Koehler Advisory.

AI is the tool that drives the cost of knowledge toward zero.

Think about why you pay for a doctor. Not for the consultation room. You pay for knowledge and experience that took years to accumulate and that you do not have. Same with a lawyer: you are buying the fact that they know the law and you do not. For most of history, knowledge like this was scarce, gated, and expensive, and that scarcity was the whole basis of the transaction. AI collapses the price of the knowledge itself. The answer a lawyer would have charged you for is now a question you can ask directly.

That sounds like it should erase a lot of jobs. It does not, and the reason is the most important idea in this note.

Free knowledge is not the same as applied knowledge.

Knowing the law is not the same as being accountable for legal advice. Knowing medicine is not the same as carrying the responsibility for a diagnosis. AI drives the cost of producing knowledge-work toward zero. It does not touch the ownership of it: the judgement about whether the answer is right, the accountability when it is wrong, the authority to act on it.

The lawyer’s afternoon of research collapses to nothing. The lawyer standing in court using everything they have accumulated to argue your case does not, and neither does their signature on the advice. What changes is not whether the role exists but where its value sits. It moves off the part that was really information retrieval and onto the part that was always judgement, enforcement, and trust. The work that survives is the work of standing behind an answer.

This has happened before.

“Computer” used to be a job. Before it was a machine, it was a person, usually a room of them, doing calculations by hand. The electronic computer did not delete those people. It moved them up a level, off the arithmetic and onto deciding what to calculate and what the answer meant. The spreadsheet did the same thing to accounting. It did not end the profession. It ended the manual ledger and turned the accountant’s day toward analysis and judgement.

AI is the same kind of shift. It does not erase the role, it relocates the effort, away from the part that was mechanical and onto the part that needs a person. The work does not disappear. It changes direction.

But there is a harder edge to this, and it is worth being direct about. AI will not replace you. A person using AI will. If a competitor or a colleague is doing the same job with these tools and you are not, they are simply faster, and the gap compounds. The point of adopting AI is not novelty. It is that it clears the mechanical work off your plate so your actual attention goes to the things that are genuinely worth your time. The people who win the next few years are not the ones who resist the tools or the ones who hand everything to them. They are the ones who use them to spend their hours on better problems.

How AI actually functions.

Most people use AI for exactly one thing: bringing the cost of knowledge to zero. How do I water this plant. What is the best aperture for portraits. They use it as a Google you can talk to, one that gives you the answer directly instead of making you hunt for it. That is genuinely useful and it is also the floor, not the ceiling.

A step up, AI stops being something you ask and becomes something you deploy. Specific tools handle the repetitive work that used to eat your week: summarising long documents, cleaning data, drafting the first version of something, processing the same task a hundred times. You are not asking a question any more. You are handing off a job.

A step beyond that, it does the work that used to require gated knowledge. Coding is the clearest example. Building something used to mean either learning to code or paying someone who could. Now you can describe what you want and get a working result, which means the ability to build is no longer locked behind years of training. The same applies to analysis, design, research. Capabilities that were specialist become things an ordinary person can simply use.

And then agents. Instead of one task, you hand off a goal and the system works through the steps to get there. This is where it stops looking like a tool and starts looking like a colleague.

The real shift is integration. Through things like MCPs and CLIs, AI connects directly into the tools you already use every day: your email, your calendar, your files, your codebase. At that point it is not a separate app you visit. It is a layer across your whole system, something that can act where your work already lives. The chat box was never the product. It was the demo.

So the question stops being “what can AI tell me” and becomes “what can I now build, automate, and own that I could not before.” Most people are still asking the first question. The leverage is all in the second.

Where I think the economics actually shift.

That is the individual picture. Inside a business the same logic holds, the stakes are just higher, and this is where I spend most of my time. So here is how I think about it.

I think the honest version of the strategy question is narrow. In most businesses there are three or four places where current models genuinely change the unit economics of an existing process, and a much longer list of places where they look transformative in a demo but break against the surrounding system. The work is to find the first list and ignore the second.

In my experience the high-leverage areas sit at the edges of a business rather than in its core: customer intake, content production, internal research, the long tail of operational work that was always too expensive to systematise. Core production systems, the things customers actually buy, are a slower and more careful conversation. The cost of a regression there is high, and the surface area for non-determinism is larger. I would rather move fast at the edges and stay patient at the core.

Where I think you should leave it alone.

The part rarely written about with any honesty is where AI does not belong. The current generation of models is impressive precisely because it produces plausible output in almost any domain, which makes it easy to mistake plausibility for fitness. A language model is not the right answer for every personalisation problem, every search problem, or every classification problem. Often a plain deterministic algorithm produces a faster, cheaper, more private, and more predictable result.

So the discipline I hold to is simple: use the simpler tool by default, and reach for a model only where rule-based logic cannot do the job, things like image understanding, open-ended language, nuanced ranking. What you want is a system that is largely deterministic, with a small surface of language-model behaviour at exactly the points where it earns its place. Most people do the opposite. They reach for the model first and bolt the logic on afterwards.

Why none of this stays still.

The harder problem is that none of these answers are stable. A choice that was right in the first quarter of the year can be obviously wrong by the third, because the price of a capability has collapsed, because a new model tier has moved the reliability frontier, or because the right pattern for a problem has changed entirely.

This is the part I think most people get wrong. They treat AI as a project, something you finish. I think it is an operating discipline, something you keep doing. Someone has to own keeping the architecture current: revisiting what was built last quarter against what is possible this quarter, and retiring the choices that no longer make sense. Without that, AI investments do not decay loudly. They decay quietly, and the cost surfaces much later as a fragile platform nobody wants to touch.

What AI is not.

Two honest caveats, because a note like this is useless without them.

The first: AI still makes mistakes. It produces wrong answers with the same confidence it produces right ones, which is exactly why the ownership point matters so much. The tool is not the thing you trust. The person standing behind its output is. And it is worth remembering how early this is. Measured against where it will be, today’s models are the iPhone 6, not the iPhone 17 Pro. They are useful now and they will be unrecognisable in a few years. Building on them means building for a moving target, not a finished one.

The second is more fundamental. Left to itself, a model works from the average. Ask it to write an email with no further instruction and you get the common denominator of every email like it on the internet: competent, generic, the mean of what it has seen. That is the default behaviour of the whole system. It reaches for the centre of the distribution. But that is a floor, not a limit. Give it twenty or thirty examples of your own writing, tell it the tone, the language, the things to do and avoid, and it will work from your specifics instead of the average. The capability is there. It just needs to be directed, and the quality of what you get back is mostly a function of how precisely you ask.

This is also the honest line on where AI sits today. It is exceptional at organising, recombining, and applying knowledge that already exists, the recorded output of human thinking in articles, papers, code, and essays. It is far weaker at producing genuinely new knowledge on its own, the kind of insight the world has not arrived at yet. That distinction, between applying what is known and creating what is not, is roughly the line between AI as it exists today and the idea of superintelligence.

So I do not treat AI as magic, and I do not treat it as a threat. I treat it as the most powerful tool I have ever been handed, and like any tool, the whole game is knowing exactly where to point it.