The machine will absorb the mechanics of riding — what it cannot absorb is judgment that has something to lose
Training gives a model breadth, reasoning capability, and distilled pattern across disciplines. It does not give it your situation — your specific context, your tacit knowledge, your actual goal buried beneath what you first typed. This piece examines what the human contribution to an AI session actually consists of, and why it determines the quality of the output far more than most users realise.
Five levers separate a sharp result from a generic one: framing the real problem rather than the surface question; supplying context the model cannot infer; injecting the unconventional angle that breaks the model’s statistical bias toward consensus; iterating with precision rather than accepting the first output; and knowing when to stop. Each is a compression of human judgment into the work.
Looking forward, the mechanical parts of the rider’s role will be absorbed — intent inference, session memory, agentic execution across multi-step workflows — within a three-to-five year horizon. What resists absorption is the judgment embedded in genuine stakes: knowing which question is worth asking, carrying tacit knowledge that was never written down, and backing an unconventional idea against consensus. These are not engineering problems. They may be category problems. Nor does the objection that the rider will soon be another machine dissolve the role: agentic loops rebuild the mechanics of riding, but the objective always enters from outside the loop, sibling models compound the very consensus bias the rider exists to break, and accountability cannot be delegated to something that cannot owe.
The professionals most exposed to AI displacement are not the weakest. They are the competent middle — those whose value is defined by reliable execution of known processes, which is precisely what models are built to replicate. Original thinkers and carriers of hard-won experiential knowledge become more valuable as everything else is absorbed.
There is a widespread and quietly dangerous assumption about AI systems — that they are either oracles to be consulted or tools to be operated, and that the distinction between the two is largely cosmetic. It is not. How you engage with a frontier model shapes the result more profoundly than most people who use it casually ever discover. The model is powerful, genuinely so. But power without direction is just noise. The rider matters.
Training gives a model breadth. An enormous, sometimes astonishing breadth. Ask it to draft a term sheet, reframe a macro argument in gold terms, summarise a prime brokerage agreement, or structure a research interview protocol — it will do all of these things competently, often expertly, drawing on patterns absorbed from trillions of tokens of text, code, mathematics, and professional literature. It does not need to be taught these domains from scratch each time you sit down.
Modern large language models compress vast corpora into billions of parameters. The result is not memory, exactly — it is something closer to distilled pattern: the shape of how ideas connect across disciplines. This is both more and less useful than it sounds.
Training also gives a model reasoning capability — not infallible logic, but a genuine capacity to follow chains of argument, identify inconsistencies, hold multiple frameworks in tension, and produce structured thinking across disciplines. When a problem is well-posed, a well-trained model can be a remarkable thinking partner.
What training cannot give it — and this is the crux — is your situation. Your specific context. Your actual goal buried three layers beneath what you first typed. Your tacit knowledge, built across twenty years of watching how institutional relationships actually work. Training encodes the general. You hold the particular.
The model knows the map. You know where you actually are, and why it matters to get somewhere specific by nightfall.
Watch two people use the same AI model on the same day. One receives generic, hedged, vaguely useful prose. The other receives sharp, actionable, precisely calibrated output that saves hours of downstream work. The model is identical. The training is identical. The divergence lies entirely in the human interaction.
This is not mysticism. It is prompt craft — the same skills that make a good brief, a precise instruction to a junior analyst, or a sharp research question to a subject matter expert. Clear thinking, communicated well.
The five levers of human influence
Consider the difference between a highly trained research analyst and a senior portfolio manager. The analyst can process information, identify patterns, draft arguments, and produce structured work at volume. The PM cannot do all of those things faster. But the PM knows which questions are worth asking, which outputs to trust, which assumptions to interrogate, and — crucially — what to do with the answer once it arrives.
AI is the analyst. A very fast one, with very broad exposure. The user is the PM. The system only functions well when both roles are played.
What is seductive — and dangerous — about AI tools is that they make it easy to feel like the PM role is being filled when it is not. You can generate ten pages of structured content in four minutes. That content can look authoritative, well-formatted, and coherent. It can still be entirely the wrong answer to the wrong question, pursued at impressive speed.
A sharp question is not just a good start. It is, in fact, almost the entire job. Everything downstream is execution.
This question deserves a direct answer rather than a diplomatic hedge. The answer is: partially yes, structurally limited, and the ceiling is not primarily technical. It is categorical.
The rider’s role is not monolithic. It is a bundle of distinct contributions — some already being absorbed by the machine, some resisting absorption on principled grounds that go beyond current engineering constraints.
Inferring intent from poorly-formed questions: models are already measurably better at this than eighteen months ago — the 2026 AI Index records agent success on Terminal-Bench, a benchmark of real-world task completion, rising from 20 to 77 per cent in a single year (Stanford HAI, AI Index Report 2026) — and the trajectory is steep. Maintaining context across a session, remembering prior positions and preferences, anticipating the follow-up before it is asked — these are engineering problems with known solution paths. The friction of re-teaching the model your situation every session is largely solvable, and it will be solved.
Agentic AI — models that do not wait for a prompt but pursue a goal autonomously across many steps — is already functional in narrow domains: code, research synthesis, scheduling, workflow execution. The loop between instruction and output is tightening rapidly. Within a few years, “prompting” may feel as archaic as programming by hand.
Within a three-to-five year horizon, a significant portion of the scaffolding work the human currently provides — framing, contextualising, iterating on mechanical outputs — will be absorbed into the model’s own inference layer. The rider will spend less time steering and more time simply stating the destination.
The harder problem is not technical. The rider’s most valuable contribution is not asking questions — it is knowing which question is worth asking in the first place. That judgment is embedded in stakes: real consequences, real relationships, real skin in the game. A model has none of that. It has no cost function attached to being wrong in your specific situation.
In institutional finance especially, the tacit knowledge that makes a senior professional irreplaceable is not procedural. It is relational and situational. Knowing that a particular LP is sensitive about fee optics this quarter because of a redemption cycle they will not announce publicly. Knowing when not to push. Knowing which room in the building holds the real decision-maker. That knowledge does not sit in training data. It lives in accumulated experience with real stakes attached.
There is also a deeper structural problem. Models optimise for the statistically expected response. Genuinely original thought — the kind that creates asymmetric value — requires breaking from the statistical grain. That impulse comes from caring about the outcome in a way that is hard to replicate without genuine agency.
The machine can learn to ask better questions. It cannot learn to care which answer is right.
The obvious next question — and by 2026 it is no longer hypothetical — is what happens when the rider itself is a machine. The current architecture of agentic systems is exactly that: a planner model frames the task, retrieval supplies the context, a critic model iterates on the output, a budget rule decides when to stop. Each of the five levers, rebuilt in silicon and stacked into a loop. Machine riding machine is not a thought experiment; it is the design pattern of the agentic stack.
What the loop cannot do is source two things from inside itself. The first is the objective. A loop optimises whatever it is handed; add a supervisor model and the supervisor optimises whatever it is handed. The regress never terminates in silicon — the outermost goal always enters from outside, from something that wants an outcome and bears its cost. The rider’s irreducible act was never steering; it was wanting. The second is divergence. Every member of the loop is drawn from the same statistical centre, so a critic model criticises toward consensus and sibling supervision compounds the conventional bias rather than correcting it. The lever that breaks from the statistical grain gets weaker, not stronger, as machine riders multiply — a monoculture checking its own homework.
And even a loop that rides well cannot owe anyone the outcome. Bindingness is a legal property, not a computational one — the argument is made at essay length in The Refusal Layer — which is why institutions will run these loops inside licensed entities rather than instead of them (The Bundle Holds). The human rider does not disappear; the ratio changes. One accountable rider over a fleet of machine riders: from riding the machine to riding the loop. The honest failure mode of that answer is not replacement but atrophy — standing mandates delegated at ever higher abstraction, until the outermost human contribution thins to a signature. The risk of the 2030s is not the rider who is replaced. It is the rider who stops riding.
Projections about AI capabilities have been wrong in both directions — too pessimistic about raw performance, too optimistic about the displacement of human judgment. With that caveat stated, the horizon looks approximately like this.
Three horizons
The riders most at risk are not the sharpest ones. They are the people in the middle — professionals who add value primarily by executing known processes with competence and reliability. That profile is exactly what AI systems are designed to replicate. Not because those people are untalented, but because their contribution is, by definition, learnable from pattern.
The riders most at risk are not the weakest. They are the competent ones in the middle — whose value is, by definition, learnable from pattern.
The profiles that become more valuable as the machine absorbs the rest: the genuinely original thinker who brings an angle nobody else had. The person with hard-won tacit knowledge — accumulated across decades of specific, consequential experience — that was never written down and therefore cannot be trained away. The connector who understands relationship dynamics that exist nowhere in any document. The judge who knows not just what is technically correct but what is actually wise in context.
The machine is becoming extraordinary. The rider’s role is narrowing. But what remains of that role, once the narrowing is complete, will be worth more than ever — precisely because so little of it can be replicated.
The machine still needs a rider. For now. The question worth asking today is what kind of rider you intend to be when it needs one less.
Written in a personal capacity. Analytical views only — not legal or investment advice.
Julian Gretzinger — Investor and writer on monetary history, real wealth mechanics, and financial markets. substack.com/@juliangretzinger