TBM 432: Bundling & Unbundling Capabilities (and AI)

John Cutler · email · 2026-07-26 Reader 🖥️

practical · expository essay building a nine-part taxonomy of capability motions, illustrated with a career anecdote and a headline-driven present-day scan, then applying it to AI — Classic Cutler 'Sunday-morning diagramming and thought-provocation': opens with a personal-career bundling/unbundling story, generalizes it into a taxonomy of nine capability motions plus a 'loops not stages' visual, walks the taxonomy against current AI headlines section-by-section, then closes -- as in TBM 433 -- by handing the practical question back to the reader ('Use them as language... The point is not to classify everything perfectly').

Structure

Key terms

• Role as a temporary bundle
  A job title or role is not a stable unit of capability -- it is a provisional grouping of capabilities that happen to sit together right now; the capabilities themselves live independently across people, teams, routines, tools, platforms, and institutions.
• Capability motions (Specialization, Diffusion, Centralization, Integration, Embedding, Rebundling, Externalization, Elimination, Loss)
  Nine distinct ways a capability can move through a system over time -- Cutler's granular alternative to talking about 'roles changing' or 'jobs disappearing.'
• Loss (vs. Elimination)
  A specific, worse motion: work disappears before the underlying capability -- the judgment, context, relationships it sustained -- has been reproduced elsewhere. Distinguished sharply from Elimination, where work genuinely disappears without that cost.
• Loops, not stages
  Capability motions do not proceed through a single linear sequence; many loops run simultaneously at different levels and speeds, each motion creating pressure and tension for other motions.
• The physics of a pattern (vs. its variables)
  Cutler's self-directed diagnostic question -- 'Does AI change the physics of this pattern?' -- distinguishing whether AI alters the underlying causal structure of an organizational dynamic versus merely changing the magnitude of its variables so dramatically that it looks new.
• 'We know more than we can tell' (Polanyi, via diffusion)
  Applied specifically to whether AI can diffuse expertise: codified knowledge transfers via AI, but tacit, contextual, embodied expertise may have a diffusion ceiling that is invisible from outside the expertise itself.

Propositions

• explicitA role is only a temporary bundle inside a system; the capabilities themselves are distributed across people, teams, routines, tools, platforms, institutions, and external partners.conf: high
• explicitWhen technology changes work, capability moves through one or more of nine distinct motions: specialization, diffusion, centralization, integration, embedding, rebundling, externalization, elimination, or loss -- and these are analytically distinct from one another.conf: high
• explicitLoss is worse than Elimination: loss means work disappears before the judgment, context, relationships, and learning pathways it sustained have been reproduced elsewhere, whereas elimination means the work simply, genuinely disappears.conf: high
• explicitContradictory-seeming AI-era headlines (PMs becoming 'AI builders,' designers becoming 'designer engineers,' titles collapsing into 'builder') are not evidence of incoherent change -- they are separate capability motions occurring simultaneously, compressed into single catchy narratives by the press.conf: high
• explicitCapability motions do not occur in a fixed linear sequence; multiple loops operate simultaneously, at different organizational levels and over different timeframes, and each motion creates pressure or tension for other motions.conf: high
• explicitdevelopedAI changes the magnitude, cost, and feasibility of each capability motion, but does not change the underlying physics -- the causal structure -- of the pattern itself; the appropriate question is motion-by-motion ('to what extent, under what conditions, with what second- and third-order effects'), not a single yes/no verdict.conf: high
  supportsTBM 430: Incubate, Compound, Refinance, Liquidate · p7 — Both propositions make the identical 'physics vs. variables' move about AI: TBM 430#p7 says AI does not eliminate complexity/coupling/coordination/decay in a software portfolio, only changes the cost/speed/feasibility of responding to them; TBM 432#p6 makes the same claim one level up, about capability motions generally. TBM 432 generalizes TBM 430's portfolio-specific finding into Cutler's cross-essay methodological stance.
  extends ←TBM 431: The Denominator That Matters · p1 — TBM 432 states AI changes the magnitude, cost, and feasibility of capability motions but not their underlying causal 'physics'; TBM 431 operationalizes that same physics-vs-accelerant thesis into a literal ratio with named variables, applied to organizational AI success generally rather than to capability motions specifically.
• explicitcorroboratedAI can diffuse codified expertise widely, but it is unknown whether it can diffuse tacit, contextual, or embodied expertise -- and if Polanyi's 'we know more than we can tell' remains true, diffusion through AI has a ceiling invisible from outside the expertise.conf: medium
  supports ←Design AI Systems That Actually Strengthen Human Reasoning · p5 — Both articulate a ceiling on how much of contextual/tacit expertise AI can diffuse or substitute for, framed via Polanyi-style tacit knowledge in tbm-432 and via lived/embodied experience here.
• explicitcorroboratedAI can reduce the translation, handoff, and queueing costs (friction) created by specialization, but it has not eliminated the social work of coordination, negotiation, and shared judgment between specialists and generalists.conf: high
  supports ←Netflix CPTO on AI and the future of product and tech roles | Elizabeth Stone · p3 — Stone's insistence that fluid prototyping still requires engineering partnership matches Cutler's claim that AI reduces specialization friction without eliminating the social work of coordination and shared judgment.
• explicitAI enables people to produce artifacts across more role boundaries (a 'higher-level generalist'), but producing more kinds of artifacts is not the same as possessing the underlying expertise those artifacts used to signal.conf: high
  supportsFragments: February 18 · p8 — TBM 432's Rebundling claim (AI enables a 'higher-level generalist' who produces artifacts across role boundaries, but producing artifacts is not the same as possessing the underlying expertise) reinforces the corpus's existing Expert Generalist thread: Fowler's Thoughtworks-retreat proposition that LLMs erode platform-specialist value because LLM-driving skills matter more than platform details is the same rebundling dynamic seen from the individual-skill side rather than the role-taxonomy side.
• explicitAI simultaneously broadens local access to a capability while concentrating control over the models, infrastructure, standards, and defaults underlying that capability in a small number of firms -- both diffusion and centralization are true at once, at different layers of the same capability.conf: high
• explicitcorroboratedAs expertise is embedded directly into tools and workflows, the human role that remains shifts toward direction, assessment, and oversight -- which are genuinely different skills from the skills the embedded expertise replaced.conf: high
  supports ←Design AI Systems That Actually Strengthen Human Reasoning · p16 — Both describe the human role shifting toward direction/assessment/judgment as a genuinely different (and valuable) skill from the skill AI substitutes for — Anthony's senior bankers who engage with the technology vs. tbm-432's 'direction, assessment, and oversight' framing.
• explicitWidespread AI adoption is not clean evidence of AI's value, because visible action and spending have themselves become competitive signals independent of value produced -- one of the defining dynamics of a bubble.conf: high
• explicitBecause AI's organizational effects touch individuals, teams, incentives, workflows, and markets simultaneously, task-level research findings (e.g., a study showing AI improves performance on a bounded task) cannot be reliably generalized into claims about how roles, teams, or organizations will change.conf: high
• explicitTreating a codebase as a pile of code rather than a cumulative record of decisions, constraints, failures, and hard-won knowledge leads people to falsely assume that generating replacement code is equivalent to replacing the system.conf: high
• explicitBecause familiar-looking artifacts (a chart, a roadmap, a status report) make specialist work look understandable to non-specialists, and AI can reproduce those artifacts long before it can reproduce the underlying judgment and coordination, AI systematically worsens the existing confusion between 'looks like something I could do' and 'is actually qualified to do it.'conf: high
• explicitAI is not one adoption curve but a stack of technical (model architecture), economic (adoption), and social (organizational adaptation) curves unfolding at different speeds, so asking whether AI is 'early' or 'late' is too simplistic a question.conf: high
• explicitThe taxonomy of capability motions is intended as shared descriptive language for what people are already observing, not a framework to be imposed or perfectly classified.conf: high
• implicitBecause Cutler frames every motion as producing genuine adaptation, harmful instability, or both without specifying the conditions that tip a given motion toward one outcome or the other, the taxonomy is currently descriptive rather than predictive -- it names what can happen, not when it will.conf: medium
• implicitThe essay implicitly treats organizational legibility (making tacit capability visible and describable) as a precondition for using the taxonomy at all -- you cannot ask 'where is expertise diffusing/centralizing/embedding' inside an organization whose actual capability distribution is invisible to its own leaders.conf: medium

Arguments

Narratives

• Cutler's own early product-management career, in which he personally performed research, data analysis, wireframing, requirements-writing, release coordination, support, and even some coding -- all loosely bundled under 'product management' -- before those capabilities separated into distinct disciplines (design, analytics, product marketing, program management, product operations) over time.
  → grounds the abstract 'role as temporary bundle' thesis in a single lived, concrete instance before generalizing it into the nine-motion taxonomy -- the essay's entire argument is an induction from this one biography.
• The MIT Work of the Future task force (Autor, Mindell, and Reynolds, 2020) warning that cutting out middle tasks can destroy workers' skills, networks, and career pathways even when immediate output looks intact.
  → supplies external, citable authority for the Loss motion's central worry, preventing it from reading as pure speculation.

Questions addressed

• ✅ Why do AI-era headlines about roles seem to contradict each other (broadening, disappearing, specializing, and merging all at once)?
  Because multiple distinct capability motions are occurring simultaneously and headlines collapse them into single catchy narratives.
• ⭕ Does AI change the fundamental patterns of organizational capability, or just the scale/speed of existing patterns?
  Cutler poses this as his own open diagnostic question and answers it only provisionally, motion by motion, without a final unified verdict.
• ⭕ Can AI actually diffuse deep expertise to more people, or only its codified surface?
  Left open, gated on whether Polanyi's tacit-knowledge ceiling holds under AI -- explicitly stated as currently unknown.
• ✅ What happens to the human role once its expertise is embedded into AI tools?
  It shifts toward direction, assessment, and oversight -- genuinely different skills from the embedded ones.
• ⭕ What conditions determine whether a given capability motion produces useful adaptation versus harmful instability?
  Not addressed; each motion's tension is described as capable of producing either outcome or both, with no criteria given for which prevails when.

Gaps

• incompleteCutler names that each capability motion 'may produce useful adaptation, harmful instability, or both' but never supplies criteria, examples, or even a hypothesis for what determines which outcome dominates in a given case -- team maturity? psychological safety? existing coordination debt? Without this, the taxonomy is a vocabulary for description but offers no diagnostic or prescriptive power: two organizations undergoing the identical motion (say, Embedding) get no guidance on why one thrives and one degrades.
  Read next: Contingency-theory organizational design literature (e.g., Galbraith on organization design fit) for criteria that predict when a given capability reallocation helps versus harms; Donella Meadows on leverage points for a framework distinguishing which interventions matter more.
  closed byTBM 431: The Denominator That Matters — TBM 432's open gap asks what determines whether a capability motion tips toward useful adaptation or harmful instability, listing 'psychological safety' among unexamined candidate factors. TBM 431 supplies Cutler's own explicit candidate answer at the level of AI adoption generally: the Social Contract (whether gains will be captured against the people who generated them) -- though this is Cutler's asserted hypothesis, not empirically adjudicated, so the gap is narrowed rather than fully closed.
• incompleteThe essay's Loss motion and its MIT-task-force citation name the risk that judgment, context, and career pathways disappear before being reproduced elsewhere -- but Cutler offers no mechanism, checkpoint, or organizational design that would let a leader detect Loss is happening before it has already happened (unlike, say, a lagging financial metric). The essay's own Conclusion asks 'which routine activities are helping people build judgment' as a discussion-starter question, implicitly admitting no answer exists yet.
  Read next: This corpus's own junior-pipeline/apprenticeship-collapse gaps (e.g. extract-research-why-some-junior-employees-work-well-with-ai-and-others-dont, the Hanselman and Reed extracts) for parallel unresolved attempts at the same mechanism question; Cappelli on internal talent architecture for a stock-and-flow framing of capability loss.
• uninformedCutler asserts that AI 'hasn't eliminated the social work of coordination, negotiation, and shared judgment' between specialists and generalists (Integration subsection), but doesn't engage the possibility -- raised elsewhere in this reader's corpus about the same author's fiefdom essay -- that AI's friction-busting on translation and handoff costs could, over a longer horizon, shift the political economy of coordination itself (who needs whom, who can bypass whom), not just its transactional cost. He treats the political layer as static while granting the cost layer is dynamic.
  Read next: extract-cutler-ai-wont-save-kingdoms#g3 poses exactly this dynamic-vs-static coordination-cost question from the fiefdom side; reading the two together tests whether Cutler's own corpus is internally consistent on this point.
• incompleteThe essay treats 'capability' as if it moves largely independent of formal organizational structure (reporting lines, budget ownership, team boundaries) -- but Cutler's own other writing (on defining teams, on the reverse Conway maneuver) argues architecture and org structure are hard constraints on how fast capability can actually be reallocated. This piece never reconciles the two: can a capability really 'rebundle' into a role if the team/reporting/architecture boundaries around it haven't moved, or does rebundling silently stall at the structural layer?
  Read next: extract-tbm-423-defining-teams (same author) directly argues architecture is a hard constraint underestimated by reorg efforts -- read the two Cutler pieces together for the missing synthesis.

Closes gaps elsewhere

• AI Won't Save the Kingdoms We Built · g3 — That gap asks whether AI changes the unit economics of coordination (cheaper synthesis/proposals) enough to shift the political calculus around fiefdoms over time, since Cutler's kingdoms essay treats AI's inability to fix coordination as static. TBM 432's Integration proposition (p8) gives a partial, motion-specific answer: AI can friction-bust the translation, handoff, and queueing costs specialization creates, but has not eliminated the social work of coordination, negotiation, and shared judgment -- confirming the gap's suspicion that the cost layer is dynamic while showing the political/social layer remains largely untouched, at least on the time horizon Cutler considers here.

Related documents

• TBM 430: Incubate, Compound, Refinance, Liquidate
• Experience Mapping Matters More the Faster You Move
  AI as Amplifier Coordination Cost Organizational Legibility
• Managers Are Not Overhead: They Are Infrastructure
  AI as Amplifier Coordination Cost Organizational Legibility
• TBM 434: How Maps Can Hide Problems
  AI as Amplifier Organizational Legibility Sociotechnical Architecture
• 🧠 Community Wisdom: How AI is changing product operating models, tracking work stress with Whoop, whether you nee…
  AI as Amplifier Coordination Cost
• AI Won't Save the Kingdoms We Built
  AI as Amplifier Coordination Cost
• John Cutler's Post
  AI as Amplifier Coordination Cost
• Software Development As Collective Learning With Hazel Weakly
  AI as Amplifier Sociotechnical Architecture
• AI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action
  AI as Amplifier Sociotechnical Architecture
• The engineering manager role is splitting in two
  Coordination Cost Organizational Legibility