationship, and nothing is underneath it if I hand it over. Same rule for my brand. I do not automate relationship building, so I built Daily 10, a relationship habit builder that puts ten people in front of me every day. It runs the queue. It has never written a comment for me.
A system can decide who you talk to, when, and about what. It cannot be the one talking.
Automate this rung by accident, and people do, and you change what you sell. You stop selling a person who knows and start selling an output anyone with the same tools can make.
And the price of a commodified output only moves one way.
Deloitte Australia showed the fast version. Last year, the firm delivered a 237-page review to a government department for around AU$440,000, and had to refund the government back.
A researcher at the University of Sydney checked the sources. The report cited academic papers that did not exist and quoted a federal judge saying something no judge had said. Deloitte refunded the final payment. The department said the analysis and recommendations still stood, which is the whole lesson. The thinking survived (inasmuch as it survived after having suffered a shameful, public, blow). The sourcing did not, and the sourcing was the part carrying the firm’s name.
But, you can guard the top rung perfectly and still lose it.
In 1983 a psychologist named Lisanne Bainbridge published a paper called Ironies of Automation.
Automate most of a process and the operator keeps only the parts that could not be automated. Their job turns into watching. They stop practicing the skills the job used to build. Then something goes wrong, a person has to take over, and that situation is hard by definition, so it needs more skill than the old job did.
They have less.
Your judgment works the same way. It is a practice, not a possession.
So sorting your ladder correctly is not enough. Going back to doing the grunt work will not fix it either. Making your own decisions visible to yourself will.
So every time you override one of your systems, write three lines. What it recommended. What you did instead. What you saw that it missed.
Ninety seconds. No tool, no template. One running document.
That log does three separate jobs:
It gives you the rep back. Naming the reason is the practice that watching takes away, and it costs a minute instead of an afternoon.
It shows you which overrides repeat. Anything you have written four times is a rule you can now state, which means it just moved down a rung on purpose instead of by accident. Feed it to the system as a check. The decision still stops with you.
It piles up into a record nobody else has. Not what you know. What you decided, when, and against what advice.
A year of that log is what you teach from. It is what you hand a new hire instead of a slide deck. For a business made of expertise, it is about as close to inventory as you get.
Make this the question you ask before you build anything:
Which rung is this task on?
Try to write down the rule you follow when you do it. If you can finish the rule, all the way to what finished looks like, it is bottom rung. Hand it over and stop touching it. If you can write the steps but not the right answer, it is middle rung. Let the machine draft, and keep the decision. If you cannot write it down at all, or if it is a point where your work meets a person, it is top rung. Leave it alone.
Then draw the line on purpose. Every time, on purpose. Set it too low and you drown in work you should have handed off years ago. Set it too high and you give away the thing that made you worth choosing.
The question today is not “What should I automate with AI?”
It’s “What should I not?”
PS. I built a short diagnostic that shows where your line probably sits and which rung you are most at risk of crossing. Takes a few minutes, and it is free.
Find your line: take the AI Readiness Diagnostic
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.