Four Categories Every Piece of Cognitive Labour Falls Into, and the Sorting the Architect Must Perform Before Someone Else Performs It on the Architect’s Behalf.
Artificial intelligence is absorbing entire categories of cognitive labour, and the sorting of which categories those are has already begun, whether the architect participates in it or not.
This is the research aligned framework for Volume IX, and it deserves a direct opening because the substance is time sensitive in a way most of the Circle’s frameworks are not. Through 2025 and into 2026, the capability of large language models and adjacent systems has moved from producing plausible drafts to producing work that, across a widening set of categories, a trained reader cannot reliably distinguish from work produced by a skilled human operating carefully. The pace of that widening is itself contested. What is not contested, across every serious account of the technology’s trajectory, is the direction. The set of tasks a machine can perform at an acceptable standard is larger this year than last year, and there is no credible account in circulation that expects the set to shrink.
The operator economy’s dominant response to this shift has been either denial, the position that the architect’s own work is safely exempt, or panic, the position that everything is about to be automated and the only rational response is anxiety. Both responses share the same underlying failure. Neither performs the actual sorting the moment requires.
This essay introduces the framework. The Reallocation Doctrine. Four categories. Irreducible Judgment. Augmentable Work. Automatable Work. Obsolete Work. Every piece of cognitive labour the architect performs, and every piece of cognitive labour the architect’s team performs, falls into one of these four categories. The architect who has not performed this sorting deliberately will discover it has been performed anyway, on someone else’s timetable, without the architect’s judgment involved in where any given piece of work landed.
Why the exemption framing and the panic framing both fail
Three reasons neither the exemption framing nor the panic framing serves the architect.
First reason. The exemption framing assumes the architect’s own work is categorically different from the work being absorbed, usually on the grounds that the architect’s work involves judgment, relationships, or taste. This is true of some of the architect’s work and false of a meaningful share of it. The architect who has not separated the judgment component from the executional component of their own labour cannot know, honestly, which share is actually exempt.
Second reason. The panic framing treats the absorption as total and immediate, which produces paralysis rather than useful action. The absorption is real, but it is categorical, not uniform. Some work is genuinely irreducible. Some work is genuinely already gone. Treating the whole of cognitive labour as a single undifferentiated category under threat prevents the architect from seeing where the actual, specific, addressable risk sits.
Third reason. Both framings are passive. Exemption waits to be proven wrong. Panic waits for the wave to arrive and hopes to survive it. Neither performs the active work of sorting, which is the only response that produces a structural advantage rather than a structural exposure.
The structural framing corrects all three failures. Reallocation, engineered deliberately, replaces both exemption and panic with an active discipline the architect controls.
Category One. Irreducible Judgment
Irreducible Judgment is the work that depends on context, accountability, and relationship in a way no current or foreseeable system can substitute for, because the substitution would require the system to bear consequence, which it structurally cannot.
This category is smaller than most architects instinctively assume, and larger than the panic framing allows. It includes the decisions where the cost of being wrong is borne personally by the architect, where the relationship carrying the decision depends on the counterparty trusting a specific person’s judgment rather than an output, and where the context required to make the call correctly is not fully documentable, existing instead in years of accumulated pattern recognition the architect has built and cannot fully externalise.
Four properties to audit when placing work in this category.
First. Consequence bearing. The work involves a decision whose failure the architect, specifically, is accountable for, in a way that cannot be transferred to a tool without also transferring the accountability, which the tool cannot hold.
Second. Relational dependency. The counterparty’s trust attaches to the specific person making the call, not merely to the quality of the output, meaning the same output produced by a system would not carry the same standing even if the content were identical.
Third. Undocumented context. The judgment draws on accumulated pattern recognition that exists in the architect’s own experience and has not been, and likely cannot be, fully captured in any document or dataset a system could train against.
Fourth. Genuine novelty. The decision addresses a situation sufficiently unlike prior situations that pattern matching against historical data, however extensive, would not reliably produce the correct call.
The dominant failure mode at this category is overclaiming it. Architects place work here that does not actually meet these four properties, usually because the work feels important or because admitting otherwise feels like a diminishment. The honest audit is uncomfortable precisely because it requires the architect to separate what genuinely requires their judgment from what merely always has, until now, required a human, for reasons that were about capability rather than about anything irreducible.
Category Two. Augmentable Work
Augmentable Work is the category where a system meaningfully improves the architect’s or the team’s output, speed, or quality, without replacing the human’s role in the work, because the human retains the framing, the judgment about which output to accept, and the accountability for the result.
This category is currently the largest and the fastest growing of the four, and it is also the category where the architect’s discipline matters most immediately, because augmentation performed carelessly degrades quality even as it increases speed, while augmentation performed deliberately compounds both.
Four properties to audit when placing work here.
First. Retained framing. The human defines the problem, the constraints, and the standard the output must meet, and the system operates inside that framing rather than substituting its own.
Second. Retained review. A qualified human reviews the output against the standard before it is used, rather than allowing the system’s output to pass through unexamined on the assumption that speed matters more than verification.
Third. Genuine time or quality gain. The augmentation produces a measurable improvement, not merely a feeling of productivity. The architect who has not measured this honestly risks mistaking activity for output.
Fourth. Skill preservation. The human operator retains, and continues to develop, the underlying skill the system is augmenting, rather than allowing the skill itself to atrophy behind the convenience of the tool, which would convert augmentation quietly into dependency.
The dominant failure mode at this category is unexamined dependency, the fourth property inverted. The architect or the team becomes so reliant on the augmentation that the underlying judgment required to evaluate the system’s output correctly begins to erode, at which point the augmentation is no longer improving human judgment. It has begun to replace the capacity for judgment that would be needed to catch the augmentation’s own errors.
Category Three. Automatable Work
Automatable Work is the category where the task can be performed by a system end to end, at an acceptable standard, without meaningful human judgment in the loop, and where the accountability for the outcome does not require a specific human’s standing.
This category has expanded substantially through 2025 and 2026 across drafting, summarisation, first pass analysis, routine correspondence, and structured research assembly, among other domains. The precise boundary of the category continues to move, and the architect’s discipline is not to fix the boundary permanently but to review it on a fixed cadence, because work that sat in Augmentable last year may have moved fully into Automatable this year, and the architect who has not rechecked the boundary is paying human attention against work that no longer requires it.
Four properties to audit when placing work here.
First. Standard achievability. The system’s unassisted output meets the required standard reliably, not occasionally, verified against actual review rather than assumed from the system’s general reputation.
Second. Consequence tolerance. The cost of an occasional error is genuinely acceptable, either because the error is easily caught downstream or because the stakes of any individual instance are low.
Third. Absence of relational dependency. The counterparty receiving the output does not require it to carry a specific human’s personal standing, meaning automation does not damage a relationship that depended on the human element being present.
Fourth. Monitoring in place. Even fully automated work carries periodic human review, not to perform the task but to confirm the standard is holding as the system, the data, or the surrounding conditions evolve.
The dominant failure mode at this category is the opposite of the one at Irreducible Judgment. Architects underclaim it, continuing to route work through expensive, slow human attention out of habit or discomfort, long after the work has genuinely moved into a category a system can carry at an acceptable standard. This failure is costly in a specific way. It is not merely inefficient. It is a misallocation of the architect’s own scarcest resource, personal attention, against work that no longer requires it, at the direct expense of the Irreducible Judgment category, which does.
Category Four. Obsolete Work
Obsolete Work is the category where the task itself no longer needs to be performed at all, by a human or a system, because the reason it existed has disappeared, and the honest audit reveals it was being preserved out of habit rather than genuine necessity.
This category is easy to overlook because it is not about who performs the work. It is about whether the work should exist. Architects sorting labour into the first three categories can miss this fourth category entirely, dutifully deciding whether a task belongs to human judgment, augmentation, or automation, without ever asking whether the task has quietly stopped mattering.
Four properties to audit when placing work here.
First. Purpose decay. The original reason the task existed, a specific report a specific stakeholder once needed, a specific check a specific risk once required, has genuinely lapsed, and the task has continued only through institutional inertia.
Second. Absence of a genuine consumer. No one actually uses the output of the task in a way that changes any subsequent decision, meaning the task’s continued existence produces no consequence either way.
Third. Substitution by upstream change. A change elsewhere in the architecture has made the task redundant, a new system, a new process, a restructured decision right, and the original task has simply not been retired to reflect it.
Fourth. Honest willingness to retire it. The organisation is prepared to actually stop performing the task once it is identified as obsolete, rather than continuing it because retiring it feels like an admission that the work was unnecessary all along.
The dominant failure mode at this category is the one hardest to correct, because it is emotional rather than structural. Obsolete work is frequently preserved because someone built genuine identity around performing it, and retiring the task feels, to that person, like a judgment on their contribution rather than an honest structural finding. The architect’s discipline is to separate the two explicitly, retiring the task while explicitly affirming the person, a distinction that requires deliberate communication rather than assumption that the distinction will be understood on its own.
The integration of the four categories
The four categories form one Reallocation Doctrine, and the integration is the actual discipline, more than any single category examined alone.
The categories interact through a single scarce resource, the architect’s own attention and the team’s collective attention, which the four category sorting exists to allocate correctly. Every hour spent on work that has genuinely moved into Automatable is an hour unavailable for Irreducible Judgment, the category that actually requires it. Every hour spent preserving Obsolete Work is an hour unavailable for either. The sorting is not an academic exercise. It is the mechanism by which the architecture’s scarcest resource is protected from being spent on work that no longer requires it.
This is the direct continuation of the Talent Architecture (Volume III), which named the layers of capability an architecture develops in its people. The Reallocation Doctrine asks a related but distinct question, not what capability a person has, but what category the specific work they are currently doing actually belongs to, regardless of the capability of the person currently doing it. A highly capable person can be spending years on work that has quietly moved into Automatable, and the Talent Architecture alone would not reveal this. Only the Reallocation Doctrine, applied to the actual tasks rather than to the person, surfaces it.
It also connects directly to this Volume’s Monday framework, the Delegation Architecture. Delegation and reallocation are related but distinct movements. Delegation moves work from the architect to another person, through the four gates already named. Reallocation moves work between the four categories this essay describes, a decision that must be made before delegation is even considered, because the correct destination for a piece of work, another person, a system, or retirement entirely, depends on which category the work actually belongs to.
The audit cadence
The Reallocation Doctrine is audited quarterly, more frequently than most frameworks in the catalogue, because the boundary between the categories, particularly between Augmentable and Automatable, is moving faster than most structural boundaries the Circle audits.
At each quarterly review, the architect and the team walk the current task list and ask, honestly, for each recurring category of work, which of the four categories it now belongs to, explicitly checking whether anything has moved since the prior quarter. The most valuable finding, most quarters, is not a dramatic reclassification but a small one, a task that has quietly slid from Augmentable into Automatable, freeing attention that had not been noticed as available.
A closing note on the architect’s own position
The architect performing this sorting on their team’s work must perform it, with equal honesty, on their own. The temptation to sort the team’s labour rigorously while exempting the architect’s own work from the same scrutiny is understandable and indefensible. The architect who has genuinely audited their own calendar against these four categories, honestly, usually discovers that a meaningful share of what currently occupies their own attention belongs in Augmentable or Automatable, not in Irreducible Judgment, and that the architect’s own greatest structural risk is not that a system will replace their judgment. It is that the architect has been spending judgment on work that never required it, at the direct cost of the work that does.
Doctrine close
Artificial intelligence is absorbing categories of cognitive labour, and the sorting of the architect’s own work and the team’s work into what remains irreducible and what does not is either performed deliberately or performed by default, on someone else’s timetable.
The Reallocation Doctrine names four categories. Irreducible Judgment. Augmentable Work. Automatable Work. Obsolete Work. Sorted honestly and reviewed quarterly, the four categories protect the architecture’s scarcest resource, the architect’s own attention, from being spent where it is no longer required.
The Circle is a room of architects who perform this sorting deliberately, rather than discovering their own obsolescence in a category they never audited.
Structure. Access. Positioning.
The Architect Circle is application only. Two hundred and fifty seats globally. createyourownmarket.com.au.