AI for independent professionals: staying the expert in the room as the tools get smarter

I use AI all day, and most of the people I work with would not know it.

It is in my writing, my research, and the running of my practice. It drafts articles and social posts, summarises what I read, follows the people I learn from across Substack, newsletters and X and brings what they publish to me, and pulls the SEO reports.

It builds a briefing document before I meet someone, sets up the client folder when a booking comes in, and manages the list. It runs through connected tools that reach into my calendar, my email, and my client records. I have a strong model for heavy thinking, a quick one for queries, and an agent that drives the tools. I cannot write code, and it has written working programs for me.

It has made the practice bigger than I could run on my own, and most of the time, no one can tell the difference. That's the point.

Stop treating it like a faster spreadsheet

Most advice about AI gets one thing wrong, and that's what matters.

It treats the tool as a faster version of what you already have. A quicker typewriter or a spreadsheet that talks. The advice is always the same: you do the thinking, the tool adds the speed.

That's a mistake, and I think it's an expensive one. AI is not software in the way your other tools are software, but closer to a mind – a co-intelligence, and a strange one at that.

You cannot tell from the outside which tasks it will be brilliant at and which it will let you down. It writes a sonnet easily and cannot reliably write a poem of exactly fifty words. It is very good at idea generation and weak at basic arithmetic. There is no instruction manual, because no one knows the full range of what it can do, including the people who built it.

The only way to learn the shape of it is to use it a lot, and that keeps moving as it gets stronger every few months. What it could not do for you in March, it can do in September.

Ethan Mollick, who studies this full time and co-authored the research I am about to describe, puts it this way: assume this is the worst AI you will ever use.

Is it good, or is it just fast?

Billy Oppenheimer works as a researcher to Ryan Holiday, and he has a four-word test for the whole thing. When someone asked him what question about AI and creativity is not being asked enough, he answered: is it actually good, or is it just fast?

Speed was never what made the work good. Oppenheimer points to a sociologist named Daniel Chambliss, who spent years watching swimmers at every level, from country-club squads to the Olympic team. Chambliss found that the best swimmers did not do more of what the rest did, faster. They did the small things differently, and they did them on purpose. Excellence, he concluded, is qualitative rather than quantitative. Not more of the same thing, but the same thing done differently.

AI is very good at the quantitative part, at producing more and doing it faster. The qualitative part, the small things done differently on purpose, is the part you've spent your career learning. That is where the value lives, and the tool doesn't, as yet, come close.

Information is cheap. Understanding is expensive.

Karl Fast, the information architect who wrote Figure It Out with Stephen Anderson, makes the same point from the other side. Information is cheap, he says. Understanding is expensive.

Those two sentences form the economics of what an independent professional sells. The internet has driven the cost of information to nearly nothing. You can pull up a thousand answers in seconds.

What has not got cheaper is turning those answers into understanding: working out which one applies to this client, this problem, this moment, and what to do about it.

Fast breaks that work into four moves.

1. Foraging is gathering the raw material. 
2. Tuning is working out which parts matter.
3. Externalising is getting the thinking out of your head and onto a page or a board so you can see it.
4. Constructing is building the understanding itself, the part where the pieces become a position you would stand behind.

AI is very strong at foraging, and weak at constructing. It gathers and sorts at a scale you cannot match. The constructing, the building of understanding, is the part that happens when a person has skin in the outcome. That is the part you sell.

And there is a warning folded into the same idea. Fast argues that when the interaction gets too easy and too quick, the understanding goes shallow. He reads paper books and marks them with sticky flags instead of digital highlights for exactly this reason. The small friction is what forces the synthesis. Take the friction away and you are left with information that has moved through your head without ever becoming understanding.

Where my understanding actually lives

Fast's point is that the mind does not stop at the skull. It extends out into the tools, the space, the people around you. That's not theory to me. It is how my practice actually runs.

There is constant creation. I am always writing notes, and I keep a set of skills I return to and fine-tune, short written pieces that hold how I do a thing so I do not have to work it out from nothing each time.

I use Obsidian as the vault that holds it all: the notes, the wiki entries I build and keep updating, the articles, the interview transcripts.

Readwise brings me the highlights from what I read. It is all connected and searchable, so when I sit down to write or to prepare for a client, I am not starting from whatever is in my head that morning. I am starting from a system that has been collecting and connecting for years.

That is the extended mind in operation, and the AI sits inside it. It helps me forage the material and keep it in order. It summarises a transcript and pulls out the part I need. But the constructing, the deciding what it all means and what to do with it, still happens where it always has: in the conversation between me and the notes.

This is why experienced professionals get more out of AI than anyone else. The tool is good on its own. It is better inside a system someone has already spent years building to think with.

The evidence is not subtle

The study that settles this ran 758 consultants at Boston Consulting Group, roughly seven per cent of the firm, through realistic work for a fictional shoe company.

The consultants who used AI finished more than twelve per cent more tasks, about twenty-five per cent faster, and their work was graded more than forty per cent higher in quality than the work of those who did not. Read that last part again. It is not just speed. The work itself got better.

The gain was largest at the bottom, and that's the part everyone quotes. The below-average consultants improved by forty-three per cent, the above-average by seventeen. The tool is a leveler. It lifts the person with the least experience the most. Mollick's comparison is the steam shovel: how good a miner was at digging used to matter a great deal, right up until it did not matter at all.

But look at the other half of that finding, the half that gets skipped. The above-average consultants improved too. Everyone got better, top included.

That is the whole argument in two numbers. The tool raised the floor, and it raised the ceiling. The reason the experienced professionals got the most value is that they could tell good output from generic, steer the thing toward something useful, and know when to push back.

Why this is your moment

There is a second finding in the research that gets far less attention, and it's the one that matters most for you.

AI bites hardest at the entry level. The junior tasks, the research summaries, the first drafts, are the first things handed to the tool, because they are the easiest to automate. Those are exactly the jobs that used to train the next generation of professionals.

When that work disappears into the machine, the pipeline that produces experienced judgment narrows with it. Fewer people are coming up behind you who have done the thousand small jobs that built your judgment. The trap is not the tool, and it is not a future problem.

That sounds like a warning about the young, and it partly is. But for you it lands as opportunity. What you have is not just hard to copy. It is getting harder to replace, because fewer people are being trained to hold it. The scarcity moves toward you.

Then there is a finding that I think is genuinely reassuring. The researchers behind the augmentation trap proved something specific about when that trap actually binds.

It binds when one person decides how much AI someone else uses. A manager who wants the output today and does not feel the cost of a dulled skill two years on will overuse the tool. But when you control your own use, that particular trap has nothing to bite on, because you feel both sides of the bargain yourself: the time you save now, and the skill you would lose later.

That's the structural advantage of independent practice, and it is a real one. The trap needs a decision-maker who is not the worker, and you are both. That does not cover you against rust, which comes from your own habits.

Two ways to work with it

When Mollick and his co-authors looked at how the best performers actually worked, they found two patterns, and both beat working alone.

The first is the cyborg. You weave your thinking with the tool's. You write half a sentence and let it finish it. You push back on its word choice, ask what a line means, make it try again until the sentence is yours. The tool is inside the thinking, not downstream of it.

The second is the centaur. You divide the work cleanly. You do the thinking, and you hand the mechanical parts over. Drafts, formatting, sorting, scheduling.

Both work. But when you read the descriptions, most independent professionals I talk to are cyborgs, whether they have the word for it or not.

Peter Hatherley, who founded an AI company in Christchurch and joined me on episode 54, treats the AI as a co-architect. It asks whether you really want this, and it works the detail with you.

Margareta Krizova, a lifelong entrepreneur near Prague who joined me on episode 90, does not ask the AI for answers. She gives it the work of a businessman she admires and asks it to ask her the questions she would not think to ask herself.

Anna Burgess Yang, a writer who joined me on episodes 65 and 77, runs an automation that drafts posts from her blog, then rewrites every one into her own voice before it goes out.

Filip Drimalka, who joined me on the podcast, expresses it like this: most people, he says, use AI as a muscle, a way to do more of what they already do. 

He wants to use it as a brain. This is my goal, this is my intent, these are my data, and I am the one who approves. He is happy for the tool to do ninety per cent, ninety-five, even ninety-nine. The part that matters is that the last decision stays his.

All four are doing the same thing. They treat the tool as a thinking partner and stay in the conversation. The work gets better, and so do they.

The move that changes the output

There is one practice that separates the people who get good work from AI from the people who get the average, and it's way simpler than it sounds.

The tool's default answer is the average answer. That is what it is built to produce, the most likely next word. So the way to get anything better is to tell it who to be before you ask.

Mollick co-wrote his book this way. He didn't hand it a draft and ask for checks. He built a small team of personas, a pompous critic, a dreamy connector, a median reader, and ran his drafts past them to be challenged. That's not handing the thinking to a black box. It is thinking out loud with a room full of sharp people who just happen to be made of code, or a robot if you prefer.

You can do the same with one line.

• You are a sceptical client who has heard this pitch twice
• You are the median reader of my newsletter 
• You are the tough conversation I am about to walk into.

The role shapes everything that comes back.

This matters more for you than for most, because you have spent a career learning what a good client, a good reader, a good objection actually looks like. That is exactly the material you can hand a model to make it work better for you. The more you know, the better the brief.

Oppenheimer watched this happen in slow motion for years. Ryan Holiday has spent twenty years working the craft of writing, and when he prompts an AI, what he thinks to ask is different from what someone else would ask. The result is better, and it is better because of the twenty years, not because of the prompt. The tool does not level that difference. It pays for it.

Do what you could not do before

The question most people ask is the wrong one. They ask, could I still do this without the tool? And if the answer is no, they pull back, worried they are building a dependency.

That question would have kept me from writing the scripts I use every day, because I cannot code. It would tell a non-programmer not to build an app. It is a rule against ambition, and ambition is the whole point of the tool.

Mollick tells his students to make their projects ambitious to the point of impossible, because they have AI. The non-coder should build the app. The person with no design background should ship the landing page. Your range of what you can do is supposed to get broader when the tool fills the gaps.

Drimalka says the most important skill now is curiosity, the willingness to play with the thing and find out what it can do. He also says AI is the first technology that can teach you how to use it. Both point the same way: stop asking what you already do that it can speed up, and start asking what it opens that was closed before.

The right question is not could I do this without it. It is what can I do now that I could not do before it.

What the client actually hands over

That question asks what you keep. The deeper question is what the client hands over to you, and why. The answer is reliability. Clients hand work over when they trust someone to decide whether it is fit for their situation and stand behind the result.

A self-directed system such as GPT-6 Astra can take a goal stated briefly and imprecisely, work for days without derailing, and return finished work containing hundreds of decisions you never watched it make. (just be careful of the cost!)

The middle of the professional's job starts to disappear. What remains is a short specification at the front and a decision at the back: does this work deserve to go out under my name? The rules covering people who act on your behalf now have to account for machines that do, because you can be answerable for decisions you never made.

The value sits in the gap between work that succeeds ninety-eight per cent of the time and work you are prepared to be personally answerable for. Clever output is plentiful. The last two per cent is checking the loose ends, deciding whether the result fits this client and this situation, and being the person who stands behind it.

As machines absorb more of the middle, small firms can take on work that once required layers of people to hold those loose ends.

Evaluation is being handed over as well. AI systems are already scoring working papers they did not create, using several models to reach a combined assessment. Remembering everything is becoming the starting point. The greater value is knowing what deserves attention before the client has to ask.

Once AI works on its own, your remaining value narrows to that last two per cent of trust: judging work you did not watch get made and accepting responsibility for what happens when it is used.

The one thing that actually goes wrong

None of this means AI is risk-free. There is one real cost, and it is worth naming because everything else about the tool is yours to gain.

The risk is not that you use it. It is that you fall asleep at the wheel.

The researchers found it in the same study. Outside the areas where the AI was reliable, the consultants using it were nineteen percentage points less likely to be right than those working alone, because they stopped checking.

Fabrizio Dell'Acqua, one of the co-authors, found the same thing among recruiters: those with the best AI made worse hiring decisions than those with a weaker tool or none, because they leaned on the recommendation instead of their own read. He called it falling asleep at the wheel.

There is a faster version of the same thing. A fluent, complete-seeming answer can feel like understanding, whether it is right or not, and that feeling stops you checking before you have thought to check. Three researchers gave it a name, Epistemia.

And there is a slow version, the one you are most likely to feel, and it needs to be identified.

When you offload the thinking over the months, your own judgment slowly loses its edge. The researchers call it intuition rust. The clearest evidence is a year-long study of cancer specialists. The doctors who leaned on an AI decision tool got faster at first, and over the year their independent diagnostic skill declined. The tool was doing the seeing, and their own seeing went soft.

Filip Drimalka caught it in himself, and he is as committed to AI as anyone I have met. He noticed his writing had lost something, so he stopped letting AI write his LinkedIn posts and went back to writing them himself, to rebuild the skill. He did not stop using AI. He stopped using it for the one thing where he needed to keep his own hand in.

So yes, rust is real. I want to say that straight rather than hide it. It is the one genuine cost in the bargain, but I think it's manageable.

The fix for all of it is the same, and it is not to use the tool less. It is to stay in the loop. Let the tool generate a hundred and twenty-two uses for a toothbrush, then use your taste to pick the two that matter. The value you bring is the expensive, calibrated judgment that selects, edits, and steers. That is not doing all the thinking yourself. It is the part of the thinking that only you can do.

The four places your work goes

Strip away the tools and the frameworks and it falls into four places. I use this with clients, and Filip walked through it on the episode. The skill is knowing which place a task belongs in, and refusing to let the volume leak into the first one.

First, the work that is only yours. Everything you have been through is the qualification: your career, your life, the people you have learned from, and the judgment calls that only make sense because you were there. Your understanding of the client whose real problem doesn't fit the one they described. This is why anyone pays you.

Second, the people who know what you do not. Collaborators, subcontractors, fractional specialists who bring their own judgement to the part you cannot or should not do yourself. You are not outsourcing to save money. You are buying expertise you genuinely lack.

Third, the volume. First drafts, research, transcripts, scheduling, the forty pages you need the shape of by Thursday. The work where speed and pattern-matching win. This is where the AI goes, as augmentation rather than a substitute for the thinking.

Fourth, the community. The peer groups and professional communities that give you an outside view of how you are organising your own work. Nothing gets invoiced out of this pile. It is how you notice the other three have drifted.

The first pile is the one that got more valuable as the other three got cheaper, and it is where you decide what is good enough to carry your name. Filip explains all four places in his own words on the hundredth episode (out now), and it is a useful picture to keep in mind while the tools keep getting better and better.

A last thought

I want to be direct about where I stand on this, because the caution in this piece is easy to over-read.

I use AI more than most people I know, all day, across my whole practice, and I wish I had started sooner. The work is better, the reach is wider, and the parts I used to dread are gone. The professionals I talk to who get the most from AI all have the same thing in common: they went all in, and they stayed in the conversation.

When Mollick describes the people who get the most out of AI, he is describing how I work now. And the honest balance is this. Rust is a real risk, and I have stated that up front. The reason it has not caught me is that I stay in the loop. I edit every line, and I reflect on what comes back, asking Grok, one of many models I use, to weigh in before anything gets filed away. The tool gives me more to work with. It does not do the deciding.

The tool gives me more to work with. It does not do the working. Without AI, I could not be doing what I am doing now. The scripts I run, the briefs I prepare, the reach I have, none of it happens on my own. That is a trade I make willingly, and I would make it again.

If you are holding back, you are leaving the biggest part of the gain on the table. Your experience is not something to protect from the tool. It is the reason the tool works better for you than for anyone else. The question worth asking is not whether you can still do your work without it. It is what you can build now that you could not build yesterday. Keeping your mind sharp as a business asset is the longer conversation. If you would like help working out how AI fits in your own practice, start with the free assessments.

Frequently asked questions

Fabrizio Dell'Acqua, Ethan Mollick, et al. - "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality" (2023). The BCG study of 758 consultants: more than 40 per cent higher quality, 25.1 per cent faster, below-average performers up 43 per cent and above-average up 17 per cent, and 19 percentage points less accurate outside the frontier. Also the source of the Cyborg and Centaur distinction.

Billy Oppenheimer - interview with Alex Dobrenko, "Inside the research mind behind Ryan Holiday and Rick Rubin", The Sublime Newsletter, 8 November 2025. The source of "is it actually good, or is it just fast?", the Chambliss swimming study, and the observation of Ryan Holiday's twenty years of writing craft changing what he asks of an AI.

Daniel F. Chambliss - "The Mundanity of Excellence: An Ethnographic Report on Stratification and Olympic Swimmers", Sociological Theory, 1989. Excellence is a qualitative, not a quantitative, phenomenon (as cited by Billy Oppenheimer).

Karl Fast and Stephen P. Anderson - "Figure It Out: Getting from Information to Understanding", Rosenfeld Media, 2020. The source of "information is cheap, understanding is expensive" (Fast, World IA Day, 2012), the four epistemic interactions (foraging, tuning, externalising, constructing), and the point that interaction which is too easy and too quick produces shallow understanding.

Caosun, M., and Aral, S. (2026) - "The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading", arXiv:2604.03501, MIT Sloan School of Management. The source of the augmentation trap and intuition rust, the slow erosion of professional judgment when thinking is offloaded to AI, and the finding that the trap binds only when the decision-maker and the worker are misaligned, so a professional who controls their own AI use is the case the trap cannot reach.

Ehsan et al. (2026) - the year-long study of cancer specialists whose independent diagnostic skill declined under AI decision support, as cited by Caosun and Aral. The concrete picture behind intuition rust.

Quattrociocchi, W., Capraro, V., and Perc, M. (2025) - "Epistemological Fault Lines Between Human and Artificial Intelligence." The source of Epistemia, the displacement of a person's own evaluation by a fluent, complete-seeming AI answer.

Fabrizio Dell'Acqua - "Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters" (2022). The recruiter study showing that leaning on high-quality AI for hiring judgment degrades it.

Nate - "Executive Briefing: How to Check Work You Didn't Watch Get Made", Nate's Substack, 2026. The source of the "last two percent of trust" and the point that once AI is self-directed, the professional is answerable for decisions they did not watch get made.

Georgetown AI Referee Paper Leaderboard - a live leaderboard scoring working papers with an ensemble of LLM referees. A working example of AI evaluating work it did not create.