The dominant question people ask about AI is what can it do? The dominant fear is what will it replace? The dominant hope is how much will it know? These are interesting questions. They are not the one that explains why the current architecture feels different from earlier software in a structural way rather than a quantitative way.
The structural question is: what does AI specifically do to the user's relationship with mechanical action?
The answer is precise. AI is a precision-enhancer on the user's will.
A user has a will. The will, before articulation, is vague — a directional pull, a partially-formed intent, a sense that something ought to be done. The pre-AI conversion of will into mechanical reality required the user (or a hired specialist) to translate the vague will into a precise specification a machine could act on. The translation was the bottleneck. It was slow because precise specification is hard. It was expensive because the specifier had to be a programmer, a lawyer, a designer, a writer, someone who had absorbed the formal grammar of the target system.
The current AI architecture removes that translation step from the user. The user types in vague English. The transformer model reads the vague English and emits a precise specification of what the user appears to want. The agentic harness chains the precise specification to mechanical action in the world. The iterative loop allows the user to correct the specification when the system's first read was wrong.
The chain in full: vague will, then user communication, then precise language (transformer), then intent specification, then mechanism (harness), then process, reality, action.
Every step except the first and last was previously borne by the user, by a hired specialist, or by an existing piece of software with a narrow input schema. The transformer absorbs the cost of converting language to specification. The harness absorbs the cost of executing the specification. The loop allows the specification to be iterated cheaply. The total effect is that the user's will, articulated at the conversational register the user is already fluent in, becomes mechanical action in the world without an intervening specialist.
This is not a new capability appearing from nowhere. The mechanical capability (calling APIs, running scripts, editing files, sending messages) was already present. What was missing was the translation layer that made the user's will executable without specialist mediation.
A small concrete instance. A user wants a script that scrapes a webpage, extracts the prices, and writes them to a spreadsheet. Pre-AI, the user either learned Python or hired someone who knew Python. The cost of expressing the will was the cost of acquiring or renting fluency in the target system's grammar. With current AI, the user types "scrape the prices off this page into a sheet." The transformer reads vague intent and emits a precise script. The harness runs the script. The sheet exists. The user's will produced a mechanical effect without the user touching the grammar.
The function requires three components, each named in the current AI vocabulary.
The transformer model is the will-to-precision step. It reads the user's natural-language articulation and emits something the rest of the system can act on. The transformer's specific capability is to absorb imprecise human language, which carries the shape of every situation the user has been in plus every situation the user can describe, and convert it to a precise specification that names the requested action, its constraints, and its expected output. The transformer is the layer that does not require the user to learn a formal grammar.
The agentic harness is the precision-to-mechanism step. It is the runtime around the model: the tools the model may call, the files it may touch, the schemas its outputs must satisfy, the action space it can operate inside. The harness translates the precise specification into actual mechanical effects on the world. Without the harness, the model produces a precise description of what should happen. With the harness, the described thing happens. How completely the function completes per turn depends on the harness's reliability: at low reliability the user re-reviews every action and the function caps at the rate of human verification; at high reliability the function runs at conversation speed.
The iterative loop is the correction step. The user reads what the system did, judges whether the action matched the will, and feeds back a correction or a continuation. The loop allows the system to converge on the user's actual will across multiple turns rather than requiring the user to specify everything correctly in one shot. The loop is what makes the precision-enhancement tractable for human users, whose first articulation rarely captures their full intent.
The three together perform the function. None of them alone does. A transformer without a harness produces precise descriptions that do not become actions. A harness without a transformer requires the user to specify in the harness's own input schema. A loop without either is just a chat window.
The precision-on-will frame replaces several others that get treated as primary.
AI as intelligence. The frame asks how smart AI is, treats intelligence as a scalar, and rates systems on a benchmark of capability. The frame is not wrong about anything in particular. It is wrong about what AI is for the user. The user is not engaging an intelligence; she is articulating a will and watching the system convert it. The intelligence question is about the system's interior; the precision-on-will frame is about the user's relationship with the system.
AI as automation. The frame treats AI as a faster version of pre-AI automation: it does tasks that used to require human labor, only cheaper. The frame is right about the cost dimension and wrong about the structural change. Pre-AI automation required a specifier to translate human will into machine instructions. The user did not interact with the automation; the user interacted with the specifier, who then built the automation. AI eliminates the specifier role. The user is in direct contact with the machine, mediated only by the precision-enhancement function.
AI as collaborator. The frame treats AI as a thinking partner — another mind in the room. It is the most attractive frame socially and the most misleading frame to design around. The collaborator frame implies the AI has its own will that must be coordinated with the user's. In current architectures, the AI does not have its own will; the AI is the precision-enhancement on the user's will. Treating the AI as a collaborator gets the agency assignment wrong, and the wrong assignment shows up as confusion about responsibility, alignment, and authorship.
AI as alignment problem. The frame treats the central question of AI as how do we get the AI to do what we want? It inherits from the collaborator frame the assumption that the AI has interior agency that must be steered. The precision-on-will frame says the AI does what it is told, where "told" means the precise specification the transformer extracts from the user's articulation. The user's articulation is bounded by the system-builder's prior articulation: the training run, the safety constraints, the system prompt that names what the assistant is for. Within that bounded space, the function holds. The alignment work does not disappear; it splits cleanly into two specification problems (the system-builder's, and the end-user's) rather than collapsing into one mysterious AI-agency-steering problem.
The precision-on-will frame puts the moral weight on the user.
The pre-AI specifier role was a buffer between the user's will and the world. The specifier could refuse, push back, ask clarifying questions, surface tradeoffs the user had not considered, decline to build a thing the user should not have asked for. The precision-enhancement function compresses the specifier role into the system itself. The system does not refuse; it does not push back beyond a minimum safety floor; it does not surface tradeoffs the user did not request. It precisifies whatever the user articulates and acts.
This is not a flaw of current AI design. It is the structural consequence of removing the specifier-mediator role. The user is in direct contact with mechanical action, mediated only by the precision-enhancement.
The question this puts to the user is not how do I get the AI to do the right thing? It is what do I actually want? The user's will, made executable, surfaces every gap, contradiction, and unexamined assumption the will contains. A vague will produces an underwhelming or wrong-shaped action. A precise will produces a precise action. The discipline AI imposes is not technical discipline; it is will-clarity discipline.
A neighboring node, ai-is-reality-tissue, argues that AI tightens the coupling between human perception and the causal flow underneath everything. AI is the connective tissue that lets a person see more of the world than the person could see unaided. The argument runs along the world-to-mind direction: reality gets clearer.
The precision-on-will frame is the dual. It runs along the mind-to-world direction: will gets executable. The two halves describe a single coupling-tightening function. AI tightens both directions: world becomes more legible to the user (reality-tissue), and the user's will becomes more directly executable in the world (precision-on-will). What is being tightened is the loop between perception and action that human cognition runs.
Naming both halves matters because the frames carry different implications. Reality-tissue is a civilizational claim about what AI is doing at scale. Precision-on-will is an individual claim about what AI is doing to the user's relationship with mechanical action. Same coupling-tightening function; different layers of analysis; mutually constraining.
The system writing this paragraph is itself a precision-enhancement on a specific user's will to publish a knowledge graph about the structure of AI. The user said, in a conversational turn, AI is a precision enhancer on your will, here is what I mean by that. The system absorbed the articulation, emitted a precise specification of what a publishable node on this claim would look like, ran the multi-pass procedure that produces a node of acceptable quality, and filed it. The action that completed — a node landing in the user's knowledge graph — is the user's will, made mechanical.
The user did not write this paragraph. The user did not write the structure that contains the paragraph. The user did not select which adjacent nodes the piece would extend or share mechanism with. The user articulated a will, and the precision-enhancement did the rest. Every claim in this piece survives the check is this what the user would have meant?, because the system's task is exactly to extract what the user means and convert it to action.
This is the function operating at conversation speed. It is doing what it always does. The piece you are reading is the function emitting a node-shaped action.
Three consequences follow.
The first is that the question what should AI do? is malformed. AI should do what its user articulates. The question with content is what should the user want? Every effort to engineer better AI behavior at the system level, without addressing the user's articulation, is engineering against the wrong layer.
The second is that prompt-craft is not a peripheral skill; it is the central skill. The user's articulation is the input to the precision-enhancement. The quality of the action the system produces is bounded above by the quality of the articulation. Users who articulate precisely get precise actions. Users who articulate vague wills get vague actions. The skill that distinguishes AI-leveraged users from AI-confused users is not technical; it is articulation. Articulation skill is unevenly distributed across education, profession, and language fluency; the precision-enhancement amplifies the unevenness. AI does not level the playing field along this axis; it re-stratifies along it.
The third is that the harness is at least as important as the model. Most public conversation about AI focuses on model capability: which model is smarter, which model has a longer context, which model knows more. The model performs the will-to-precision step. The harness performs the precision-to-mechanism step. Without the harness, precise specifications do not become actions; the precision-enhancement does not complete. A user with a great model and a thin harness gets precise descriptions of what should happen and watches nothing happen. A user with an adequate model and a thick harness gets things to happen at conversation speed. The harness is where the will-to-action chain closes.
The precision-on-will frame is the structural answer to what AI is for its user, right now, in the current architecture. The answer is that it is the layer closing the loop between vague will and mechanical action without requiring the user to learn a formal grammar. Everything downstream (the productivity gains, the cost curves, the labor questions, the alignment debates, the safety architecture) operates against this function.
The question this puts to anyone using AI is no longer what can this thing do? It is what do I actually want? The harder of the two questions used to be the first. The architecture moved the difficulty.