Inheritance · delegation · invention
The Executable
Institution Model
A theory of how organizations preserve operating capability, delegate work to agents, and turn learned behavior into durable digital capability under a new cost structure.
- Inheritance
- What can be preserved?
- Delegation
- What can be entrusted?
- Invention
- What becomes possible now?
00 · Problem statement
“How much organizational work can AI replace?” collapses several different problems into one slogan.
It mixes the limits of organizational design, knowledge materialization, representation, model capability, verification, human attention, and accountability. Those boundaries move for different reasons and belong to different owners. A useful theory must separate them before it can guide product or organizational decisions.
The inheritance view
How much of an existing organization can be preserved as an institutional asset, and how much of that asset can agents reliably carry? This is the migration path.
The invention view
Which operating models never existed because they were uneconomical under human costs, but become viable under agent economics? This is the source of advantage.
Inheritance asks what survives translation. Invention asks what changed economics makes newly viable. Calling the second “more automation” would preserve the very assumptions—hierarchy, sampled review, rare change, person-bound capability—that now deserve redesign.
01 · The core model
Expression is a chain. Execution is a function.
The expression chain asks whether operating capability can be separated from people and represented. Rₜ is then the principal institution-specific input to execution, alongside model priors, verification, and human response. Eₜ is their combined runtime result—not a fifth layer and not a subset of Rₜ.
Expression chain
Rₜ ⊆ O* ⊊ O ⊊ CₜThe outer two gaps are definitional. The inner gap is an implementation gap that can narrow, widen, and change form.
Delegation function
Eₜ = f(Rₜ, Pₜ, Vₜ, Hₜ)02 · Inheritance
Preserve the purpose, not the human rule form.
O* names the purposes that can be detached from particular people and inherited by an institution. Rₜ is a new implementation for an agent executor. It should not copy every reminder, approval layer, or redundancy created to compensate for human limitations.
Complete Value-Creation Capability
People, relationships, capital, physical capabilities, market position, and digital assets including business data, traces, software systems, and business specifications.
Scope boundary · Cₜ → O
People and assets do not enter the Operating Model. How they are recognized, called, configured, coordinated, and governed does.
Operating Model
The structure that turns capability into coordinated, repeatable outcomes: goals, roles, authority, process, standards, incentives, information flows, precedents, escalation, and feedback.
Materialization boundary · O → O*
Some operating capability remains attached to people: private trust, physical skill, tacit intuition, taste, and situated perception.
Materializable Operating Model
The purposes within O that can be separated from particular individuals and preserved, inherited, and reused without losing what they are meant to accomplish.
Representation boundary · O* → Rₜ
O* specifies stable organizational purposes. It does not prescribe a document form, a rule shape, or whether the executor is human or agent.
Represented Operating Model
The part Airic currently expresses through roles, tasks, processes, standards, precedents, authority boundaries, decision requests, sessions, and links to systems.
Implementation layer
Representation should preserve purpose, not copy every human-facing rule. Rₜ is allowed to grow, change form, and shrink.
Two causes of a smaller Rₜ
Representation loss
Airic intends to represent a purpose but cannot carry it accurately through current documents, metadata, links, dynamic state, or conflict resolution. This is a real loss.
Deliberate non-representation
A human-facing rule is omitted because every function it served is covered by a model prior, another rule, or a stronger mechanism. This can be lossless.
The invariant is purpose coverage—not rule preservation.
The hidden-function test
Before retiring a rule, list every function it performs and name the new home for each one. Sampling, for example, controls quality but also absorbs legitimate exceptions. Replacing it with full automated checks while forgetting the exception channel preserves the declared function and loses the hidden one.
03 · Company assets
The Operating Model is not the company’s digital capability.
The expression chain follows one class of organizational asset: the structure that coordinates work. Business data, traces, software systems, and business specifications also belong to Cₜ. They are not inside O. The Operating Model calls, configures, and governs them.
The relation also runs outward. When an agent-carried responsibility becomes sufficiently deterministic, its purpose should move into software and its rule should leave Rₜ. The Operating Model is therefore an intermediate state in capability maturation, not the final destination of every digital behavior.
Complete Value-Creation Capability · Cₜ
people · capital · relationships · physical + digital assets
Business data
Trace
Software systems
Business specifications
Operating structure
O → O* → Rₜ
Describes how the organization works and how its assets are coordinated.
deterministic(Rₜ) → software ∈ CₜBusiness data
Called and governed by Rₜ; the main source of ground truth for Vₜ.
Trace
Produced by execution and returned to reflection—the only self-referential asset.
Software systems
Receive responsibilities that become deterministic enough to solidify out of Rₜ.
Business specifications
State how the business must behave for human readers and software development.
A practical boundary
The Operating Model describes how the organization works. A business specification describes how the business must behave. The former is implemented for an agent executor; the latter remains a human-readable requirement and an input to software development. Conflating them makes Rₜ larger without making execution more reliable.
04 · Reflective Bootstrapping
Rₜ must be able to shrink.
Trace, evidence, attribution, and candidate patches are agent work. Deciding what judgment deserves authority, why it should be deposited, and which exception should become precedent remains human work.
This does not eliminate dependence on people. It converts high-quality judgment from a one-time expense into a reusable asset—and therefore raises the standard of judgment required.
Add or correct
A failure, exception, or human correction during work.
Add a missing purpose or repair an inaccurate representation.
Retire
A model prior or new mechanism makes a rule redundant.
Remove the carrier only after every function it served has another explicit home.
Recheck premises
Periodic review rather than a visible failure.
Find cold drift: rules whose assumptions expired while outputs still looked plausible.
Solidify outward
A responsibility has become sufficiently deterministic.
Move its purpose into software, then remove the rule that carried it from Rₜ.
Coverage is not the governing metric
Rₜ is review-bandwidth constrained while O* drifts with reality. The meaningful measure is their rate relationship, not a snapshot coverage percentage. A high-coverage model that is six months stale is more dangerous than a smaller model that remains synchronized, because stale rules are executed with perfect consistency.
05 · Delegation boundary
Expressible does not mean delegable.
Rₜ supplies the organization-specific methods, precedents, authority, and division of labor that make delegation possible. It is necessary for institution-specific work, but it is neither sufficient nor a container for execution. Eₜ is a runtime outcome composed from representation, model priors, verification, and human response.
Eₜ = f(Rₜ, Pₜ, Vₜ, Hₜ)Represented model
The institution-specific purposes, constraints, authority, and precedents available to the agent.
Model priors
General capability the model already possesses without organization-specific instructions.
Verification environment
The tests, schemas, evaluators, feedback, and other means by which errors can be found cheaply.
Human-in-the-loop capacity
The speed and quality with which people answer decision requests, review changes, and absorb exceptions.
The marginal value of representation is concentrated where organizational purpose is not already supplied by Pₜ. As model priors improve, parts of Rₜ become redundant and should be retired rather than carried forever. The Operating Model is therefore not a permanent prompt library. It is the changing difference between what this institution needs and what its executor already knows.
Vₜ is often the harder constraint. A model that is correct nine times out of ten can still have negative value when the tenth error is silent. Tests, schemas, typed tools, reversible actions, regression cases, and observable business rules lower the delegation threshold. Because correctness ultimately needs ground truth, this is largely data and environment engineering—progress the organization can make without waiting for a generally smarter model.
Delegation quality
Escalation precision
The agent must identify the delegation boundary at runtime. Failure has two directions: crossing a boundary without escalation, and escalating so often that judgment cost is pushed back to people. An agent that escalates everything is perfectly safe and economically useless. Both precision and recall matter.
Representation quality
Rₜ is not monotonic
More rules create superlinear interaction costs: context competition, retrieval failure, and contradiction. Beyond an optimal scale Rₜ*, another document can make Eₜ smaller. Scope and routing mechanisms therefore matter more than document count. Solidifying stable responsibilities has a double return: they leave the capability boundary and free capacity for the rules that still require judgment.
Three causes of the execution boundary
PₜCapability
Model + harnessUnderstanding, context assembly, retrieval, tools, and environmental feedback improve with models and harnesses.
VₜVerification
Data + environmentDelegation expands when mistakes become cheap to detect. Ground truth, tests, schemas, and feedback are largely an organizational engineering problem.
HₜResponse
Attention + disciplineA decision request that no one answers breaks the responsibility. Response speed and judgment quality are runtime dependencies.
Idealized upper bound
Eₜ → agent-assigned(Rₜ)Rₜ itself records the division of labor. Technology determines how much can be realized inside the line; people decide where the line is drawn.
Operational measure
The sets are not directly measurable. Track human attention per unit of work: decision-request frequency, manual corrections, context restatement, and patch review time. The critical signal is their downward slope while outcome quality remains stable.
Division of labor must also be separated from discretion. Within a responsibility already assigned to it, an agent may gain broad freedom to choose methods, sequence work, and resolve familiar tradeoffs. Capability can move that internal boundary a long way. It does not decide which responsibilities were assigned in the first place.
06 · Invention under agent economics
Existing organizations optimize for human costs.
The ceiling is not the portion of today’s organization that can be copied. Existing operating models are local optima under expensive attention, lossy memory, high-friction trust, synchronous coordination, and years-long capability transfer. Agents alter those cost curves—and make previously uneconomical structures viable.
Cost collapse 01
Observation cost approaches zero
Standards can become continuously enforced constraints. Reflection can happen while trace evidence is still precise, and every rule can be periodically rechecked for cold drift.
Failure mode
Over-rigidity. Human organizations use ambiguity to absorb legitimate exceptions. Explicit exception channels and paths from exception to precedent must be designed at the same time.
Cost collapse 02
Replication cost approaches zero
One task and a thousand parallel tasks can run the same represented model without the behavioral decay that normally creates management layers.
Failure mode
Monoculture. A flawed rule is also copied perfectly. Progressive rollout, independent checks, and deliberate diversity must replace the accidental hedge of inconsistent human execution.
Cost collapse 03
Change cost approaches zero
Operating models can branch, run by scope, be compared against trace evidence, and roll back. Organization design becomes an experimental discipline.
Failure mode
Change outruns validation. A change budget and cooling period are required so that rules remain stable enough for people and agents to rely on them.
Cost collapse 04
Transfer cost approaches zero
Reusable operating capability can be inherited, forked, subscribed to, or purchased instead of rebuilt from scratch by every organization.
Failure mode
Anything transferable becomes available to competitors. Differentiation moves back toward real relationships, exclusive data, capital, market position, and judgment deposited from proprietary work.
These gains are architectural. Near-zero observation turns aspiration into constraint; replication removes layers built to fight communication loss; change makes organization design testable; transfer makes capability distributable.
Each also removes hidden safety. Inconsistency limits a bad rule’s blast radius; slow change stabilizes precedent; sampled review leaves room for exceptions. Agent-native design must restore these functions through staged rollout, exception channels, independent checks, change budgets, and cooldown periods.
The prerequisite
Deposit requires ownership.
Wage employment pays repeatedly for capability attached to a person. Efficient extraction turns that exchange into a one-time transfer; after the method is deposited, the contributor’s marginal value can collapse.
The rational response is shallow deposit: work gets done, but judgment, exception logic, and reasons never enter reflection. Normal outputs hide the loss.
People deposit deeply only when they retain a durable share in the value of what they deposit. The differentiator is not merely owning an Operating Model, but building an ownership structure in which people want to deepen it.
None of the four cost collapses bypasses delegation. Enforcing an unverified standard only scales an untested judgment; experimentation without measurable effects becomes random walk. Invention operates inside the region Eₜ can support, but changes the institution rather than reproducing it.
07 · Four paths
Two paths approximate. One reconstructs. One migrates.
The distinction is not difficulty but operator. Reflective Bootstrapping and model progress approach an existing target. Redesign asks whether the target’s shape still makes sense after the cost structure changes. Solidification changes the substrate carrying a responsibility.
Reflective Bootstrapping
Rₜ → O*Trigger · Work events
Expand coverage of organizational purposes through trace, attribution, governed behavior patches, and human judgment deposits. The target is not document volume.
Model + Harness Progress
Eₜ → RₜTrigger · Models, environment, discipline
Reliably perform more of the responsibilities assigned to agents through stronger priors, context assembly, verification, tools, and human response. The arrow is shorthand, not set containment.
Operating Model Redesign
expand OTrigger · Intentional review
Question the shape and premises of the Operating Model itself. Evidence comes from global patterns, changing cost curves, and possibilities that no single failure trace can reveal.
Solidification
deterministic(Rₜ) → software ∈ CₜTrigger · Human timing decision
Move stable responsibilities from probabilistic agent execution into deterministic software. Flexibility falls, while reliability and available Rₜ capacity rise.
A shared mechanism, three uses
Periodic whole-model review is simultaneously premise rechecking, cold-drift governance, and Operating Model Redesign. Agents prepare global evidence and candidate structures; people decide whether a rule is wrong—or whether the reason for having that kind of rule has disappeared.
The target quantity
System-borne work = Dₜ + EₜDₜ is work carried by deterministic software. Eₜ is work reliably carried by an agent. Maximizing Eₜ would perversely make solidification look like regression, because a responsibility leaves Eₜ when it becomes code.
New work enters Eₜ while judgment and adaptation are still required. Mature work settles into Dₜ, releasing agent capacity for the next frontier. Eₜ is best understood as a throughput aperture through which uncertainty becomes durable capability.
Together the four paths form a cycle: redesign decides what the organization should be able to do; reflection represents that purpose; model, environment, and discipline make it executable; solidification moves the stable portion into deterministic software and makes room for another round.
08 · The new company boundary
Certainty becomes copyable. Capacity to bear consequences does not.
A company adds value beyond an individual through two historically bundled pillars. Operating Model automation separates them.
Pillar 01 · Certainty
Predictable, repeatable, scalable outcomes.
Certainty comes from the Operating Model. Once represented, it can be inherited, distributed, purchased, and executed without losing fidelity as scale grows. A small team can run operating capability that once required a large hierarchy.
Pillar 02 · Liability capacity
The ability to absorb real consequences.
Capital, assets, legal structure, insurance, regulatory standing, longevity, and brand as a durable guarantee belong to Cₜ. None can be obtained by copying an Operating Model.
Accountability
Who owns the judgment and its consequences. It is bound to decision rights and ownership and cannot be transferred.
Liability capacity
How much an organization can pay and endure. It is bound to capital, legal structure, insurance, and persistence, and can be priced or outsourced. Insurance can pay for a decision; it cannot make the decision yours.
What remains differentiated in Cₜ
| Asset | Accumulates in operation | Purchasable | As models improve |
|---|---|---|---|
| Capital | Slowly | Yes | No depreciation |
| Relationships + brand | Yes | Difficult | No depreciation |
| Data + trace | Continuously | Usually no | Often appreciates |
| Operating Model | Yes | Yes | Partly depreciates as Pₜ grows |
Data is unusual: it accumulates through real work, is usually unavailable for purchase, and becomes more valuable as models improve. Trace is the mechanism that connects this proprietary evidence back to organizational learning. An organization able to carry greater liability also gains access to higher-value real-world work—and therefore to data that smaller competitors cannot manufacture.
Competition shifts from who can organize certainty to who can bear the consequences—and who accumulates data that cannot be bought.
Large organizations lose part of their coordination premium but retain balance sheets, regulatory capacity, and brands that guarantee persistence. Small owner-operator teams can match organizational certainty and bind judgment to consequences more tightly, but remain limited by the size of the liability they can carry and the work from which proprietary data can be earned. Once the Operating Model becomes copyable, what remains outside it becomes the reason the company still exists.
09 · Human attention
Scarcity determines the human role.
Human attention constrains both learning speed and execution reliability. The goal is not simply to use less of it. As Dₜ + Eₜ carries more production, attention should move toward the work where its marginal value compounds.
Execute
Perform the work directly.
Marginal value falls as Dₜ + Eₜ expands.
Restate
Rebuild context, repeat agreements, and correct the same class of deviation.
Marginal value should approach zero.
Deposit
Turn judgment, exception handling, and purpose into governed institutional assets.
Value compounds across every future execution.
Execution and restatement are repeated expenses. Deposit is leverage: one high-quality judgment can influence every future relevant execution, and its stable portion can eventually settle into deterministic code. The saved attention should fund better observation, reasoning, relationship-building, direction-setting, and new value creation—not lower-quality human involvement.
Outside the allocation problem
Final accountability is not another optional use of attention. It is a non-transferable obligation attached to judgment and consequence.
Conclusion
Toward the executable institution.
The inheritance view is table stakes: without it, an organization cannot preserve or delegate what it already knows. The invention view is the source of advantage: it redesigns the institution around costs that no longer apply. Neither view implies that the Operating Model is the company’s total digital capability.
The executable institution learns through Rₜ, acts through Eₜ, and turns mature behavior into Dₜ. Its Operating Model coordinates data and systems without absorbing them. Human attention leaves repetition for exploration, judgment, relationships, and new value creation, while accountability remains where real consequences land.