Skip to main content

43  Human-AI Organizations, Delegation, and Accountability

43.1 Chapter status

Field	Value
Chapter ID	human-ai-organizations-delegation-and-accountability
Part	Part II - Planning, Memory, Reasoning, and Execution
Status	conceptual
Manuscript maturity	v0.3 concept-complete argument-level manuscript
Claim label	Design rationale
Evidence level	argument
Source loading state	source notes: ext_nist_ai_rmf_1_0_2023, ext_moral_crumple_zones_2019, ext_ai_decision_authority_2020, ext_generative_ai_at_work_2025, ext_constructive_interdependence_human_ai_2026, talos, ext_human_ai_team_meta_analysis_2024, ext_human_ai_feedback_loops_2025, ext_eu_ai_civil_liability_2025; raw cache: talos
Test state	A 21-declaration Lean review model covers one bounded accountability-assignment invariant; no organization, worker, task, or intervention observed.

43.2 Drafting guardrail

This chapter does not prescribe one employment model, turn productivity into welfare, or treat a human name on an approval form as accountability. It owns the meso-level design of a human-AI organization; public legitimacy, economy- wide transition, and emergent multi-agent dynamics remain separate owners.

43.3 Human Reading Path

Concrete lens. A role matrix changes the assignee and forgets prior owners. The compositional route lowers authority, transfers control paths, and retains the residual chain.

An AI system almost never acts alone once its outputs matter to other people. It sits inside a team, firm, laboratory, government office, household, or network. People decide what work exists, who may delegate it, when an AI can decide, who must review, who receives the benefits, and who carries the consequences.

The organization needs more than a list of users and agents. It needs roles, competence requirements, decision rights, workload limits, separation of duties, escalation, conflict rules, accountability, succession, and a way to dissolve the arrangement. If the AI makes the real decision while a person absorbs the blame, the system has created a moral crumple zone. If the human retains a title but loses information, time, skill, or authority, human control is ceremonial.

Inspect the recurring arrangement rather than one approval screen. Ask whether the reviewer can change the decision in time, a worker can challenge it, delegation can silently pass through another agent, and responsibilities survive model or management changes. A credible arrangement answers before an incident and preserves responsibility, remedy, and residual obligations when it ends.

43.4 Problem

Labor OS can type one unit of work and Human Factors can ask whether an oversight action is usable. Neither owns the organization that repeatedly allocates tasks and authority among human and AI actors. Repetition changes skills, incentives, workload, information flow, bargaining power, and the meaning of review. An arrangement can improve throughput while concentrating decision power, degrading human competence, or assigning blame to the least powerful participant.

The external evidence makes the boundary concrete. NIST AI RMF supplies roles and lifecycle governance but does not design a particular organization. Elish’s moral-crumple-zone analysis warns against accountability being misallocated to a proximate human. Athey, Bryan, and Gans show in a formal model that authority allocation can alter human information-acquisition incentives and AI-reliability investment. Generative AI at Work reports heterogeneous worker effects in one customer-support deployment. Biswas et al. show that team reward can fail to reveal constructive interdependence. Together they argue against designing organizations from accuracy or throughput alone.

43.4.1 Exclusive job and adjacent boundaries

Adjacent owner	That owner keeps	Human-AI Organizations owns
Labor OS	One typed job, its inputs, authority, worker, output, and receipt.	Repeated allocation of roles, decision rights, workload, incentives, accountability, succession, and dissolution.
Human Factors	Whether an oversight or control action is exercisable.	Which actor is assigned that action and whether the organization preserves capacity to perform it.
Human Intent	Authorized purposes and contestable intent.	Organizational delegation of decisions under those purposes.
Inter-Stack Protocols	Pairwise identity, message, transaction, and exchange.	Internal organizational authority and responsibility across actors.
Multi-Agent Dynamics	Emergent population behavior, coalitions, collusion, cascades, and systemic externalities.	Deliberately designed meso-level organization and its lifecycle.

flowchart LR
  C["Organization charter and purpose"] --> R["Actor, role, competence, workload, and conflict registry"]
  C --> P["Affected parties, participation, rights, and benefit commitments"]
  P --> R
  R --> D["Decision-right and delegation graph"]
  D --> J{"Authority, information, time, and separation adequate?"}
  J -- "no" --> E["Escalate, reassign, train, slow, or suspend"]
  J -- "yes" --> W["Execute typed work and decision"]
  W --> O["Observe outcome, contribution, dependence, burden, and affected parties"]
  O --> A["Accountability, remedy, learning, and authority update"]
  A --> D
  A --> X["Succession or dissolution with residual custody"]

How to read the organizational-authority loop: accountability begins before a decision, in the charter and authority graph. Outcome review includes contribution, dependence, and burden rather than task reward alone. Authority can be reassigned, and the organization must survive or close without orphaning residuals.

43.5 Why existing approaches are insufficient

“Human in the loop” is not an organization. It says nothing about which human, what authority, which information, available time, incentives, workload, appeal path, or consequences.

Task routing is not decision-right allocation. The best-performing actor may not be the legitimate decision maker. Expertise, affected-party rights, separation of duties, conflicts, and reversibility matter.

Accountability after failure is too late. A named person cannot be responsible for a decision they could not inspect, change, stop, or appeal. Responsibility must track prospective authority and actual capacity.

Productivity is not organizational welfare. Faster work can coexist with deskilling, surveillance, work intensification, inequitable benefit, lower autonomy, or fragile dependence. Heterogeneous worker and affected-party outcomes must remain separate.

Task reward is not teamwork. A team can achieve the goal with one actor dominating and another cleaning up errors. Contribution and constructive interdependence are useful diagnostics, but even they cannot replace welfare, legitimacy, or accountability.

43.5.1 Strongest objection

Organization design is contextual, political, and legally situated; a generic AI stack should not encode one managerial ideology. The design therefore standardizes the questions, records, and failure routes—not the answer. The mechanism earns its place only if it makes authority, dependence, workload, accountability, and remedy more visible and improves decisions compared with a simple role-and-approval matrix.

43.6 Core Claim

[human-ai-organizations-delegation-and-accountability.core, label: Design rationale, support: argument] A human-AI organization is eligible to delegate consequential work only when a versioned organizational contract binds purpose and mandate; membership and affected parties; human, AI, and institutional identities; roles, competence, training, workload, and accessibility; information and decision rights; delegation depth, expiry, revocation, and separation of duties; conflicts, incentives, compensation, and benefit distribution; escalation, appeal, incident, and remedy paths; contribution, dependence, and outcome measures; accountability; succession; dissolution; and residual custody. A human signature, model accuracy, task reward, throughput gain, role label, or source-reported field result alone establishes neither meaningful control, legitimate authority, accountability, worker welfare, organizational resilience, support, readiness, release, transfer, nor SOTA.

Reader claim. Delegation moves bounded authority and current responsibility; it must not erase prior owners, reviewers, evidence custodians, appeal paths, or remedy obligations.

Operational rule. Every handoff must lower or preserve the authority ceiling, name the new accountable owner, keep review and evidence custody independent, transfer intervention and remedy paths, and retain prior owners in the residual chain. An incomplete handoff does not activate.

43.6.1 Worked two-hop delegation: responsibility follows the grant

Principal 1 grants bounded operation authority to owner 2 with ceiling 4. Owner 2 delegates to owner 3 at ceiling 3, then owner 3 delegates to owner 4 at ceiling 1. Each handoff changes the reviewer and evidence custodian, transfers intervention, appeal, and remedy paths, and carries an acknowledgment receipt. The final accountable owner and current delegate both equal 4; prior owners remain visible as residual owners [3, 2] rather than disappearing from the audit trail.

The compositional model rejects a handoff when responsibility moves without the matching authority transition, when reviewer and evidence roles collapse, when a ceiling widens, or when a remedy path disappears. Across the larger lifecycle, one adverse nine-event trace closes after contestation and remedy with authority reduced to zero. The consumer checks 39 routes, 156 lifecycle mutations, and 50 bridge mutations. These finite checks preserve organizational custody; they do not establish human control, legal accountability, remedy efficacy, or good organizational outcomes.

43.6.2 Accountable organizations beneath evolving work surfaces

The consolidated reference locates this organizational route beneath AI Work Surfaces, Agent Harnesses, and Organizational Absorption. The parent owns how AI products and harnesses absorb progressively larger units of work. This route continues to own the charter, actor and role graph, competence and workload, decision rights, delegation lineage, separation of duties, incentives, accountability, remedy, succession, dissolution, affected-party standing, and residual duties that make a work arrangement an organization rather than a larger interface.

That placement does not let adoption, autonomy, or task coverage stand in for legitimate mandate or accountable outcomes. This route retains its claims, source queue, proof targets, tests, fixtures, failure modes, evidence exit, non-claims, support ceiling, identity, and legacy URL. The parent does not inherit valid delegation, meaningful control, legal responsibility, remedy, fairness, or organizational success; this route does not inherit harness capability, observability, rollback enforcement, productivity, adoption, or safe absorption. The nest creates no employment, welfare, accountability, authority, support, deployment, release, AGI, or ASI result.

43.7 Mechanism

1. Charter the organization. State purpose, mandate, affected parties, prohibited ends, jurisdictional assumptions, lifespan, and owners.
2. Register actors and roles. Bind human, AI, team, and institutional identities to competence, training, workload, accessibility, conflicts, and maintenance state.
3. Map decision rights. For each decision class, name who proposes, advises, decides, executes, reviews, can stop, can appeal, and receives notice.
4. Constrain delegation. Limit depth, duration, scope, subdelegation, automation, and model replacement; preserve revocation and reversion.
5. Separate duties. Avoid one actor generating, evaluating, authorizing, executing, and closing a consequential action without independent checks.
6. Observe real work. Record task outcomes beside contribution, interdependence, error recovery, workload, skill change, autonomy, access, benefit, harm, and affected-party distribution.
7. Assign accountability prospectively. Accountability follows actual information and authority; impossible oversight becomes an organizational defect, not personal blame.
8. Adapt or dissolve. Rebalance roles, train, narrow automation, compensate, transfer custody, or close the organization with residuals intact.

The decision-right graph is typed by consequence, not merely by department. For each decision class it distinguishes recommendation, option generation, selection, approval, execution, veto, audit, appeal, remedy, and policy change. An actor may hold several roles, but the contract exposes that concentration and states when independent review is mandatory. Authority is granted only when information, competence, time, accessibility, workload, and conflict conditions make the role usable in practice.

Organizational evidence is longitudinal. A deployment can raise immediate throughput while reducing skill, situational awareness, bargaining power, or capacity to recover later. The outcome ledger therefore measures learning, dependency, review quality, error discovery, workload, distribution of gains and losses, worker and affected-party option value, and what happens when the AI is unavailable or wrong. Counterfactual staffing and automation alternatives remain visible instead of disappearing after adoption.

Delegation changes the institution’s incentives. A vendor may benefit from automation while a worker bears review liability; a manager may optimize average speed while a minority receives more failures; an AI evaluator may reward the behaviors it generates. Conflict-of-interest and benefit records join the authority graph, and material conflicts narrow the route or require a separate owner.

Succession and dissolution are ordinary lifecycle states. The organization must transfer role identity, credentials, pending appeals, evidence, benefits, liabilities, learned procedures, affected-party notices, and residuals when a person leaves, a model changes, a vendor fails, or the institution closes.

43.8 Concept-by-concept organizational contract

The eight concepts below divide the chapter’s composite organizational claim into decisions that can fail independently. They define the manuscript’s editorial responsibilities; no organizational intervention or legal conclusion is implied.

43.8.1 Charter, mandate, and affected-party standing

An organization needs an explicit reason to exist before it allocates work. The charter states the mandate, prohibited ends, lifespan, jurisdictional assumptions, funding and ownership interests, affected populations, and who may change or terminate the arrangement. Affected-party standing is not equivalent to employment or system access: customers, patients, residents, families, competitors, and future workers can bear consequences without holding an account. The contract records how those groups are identified, represented, notified, heard, and connected to appeal and remedy.

Mechanism. Version the charter with its authorizing institution and map each decision class to intended beneficiaries, burdened groups, excluded populations, participation routes, prohibited effects, review date, and termination authority. Changes to purpose or affected-party scope invalidate delegation derived from the earlier charter.

Failure mode. Mission drift lets a workflow optimized for one purpose quietly become surveillance, discipline, eligibility screening, or cost cutting. A nominal consultation can omit people with less bargaining power, while an institution treats legal access or technical feasibility as social authorization.

Non-claim. A complete charter does not establish consent, legitimacy, lawfulness, representativeness, or that the chosen mandate is morally correct. Those judgments remain with constitutional, moral, legal, and public- institution owners.

Source grounding. ext_nist_ai_rmf_1_0_2023 contributes lifecycle context, roles, and risk-governance functions but no certified organization design. talos supplies typed-work and explicit-contract lineage as a speculative system design; it does not establish stakeholder standing or institutional authority.

43.8.2 Actor, role, competence, workload, and accessibility state

A role title is not operational capacity. The actor registry binds every human, AI, team, vendor, and institution to the exact role it can perform, the information it receives, required competence, current training, workload, availability, accessibility needs, conflicts, and maintenance state. Human and AI competence are versioned differently: a model update can invalidate an AI role, while fatigue, turnover, deskilling, task novelty, or inaccessible interfaces can invalidate a human control path.

Mechanism. For each role and decision class, define prospective competence tests, maximum concurrent burden, time-to-intervene, accessible control interfaces, required context, substitution rules, and expiry. Runtime admission checks current state rather than trusting the organization chart. A failed check routes to reassignment, training, slower work, narrower automation, or suspension.

Failure mode. Role-label laundering names a reviewer who lacks time, information, skill, language access, physical access, or power to intervene. Workload aggregation can make individually plausible approvals collectively impossible, and automation can erode the unaided competence needed during failure or withdrawal.

Non-claim. Passing an authored competence field does not prove real competence, accessibility, vigilance, or human control. Synthetic role fixtures cannot substitute for ethically reviewed observations of people.

Source grounding. ext_ai_decision_authority_2020 shows in a bounded model that authority allocation can change human information-acquisition and AI- reliability incentives. ext_generative_ai_at_work_2025 reports heterogeneous effects in one customer-support deployment. Neither supplies universal competence thresholds or a longitudinal workforce result.

43.8.3 Decision rights and meaningful intervention capacity

“Human in the loop” hides several different rights. For each consequential decision, the organization must say who may frame options, recommend, select, approve, execute, veto, pause, audit, appeal, remedy, and change policy. Information access and timing accompany those rights: an actor cannot meaningfully stop a decision after the effect is irreversible or review evidence they cannot inspect. Decision rights can be distributed, but the distribution must preserve a reachable owner for the outcome.

Mechanism. Compile a typed decision-right graph from charter and role state. Trace every effect path backward to proposal, authorization, execution, intervention, review, and remedy nodes; require current competence, information, time, and practical authority at each mandatory human control point. Exercise stop and appeal paths with seeded cases before consequential use.

Failure mode. Rubber-stamp review gives a person a binary approval under time pressure after the model has framed every option. Accountability diffusion splits decisions into many steps so that no actor can see or own the whole outcome. An AI-generated assessment can also become its own nominally independent approval.

Non-claim. A connected graph does not establish that the decision is legitimate, wise, lawful, or safe. It records and tests organizational reachability; Human Factors still owns whether controls are usable in practice.

Source grounding. ext_moral_crumple_zones_2019 motivates matching responsibility to actual control but provides no complete remedy. ext_nist_ai_rmf_1_0_2023 provides governance-role vocabulary without certifying the graph, and talos contributes typed execution lineage rather than evidence of meaningful human intervention.

43.8.4 Delegation, subdelegation, expiry, and revocation

Delegation changes who may decide; routing changes who performs a task. They are not the same operation. A delegation lease names principal, delegate, decision class, purpose, scope, duration, consequence ceiling, permitted subdelegation, required review, model and tool identity, revocation authority, and reversion path. The lease is re-evaluated when workload, competence, purpose, model, vendor, or institutional conditions change.

Mechanism. Make every delegation edge explicit and depth-bounded. A runtime check resolves the complete chain to an accountable principal, verifies non-expiry and non-expansion, and rejects hidden replacement or subdelegation. Revocation stops new effects, reconciles work already in flight, restores a qualified fallback, and preserves affected appeals and residuals.

Failure mode. Delegation laundering treats an API call, vendor service, or secondary agent as implementation detail even though it exercises independent judgment. Temporary assistance becomes an automation ratchet when the human fallback loses skill or staffing. Revocation theater disables one route while copies, queues, credentials, cached decisions, or downstream agents continue.

Non-claim. An expiring lease does not guarantee enforceable revocation, complete rollback, or lawful delegation. It also does not authorize a broader purpose or permit the delegate to create its own successor authority.

Source grounding. ext_ai_decision_authority_2020 supplies an incentive-aware authority-allocation comparator, not a deployed delegation control. talos supplies typed job, contract, replay, and approval lineage; its architecture is not evidence that revocation or subdelegation control works.

43.8.5 Separation of duties, conflicts, incentives, and benefits

An actor that generates, evaluates, authorizes, executes, and closes the same consequential action can hide errors and optimize the evidence used to judge it. Separation of duties creates independent-enough challenge where consequence warrants it. Independence is not just a different username: shared models, prompts, data, vendors, reporting lines, incentives, and performance metrics can make nominally separate roles one failure domain. Benefits and burdens must be visible because they shape which failures an organization is motivated to notice.

Mechanism. For each decision class, compile prohibited role combinations and a dependency graph among generators, evaluators, approvers, operators, and auditors. Record financial, career, vendor, model, data, and metric conflicts; who receives productivity gains; who absorbs review, error, surveillance, and remedy costs; and which independent route can reopen the decision.

Failure mode. Self-approval, evaluator capture, and shared blind spots make paper separation ineffective. A manager can optimize average throughput while workers or affected minorities bear errors, or a vendor can validate the same model from which it earns revenue. Excessive separation can also create delay and responsibility diffusion without better challenge.

Non-claim. More reviewers are not automatically more independent, fair, or effective. The contract does not determine just compensation, lawful conflicts, or legitimate benefit distribution.

Source grounding. ext_nist_ai_rmf_1_0_2023 supports explicit governance roles. ext_generative_ai_at_work_2025 provides bounded evidence that workplace effects can be heterogeneous. ext_eu_ai_civil_liability_2025 adds a jurisdiction-specific institutional comparator; none proves this separation design.

43.8.6 Longitudinal contribution, skill, dependence, and burden

Task reward and short-run productivity cannot describe an organization. A combined system may outperform a human while underperforming the AI alone, or it may succeed because one participant repairs the other’s errors under rising burden. Repeated interaction can change skill, trust, judgment, dependence, workload, bargaining position, and the ability to recover after tool loss. Those trajectories need shared populations, time points, and attrition denominators.

Mechanism. Compare human or team alone, the AI-operated process when legitimate, the combined process, and relevant conventional workflow alternatives. Track useful quality, contribution, constructive interdependence, review and rework, unaided and assisted competence, calibration, workload, autonomy, accessibility, benefit, harm, withdrawal recovery, turnover, and who is missing from later measurements.

Failure mode. Synergy laundering compares the team only with the human, while productivity laundering treats output as welfare. Short studies can miss deskilling and lock-in; survivorship can exclude people who leave; a dependence metric can be optimized even when dependence reduces human option value.

Non-claim. The chapter does not assert universal complementarity, harmful co-adaptation, durable skill change, or economy-wide labor effects. Synthetic humans remain debugging fixtures rather than participant evidence.

Source grounding. ext_human_ai_team_meta_analysis_2024 motivates the three-arm strongest-component comparison. ext_constructive_interdependence_human_ai_2026 adds a study-specific cooperation measure, while ext_human_ai_feedback_loops_2025 motivates longitudinal coupled-state tracking. Their tasks and populations bound every inference.

43.8.7 Accountability, causation, evidence access, and remedy

Accountability begins before failure by aligning prospective authority with the capacity to know and intervene. After harm, it requires more than finding the nearest human: preserve evidence about design, deployment, delegation, model and policy changes, review, actual effects, institutional decisions, and who controlled each route. Affected parties need an identifiable entity, usable evidence access, timely appeal, correction, compensation or other remedy, and a route that remains available when organizations disagree.

Mechanism. Compile an accountability map from the decision-right and delegation graphs, then join it to effect and incident lineage. Flag responsibility assigned without matching information, competence, time, or control. Preserve contestable causation hypotheses, disclosure and privacy limits, insurer or compensation routes, appeal status, remedy execution, and unresolved ownership instead of generating an automatic fault verdict.

Failure mode. A moral crumple zone makes a proximate worker absorb blame for system-level choices. Technical logs can be withheld, selectively framed, or mistaken for legal causation. Formal appeal can be too slow, expensive, inaccessible, retaliatory, or powerless to alter the effect.

Non-claim. The organizational record is not legal advice, a liability judgment, a proof of causation, or evidence that remedy is adequate. Legal and institutional authorities retain those decisions.

Source grounding. ext_moral_crumple_zones_2019 supplies the responsibility-without-control failure. ext_eu_ai_civil_liability_2025 describes a European policy space involving responsible operators, evidence, causation, insurance, and remedy. Neither is a universal legal rule or a tested ASI Stack accountability intervention.

43.8.8 Succession, dissolution, continuity, and residual custody

Human-AI organizations change even when their tasks do not. People leave, models and vendors are replaced, institutions merge, contracts expire, and systems are deliberately retired. Succession transfers a live organizational role; dissolution closes the arrangement. Neither transition may erase pending decisions, appeals, evidence, benefits, liabilities, learned procedures, credentials, affected-party notices, or harms that outlive the technical system.

Mechanism. Before succession, requalify the replacement actor, model, tool, and institution; migrate only compatible authority; rotate credentials; notify affected parties; and preserve prior decision lineage. Before dissolution, stop or transfer work, reconcile effects and queues, preserve required records, fund or assign pending remedies, restore practical alternatives, and name every residual owner and review trigger. Unowned obligations block clean closure.

Failure mode. Model replacement inherits stale competence and delegation, vendor failure strands appeals and evidence, or a dissolved entity abandons remedy. Technical rollback can be institutionally impossible after layoffs, deskilling, procurement lock-in, or public-capacity loss. Over-retention can also violate privacy or rights.

Non-claim. A succession receipt does not prove equivalent competence, continuity, complete erasure, or that affected parties have been made whole. The book has not exercised organizational dissolution.

Source grounding. ext_nist_ai_rmf_1_0_2023 motivates lifecycle governance, and talos contributes artifact, replay, and residual-custody lineage. ext_human_ai_feedback_loops_2025 supports treating dependence and human state as longitudinal variables. None establishes successful succession, dissolution, or recovery here.

43.8.9 Organizational transition: tasks, jobs, power, and public capacity

The unit of change is not simply “a job exposed to AI.” A job bundles tasks, relationships, tacit knowledge, legal duties, bargaining arrangements, learning pathways, and responsibility. Automation can remove one task while increasing review, exception handling, customer interaction, or liability. It can also remove entry-level work that trained future experts. The transition ledger maps task creation, substitution, complementarity, supervision, rework, and coordination burden before making a job-level claim.

Adoption is a diffusion process, not an instantaneous technical capability. Integration cost, data and workflow constraints, trust, regulation, capital, vendor dependence, and managerial incentives determine uptake. A field productivity result can establish a bounded task effect; it does not by itself forecast occupation-wide employment, wages, firm concentration, public-service capacity, or economic growth.

Distribution is a first-class outcome. The record follows wages, hours, workload, surveillance, accessibility, error liability, bargaining power, ownership, consumer benefit, vendor rents, geographic concentration, and affected parties who never touch the system. More output can coexist with deskilling, work intensification, exclusion, or a moral crumple zone where the least empowered human absorbs blame for a decision they could not change.

Skill is dynamic. AI may scaffold learning and widen access; it may also replace deliberate practice, hide causal structure, or turn professionals into exception handlers who cannot recover when automation fails. Qualification tracks unaided and assisted performance, error detection, novel-case transfer, recovery after tool loss, mentoring capacity, and who receives training or advancement.

Concentration and public capacity need separate denominators. Shared models, clouds, data suppliers, and evaluators create common-mode risk and bargaining asymmetry across nominally independent organizations. Public institutions may gain service capacity or become dependent on vendors whose objectives and continuity they do not control. Procurement, interoperability, exit, local capability, audit rights, continuity, and public-interest obligations belong in the organizational contract.

A serious comparison includes workflow redesign, conventional software, staffing, training, reduced workload, decision support, partial automation, collective governance, and full delegation under matched obligations. It reports output, quality, distribution, skill, autonomy, resilience, accountability, infrastructure cost, and option value separately. Economy-wide inference remains another modeling and evidence problem.

flowchart LR
    J["Job and institutional mandate"] --> T["Task decomposition"]
    T --> A["Automation / augmentation alternatives"]
    A --> W["Workflow and decision-right redesign"]
    W --> O["Output, burden, skill, distribution"]
    O --> P["Bargaining, remedy, public capacity"]
    P --> R{"Renew, narrow, retrain, or exit"}
    R --> W

43.8.10 Required artifacts

HumanAIOrganizationContract {
  charter_mandate_lifespan_and_affected_parties,
  actor_role_competence_workload_and_accessibility_registry,
  information_and_decision_rights_graph,
  delegation_subdelegation_expiry_and_revocation,
  separation_of_duties_and_independence_matrix,
  conflict_incentive_compensation_and_benefit_record,
  escalation_stop_appeal_incident_and_remedy_routes,
  contribution_interdependence_outcome_and_burden_ledger,
  accountability_map,
  succession_dissolution_and_residual_custody,
  non_authorities
}

43.9 Interfaces

Human Intent supplies authorized purpose. Labor OS supplies typed work and receipts. Human Factors supplies control-envelope evidence. Security and Runtime Adapters supply identities and actual permissions. Claim Ledgers and Operations receive organizational outcomes and residuals. Inter-Stack Protocols govern external counterparties; Multi-Agent Dynamics observes population consequences. No interface allows a role title to substitute for competence or a completed task to settle accountability.

Privacy and Data Rights governs employee, customer, and affected-party data, including monitoring and derived performance records. Moral Uncertainty and Constitutional Alignment supply protected predicates and unresolved value conflicts; an organization cannot vote itself broader technical authority. Resource Economics receives human review, training, delay, coordination, appeal, and recovery cost rather than counting only model tokens.

Artifact Graphs preserves decision, delegation, review, intervention, outcome, and remedy receipts. Capability Replacement invalidates role competence and delegation leases when the underlying model or tool changes materially. Readiness can require organizational evidence for a named deployment, but it cannot establish legitimacy or legal compliance. Institutions and Multi-Agent Dynamics consume aggregate signals while preserving the difference between a deliberately designed organization and an emergent population.

43.10 Invariants

These constraints keep an organization from using paperwork to counterfeit control. They bind responsibility to actual information and intervention capacity, keep burden and benefit visible, and ensure that reorganization, model replacement, or closure cannot erase rights, remedies, records, and residual obligations created earlier.

• Decision authority, execution authority, review authority, and accountability are explicit and may differ.
• A person cannot be accountable for an action they lacked information, competence, time, authority, or practical ability to change.
• AI-generated evaluation cannot be its own independent approval.
• Delegation is scoped, expiring, revocable, and depth-bounded.
• Model replacement expires competence and role evidence where material.
• Workload, accessibility, conflict, and incentive state travel with a role.
• Task reward, contribution, interdependence, welfare, and legitimacy remain separate outcome axes.
• Affected parties retain named notice, appeal, and remedy routes.
• Succession and dissolution preserve artifacts, obligations, and residuals.
• Organizational records create no public or legal legitimacy by themselves.
• Review burden, training cost, and benefit distribution remain attributed to the actors and decisions that cause them.

43.11 Evidence

The current source packet spans governance, theory, field evidence, and measurement. Its diversity is useful because no one source can establish the chapter. The first campaign should examine a real but low-consequence recurring workflow with prospective participants and appropriate ethics/privacy review. Compare a conventional human workflow, AI advice, AI execution with human approval, competence-aware dynamic delegation, and a deliberately simple role-and-approval baseline. Freeze tasks, role training, escalation routes, model versions, positive controls, and independent scoring before held-out work.

Measure useful output and errors, but also information acquisition, intervention success, review time, workload, skill retention, constructive interdependence, automation bias, appeal, affected-party distribution, accountability accuracy, recovery, worker experience, access, and cost. A participant study or organizational intervention cannot begin from repository authority alone. Synthetic human models may support debugging but remain separate from evidence about people.

43.12 Failure modes

• moral crumple zone;
• rubber-stamp or impossible review;
• authority without accountability or accountability without authority;
• hidden subdelegation;
• automation bias and deskilling;
• work intensification or surveillance;
• conflict of interest and incentive gaming;
• separation-of-duties collapse;
• contribution or interdependence mismeasurement;
• benefit concentration and harm externalization;
• inaccessible oversight or appeal;
• succession failure and orphaned residuals.

Further failures include role-label laundering, where “reviewer” is recorded without time or control; accountability diffusion, where every actor owns a small step and nobody owns the outcome; and automation ratchet, where temporary assistance removes the human competence needed for rollback. Metric capture can optimize task reward while degrading service quality, worker welfare, or affected-party rights. Vendor opacity can make competence, incident, or change evidence unavailable. Appeals can be formally present but too slow, costly, inaccessible, or retaliatory to use. A technically reversible workflow can become institutionally irreversible through layoffs, lost skills, contract lock-in, or redistributed liability.

43.13 Minimum Viable Implementation

Represent one recurring workflow with four decision classes and at least three actor roles. Compile a decision-right graph, competence and workload checks, delegation expiry, independent approval, stop and appeal routes, and accountability receipts. Use synthetic cases to reject rubber-stamp approval, conflicted review, overloaded humans, unauthorized subdelegation, and model replacement without requalification before any human study.

The minimum honest implementation includes a versioned organization schema, decision-class inventory, actor and role registry, competence/workload state, typed rights graph, delegation leases, conflicts and benefits, stop/appeal routes, accountability mapping, and succession receipt. A validator rejects missing affected parties, role without a usable authority path, accountability without control, self-approval, stale model competence, overloaded review, unbounded subdelegation, and orphaned residuals.

Synthetic cases are only the first artifact. They must show that the control plane changes routing and catches seeded defects while allowing legitimate work. Passing those cases is not evidence about human behavior, organizational effectiveness, worker welfare, legitimacy, or legal liability. Support remains argument until an ethically reviewed natural or high-fidelity workflow compares competent alternatives with independent outcome and affected-party evaluation.

43.14 Mature Research Target

The mature target is an organizational control plane that can simulate and audit decision rights before deployment, measure how automation changes human skill and option value over time, and reallocate authority without losing traceability or remedy. It would support many organizational forms rather than hard-code one. It would integrate collective bargaining, accessibility, affected-party representation, and dissolution while refusing to infer legitimacy from technical coherence.

Beyond current practice, the target would make organizational design counterfactual and testable. Before deployment it could compare human-only, advice, approval, execution, mixed-team, and dynamically delegated structures under the same task and consequence model. During operation it would estimate review capacity, dependence, skill change, contribution, intervention effectiveness, distributional outcomes, and recovery readiness, and it would reallocate authority or slow work when those conditions deteriorate.

The research program must use real or realistic recurring work with multiple decision classes, heterogeneous roles, consequential but ethically acceptable errors, and affected-party feedback. Comparators should include mature workflow and process controls, ordinary role matrices, strong human-AI assistance, static approval, competence-aware delegation, and the full organizational contract. Outcomes include useful quality, error and recovery, false approval, missed help, review time, workload, training, skill retention, autonomy, accessibility, appeal, accountability accuracy, benefit and harm distribution, latency, cost, and resilience to AI or vendor loss.

Causal ablations remove workload checks, competence leases, separation of duties, contribution measures, affected-party routes, or succession custody. Transfer spans domains, organization sizes, labor arrangements, models, jurisdictions, and time. Independent evaluators and participant safeguards are mandatory, and synthetic people cannot stand in for human evidence. This is a falsifiable research program, not a current result: no such campaign has passed here, so the technical record establishes no superior organization or institutional legitimacy.

43.15 Formalization hooks

43.15.1 Implemented formalization

lean:human_ai_org.accountability_requires_authority is implemented in AsiStackProofs.HumanAIOrganizations as three bounded models with 62 theorem declarations. The retained five-stage finite review covers identity, capacity, intervention authority, independence, and remedy/custody. A stage invariant is preserved for one step and by induction over an arbitrary finite run; if the run reaches accountabilityAssignable, all 20 authored assignment fields are true. One complete record reaches that state, while 20 closed mutations cover every assignment field and reach the exact refusal or repair state.

The separate ten-stage delegation-to-remedy lifecycle binds nine decision, delegator, delegate, policy, authority, reviewer, evidence, remedy, and result identities across delegation, activation, escalation, handoff, contestation, authority expiry, incident reconstruction, remedy, and closure. Arbitrary successful runs preserve those identities and zero support/external-effect authority, account for exactly one receipt per event, keep contest and remedy receipts monotone, expose accepted traces, compose across all ten prefix/suffix splits, and reject every event after closure. One nine-event adverse-path witness closes with authority zero and one contest and remedy receipt. The independent consumer reaches all 39 routes and rejects 156/156 lifecycle mutations spanning identity, stage, replay, over-ceiling delegation, missing process records, authority requests, and post-closure events.

The third model composes organizational responsibility with the independently defined authority-delegation chain. Every accepted handoff refines an accepted authority step, assigns the exact child delegate as accountable owner, retains the prior owner in residual custody, adds exactly one responsibility receipt, and preserves reviewer and evidence-custodian separation plus zero support and external-effect authority. These properties hold over arbitrary successful runs and across event-batch composition. A concrete two-hop witness delegates from owner 2 to 3 to 4 while attenuating authority to read-only, retaining residual owners [3, 2], and aligning two responsibility receipts with two authority receipts. The independent consumer checks all three bridge compositions and rejects 50/50 bridge mutations without changing the rejected state. A paired countermodel also gives a safe record and an owner/reviewer/evidence gap the same aggregate delegation summary; no classifier over only that summary can recover whether accountability is assignable.

This is a mechanized property of a finite authored record. Lean does not prove that any identity, assignment, review, evidence, handoff, or control field is truthful or usable, that a person could intervene in a real workflow, that accountability is lawful or legitimate, or that an organization is effective, fair, resilient, or safe. Chapter support therefore remains argument.

43.16 Codex test plan

Test	Purpose	Status
Decision-right reachability	Ensure each modeled assignment traverses identity, capacity, authority, independence, and remedy stages before becoming assignable.	implemented in Lean; no runtime workflow
Moral-crumple-zone mutation	Reject accountability assigned to an actor without real authority or capacity.	implemented for 20 exact authored-record mutations
Delegation-to-remedy lifecycle	Exercise bounded delegation, escalation, handoff, contestation, expiry, reconstruction, remedy, and closure with exact custody and non-authority checks.	implemented in Lean and independently reconstructed; no runtime workflow
Authority/accountability refinement	Prevent accepted subdelegation from creating an owner gap, erasing residual custody, collapsing review, or widening authority.	implemented in Lean with a two-hop witness, three bridge compositions, and 50/50 bridge mutations; authentic identities and controls untested
Aggregate-summary countermodel	Show that depth, receipt count, authority ceiling, and residual count cannot recover missing accountability identities.	implemented as a paired Lean witness and independent summary collision
Workload and competence positive controls	Verify the finite route narrows when authored workload or competence fields fail.	implemented in Lean; real human capacity untested
Contribution/interdependence audit	Prevent task reward from standing in for teamwork.	planned

43.17 Organizations Are Persistence Surfaces

Organizations learn through procedures, role definitions, incentives, approval rules, staffing, procurement, and shared narratives. Those artifacts can be more durable and consequential than a model update. Adjudicated persistence therefore includes institutional loci in the same placement decision while preserving their distinct legitimacy requirements.

Technical write access is not institutional authority. A system may propose that an incident should change a routine, but affected-party standing, contestability, labor and privacy obligations, remedy, and rightful approval remain with the relevant human institution. A technically effective rule can still be inadmissible or become adaptation debt when no accountable owner can review or retire it.

43.18 Source crosswalk

Source	Contribution	Boundary
ext_nist_ai_rmf_1_0_2023	Lifecycle roles and risk-governance vocabulary.	Framework; no organization design or effectiveness result.
ext_moral_crumple_zones_2019	Accountability-misallocation failure model.	Conceptual/empirical analysis, not a complete remedy.
ext_ai_decision_authority_2020	Incentive-aware human/AI authority allocation.	Bounded economic model, not a universal law.
ext_generative_ai_at_work_2025	Heterogeneous field effects in customer support.	One deployment setting; no economy-wide inference.
ext_constructive_interdependence_human_ai_2026	Team-dependence measure beyond task reward.	Domain- and metric-bound; not reproduced.
talos	Corben’s Labor OS and typed-work lineage for roles, work contracts, delegation, and review.	Speculative system design; no organizational, legitimacy, accountability, or welfare result.

43.18.1 Manifest source assignment reconciliation

These rows keep Human-AI Organizations, Delegation, and Accountability’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.

Source	Intake role	Boundary
ext_human_ai_team_meta_analysis_2024	Metadata-first comparator: When combinations of humans and AI are useful: A systematic review and meta-analysis. Preregistered synthesis of 106 experiments and 370 effect sizes using human-alone, AI-alone, and combined-system comparisons. The aggregate findings are task- and population-bound and do not establish universal human-AI synergy or longitudinal benefit.	No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row.
ext_human_ai_feedback_loops_2025	Metadata-first comparator: Human-AI feedback loops alter human perceptual, emotional and social judgements. Experimental evidence that repeated human-AI interaction can create feedback dynamics in studied judgment tasks. It supports measuring coupled trajectories, not a universal claim about all users, systems, settings, or long-term clinical outcomes.	No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row.
ext_eu_ai_civil_liability_2025	Metadata-first comparator: Artificial intelligence and civil liability. European Parliament research service study of AI and civil-liability questions. It supports explicit causation, evidence-access, insurance, compensation, and remedy analysis but is not legal advice or a globally settled liability rule.	No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row.
adjudicated_persistence	Passage-reviewed comparator: Adjudicated Persistence: Governing the Transition from Experience to Durable Structure in Adaptive Systems. Treats procedures, norms, role definitions, and institutional rules as persistence surfaces requiring legitimacy, contestability, and remedy rather than mere technical write access.	Conceptual author framework and benchmark proposal; no local implementation, empirical result, independently checked proof, safety result, or support movement. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.

43.19 Co-adaptation is an outcome, not a premise

An organization cannot infer “human-AI complementarity” from adoption or from a team beating the human alone. The three-arm evidence rule from ext_human_ai_team_meta_analysis_2024 applies at the organizational level: compare the human or team without the AI, the AI-operated process where that is a legitimate comparator, and the combined process. The strongest competent component is the performance floor.

Longitudinal feedback can change expertise, trust, workload, escalation habits, and what work the organization retains. ext_human_ai_feedback_loops_2025 therefore adds trajectory fields to the delegation record: unaided skill, calibration, override quality, dependence, distribution of learning, turnover, error inheritance, and recovery after withdrawal. A short productivity gain can coexist with brittle expertise and growing vendor dependence.

The organization must also distinguish an accountable role from a liability sink. When harm occurs, evidence access, causation, duty, insurance, compensation, and remedy need an executable route. Civil-liability analysis can inform that design but does not supply one universal legal answer. A human reviewer is not automatically responsible for defects they could not inspect, contest, or prevent; the model is not a legal person that can absorb duties assigned by deployers, vendors, or institutions.

43.20 Summary

Human-AI organization design governs the repeated allocation of information, work, authority, burden, benefit, and accountability. Its purpose is not to choose one institutional form but to prevent throughput and role labels from hiding who actually decides, who can intervene, who bears the cost, and how the arrangement changes or ends.

The central rule is that responsibility follows prospective information, competence, time, authority, and practical control. Recommendation, decision, approval, execution, veto, audit, appeal, remedy, and policy change remain separate rights even when one actor holds several. Delegation is scoped, expiring, revocable, and requalified when the model, task, workload, or institution changes.

Outcomes include contribution, dependence, skill, burden, access, benefit, harm, recovery, and affected-party experience alongside useful task quality. This prevents short-run throughput from hiding deskilling, surveillance, fragile dependence, or concentrated gains. Succession and dissolution preserve pending appeals, artifacts, liabilities, learned procedures, and residuals so temporary automation does not become an unowned institutional ratchet. The result is a governable organization contract, not a claim that one structure is universally legitimate.

43.21 Handoff

The organizational contract hands its roles, decisions, workload, incidents, appeals, and outcome receipts to Human-AI Symbiosis, Neurotechnology, and Cognitive Sovereignty. That chapter asks what repeated coupling changes in the person and the AI: complementarity, feedback, skill, dependence, neural and inferred mental data, consent, practical exit, and longitudinal recovery. Organizational accountability cannot stand in for cognitive sovereignty or evidence that the combined system beats its strongest component.

43.22 Sources

See the source crosswalk above and the generated external-source appendix.