The Human^AI and Org^AI framework, layer by layer

ENGIntellectual augmentation

Engelbart, 1962

The augmentation tradition treats the computer as part of a system through which people formulate, understand and solve more complex problems. Human^AI continues this perspective while adding the task, context, interface, validation and delegation to the unit of analysis.

• A model in the abstract cannot explain a person’s performance or risk.

Depends on: —
Feeds: Iᵢ Interface and mode of interaction

DCDistributed cognition

Hollan, Hutchins and Kirsh, 2000

Cognitive activity can be distributed among people, artefacts and the environment. The framework takes up the idea but separates situated collaboration from institutional capacity.

Depends on: —
Feeds: X Task, context and constraints

HIHybrid intelligence

Dellermann et al., 2019

Configurations in which human and artificial contributions are complementary. Complementarity is not assumed: it is measured against comparators.

Depends on: —
Feeds: —

ORGOrganisation, authority and knowledge

Jarrahi, 2018 · Faraj et al., 2018 · Raisch and Krakowski, 2021 · Nonaka, 1994

At organisational level, AI changes roles, workflows, coordination and authority, not only individual productivity. Organisational knowledge is not an accumulation of documents: it requires ownership, context, validity, memory and mechanisms through which conclusions are validated before reuse.

Depends on: —
Feeds: Korg Authorised institutional knowledge

FRTPerformance, frontier and homogenisation

Vaccaro et al., 2024 · Brynjolfsson et al., 2025 · Dell'Acqua et al., 2026 · Doshi and Hauser, 2024 · Bommasani et al., 2021

Evidence on human–AI combinations is heterogeneous. Improvement over an unaided person does not demonstrate synergy over the best available comparator. Better individual generation may reduce collective diversity, and defects in a foundation model can propagate into downstream systems.

• Evaluation targets the implemented configuration, not the reputation or benchmark scores of a model.

Depends on: —
Feeds: SYN The synergy condition

INFThe informational environment

Șcheau and Angheluș, 2025

Personalised feeds, synthetic media, persuasive conversational systems and emotion-aware targeting can affect source discrimination and perceptual autonomy. In the framework, these effects are outcome and risk variables.

• Organisational controls require provenance, independent verification and monitoring of repeated narrative exposure.

Depends on: —
Feeds: SIG Observable signals

EPIThe status of claims

definitions · models · propositions · scenarios · controls

Definitions are stipulative and support consistent use of the notations. Analytical models describe operationalisable relationships, not universal laws. Propositions are testable or falsifiable. Scenarios are explanatory applications, not empirical findings. Controls are design requirements, not automatic claims of compliance.

• The framework is judged by clarity, discriminatory power, operationalisability and utility, not by exclusivity of a graphical form.

Depends on: —
Feeds: P1–P8 The framework’s propositions

NOTThe notations and their documented uses

Duda Guerra, 2025 · ASFOR, 2026 · Devonport · Esparza, 2026

The Human^AI notation was used publicly before, in contexts associated with amplifying human–AI collaboration; these uses are acknowledged as antecedents, without a contest over priority. No datable use of the exact Org^AI notation with the meaning of governed organisational capacity was located, an observation limited to the sources consulted.

• The ASCII forms HumanAI and OrgAI may appear in products or proper names; graphical identity does not imply conceptual identity.

Depends on: —
Feeds: —

HᵢThe person

Hᵢ — expertise, experience and state

The expertise, experience and state of person i, assessed relative to the domain and the task. The same model produces different profiles for different people.

Depends on: —
Feeds: —

MModel and AI configuration

M — model, tools, configuration

Includes version, parameters, sources and tools. At organisational level several systems appear, written Mⱼ.

• Not to be confused with the maturity levels M0–M5.

Depends on: —
Feeds: —

XTask, context and constraints

X — the domain of validity

Defines the domain in which claims about the configuration are valid. Without a delimited task there is no unit of analysis.

Depends on: DC Distributed cognition
Feeds: —

IᵢInterface and mode of interaction

Iᵢ — prompts, feedback, explanations, friction

The concrete way the person interacts with the system. Friction, or its absence, shapes trust calibration.

Depends on: ENG Intellectual augmentation
Feeds: —

VᵢValidation and trust calibration

Vᵢ — validation practice

The practice through which the person checks outputs, in proportion to risk. It is not reducible to formal human acceptance.

• It is the variable that links the individual configuration to epistemic decoupling.

Depends on: —
Feeds: Validation, Δepi Epistemic decoupling

DᵢDelegation and autonomy

Dᵢ — what AI may recommend, decide or execute

The rules that set the boundary between recommendation, decision and execution. The human role and decision authority must remain identifiable.

Depends on: —
Feeds: Authority

KᵢSources and knowledge available

Kᵢ — sources, provenance, use limitations

The sources the person can reach, with their provenance and use limitations. They are not yet authorised institutional knowledge (Korg).

Depends on: —
Feeds: Knowledge

tTime and period

t — the time index

Human^AI is episodic when it concerns a bounded task and stable when repeated use changes the person’s practice, trust, competence, autonomy or validation behaviour.

Depends on: —
Feeds: —

dₓ ≥ θₓThe material contribution condition

dₓ(YᵢAI, Yᵢ⁰) ≥ θₓ

Yᵢ⁰ is the reference outcome without AI contribution, Yᵢ^AI the outcome in the evaluated configuration, dₓ a task-appropriate difference function, and θₓ the declared materiality threshold. The difference may concern quality, time, errors, range of options, learning, autonomy or risk.

• The threshold is defined before evaluation, to prevent opportunistic classification.
• The contribution may be positive, neutral or negative: dependence or skill erosion do not remove a case from the category, they describe its outcome profile.

Depends on: —
Feeds: AI contribution, MET H Human^AI metrics, Yᵢ The individual outcome profile

YᵢThe individual outcome profile

Yᵢ := ⟨Qᵢ, τᵢ, Eᵢ, Cᵢ, Aᵢ, Lᵢ, Ωᵢ⟩
Yᵢ(X,t) = F(Hᵢ, M, X, Iᵢ, Vᵢ, Dᵢ, Kᵢ, t)

Quality, time, error, calibration, autonomy, learning and risk exposure. The profile is not compressed automatically into a score.

• Aᵢ here means autonomy and Lᵢ individual learning — not access A in Δepi and not L in Aorg.

Depends on: dₓ ≥ θₓ The material contribution condition
Feeds: Yorg The organisational outcome profile

UNITIdentifiable unit

H: person, system, task, context
O: process, actors, systems, knowledge

For Human^AI, the person, system, task and context must be identified; for Org^AI, the organisational process, actors, systems and knowledge.

Depends on: —
Feeds: Processes and control points

AIAI contribution

H: materially changes execution or judgement
O: integrated into institutional capacity

At individual level, the contribution materially changes execution or judgement. At organisational level, it is integrated into institutional capacity.

Depends on: dₓ ≥ θₓ The material contribution condition
Feeds: —

AUTHAuthority

H: human role and decision can be attributed
O: roles, approvals and escalations defined

Authority remains attributable. A system-only decision without identifiable authority and accountability does not satisfy the governed framework.

Depends on: Dᵢ Delegation and autonomy
Feeds: Roles, rights and accountability

KNOWKnowledge

H: relevant sources can be delimited
O: sources authorised, versioned and owned

Access to an internal corpus does not demonstrate provenance, validation or authority.

Depends on: Kᵢ Sources and knowledge available
Feeds: Korg Authorised institutional knowledge

VALValidation

H: the person can verify in proportion to risk
O: validation independent and institutionalised

Validation is competent, reproducible, risk-proportionate checking, distinct from formal approval.

Depends on: Vᵢ Validation and trust calibration
Feeds: Governance, validation, contestability, audit

TRACETraceability

H: the episode can be reconstructed
O: workflow and decision can be audited

Reconstruction is proportionate to risk: the graver the consequence, the more complete the trail must be.

Depends on: —
Feeds: T Configuration and decision traceability

LEARNLearning

H: effects on the person can be tracked
O: validated feedback updates memory in a controlled manner

At organisational level, only adjudicated feedback may update memory; raw feedback written directly produces contamination and self-confirmation.

Depends on: —
Feeds: L Validated organisational learning

E/SEpisodic or stable

episodic: bounded task
stable: practice changes through repetition

The same person may have isolated Human^AI episodes or a stable configuration in which repeated use changes their practice, trust, competence, autonomy or validation.

Depends on: —
Feeds: —

✕Excluded cases

5 cases that do not satisfy the constructs

They resemble Human^AI or Org^AI but do not pass the test.

• An ignored AI suggestion with no material effect does not constitute Human^AI in that episode.
• A deterministic automation without an AI component is not included merely because it is complex.
• Multiple independently used AI licences do not constitute Org^AI.
• A chatbot connected to unverified documents does not demonstrate authorised organisational knowledge.
• A system-only decision without identifiable authority and accountability does not satisfy the governed framework.

Depends on: —
Feeds: —

KorgAuthorised institutional knowledge

Korg — owner, version, validity, access rules

The source has an owner, provenance, version, validity and access rules. It corresponds to the knowledge criterion of the membership test.

Depends on: Organisation, authority and knowledge, Knowledge
Feeds: —

PProcesses and control points

P — integration into the real workflow

Connects AI to actual work. It corresponds to workflow integration in the membership test.

• Not to be confused with the propositions P1–P8.

Depends on: UNIT Identifiable unit
Feeds: —

RRoles, rights and accountability

R — decision-maker, validator, owner

Includes who decides, who validates and who owns the source or the process. It corresponds to authority and accountability.

Depends on: Authority
Feeds: Roles and separation of duties

GGovernance, validation, contestability, audit

G — rules, exceptions, oversight

Defines rules, exceptions and oversight. At the minimum threshold, feedback control is exercised through G.

Depends on: Validation
Feeds: —

SSecurity and data protection

S — identity, access, isolation, execution

Controls identity, access, isolation and execution.

Depends on: —
Feeds: THR The threat model

TConfiguration and decision traceability

T — reconstruction proportionate to risk

Supports reconstruction of the configuration and the decision.

• Not to be confused with the task set T in Δepi(T;X,t).

Depends on: Traceability
Feeds: AUD Evidence and audit trail

LValidated organisational learning

L — raw feedback ≠ accepted updates

The seventh component of the architecture. It separates raw feedback from accepted updates and is required for the cumulative form, not for the minimum threshold.

Depends on: Learning
Feeds: 5 Memory update through adjudication, M0–M5 The maturity model

IorgThe constitutive Org^AI threshold

Aorg := ⟨Korg, P, R, G, S, L, T⟩
Iorg = 1[Korg ∧ P ∧ R ∧ G ∧ S ∧ T]

The indicator expresses the minimum conceptual threshold. If any constitutive component is missing, the configuration may represent connected or aggregated AI use, but it does not demonstrate governed Org^AI.

• L does not enter the threshold: it is required for the cumulative form.

Depends on: —
Feeds: ≠ Σ Org^AI is not the sum of episodes, M0–M5 The maturity model

≠ ΣOrg^AI is not the sum of episodes

OrgAI(X,t) := ⟨{HumanAIᵢ}, {Mⱼ}, Aorg, X, t⟩
OrgAI ≠ Σ HumanAIᵢ

Org^AI is not a product, model, agent or document repository; it is a property of the organisational configuration. More users without integration do not satisfy the test.

• A licence count does not answer the question of which capacity has been created.

Depends on: Iorg The constitutive Org^AI threshold
Feeds: Yorg The organisational outcome profile

YorgThe organisational outcome profile

Yorg := ⟨Qorg, τorg, Kcov, Coord, Res, Div, Ωorg⟩
Yorg(X,t) = Φ({Yᵢ}, {Mⱼ}, Aorg, X, t)

Quality, time, knowledge coverage, coordination, resilience, diversity and systemic risk. Organisational outcomes depend on integration, coordination and control, not merely on individual outputs.

Depends on: Yᵢ The individual outcome profile, ≠ Σ Org^AI is not the sum of episodes
Feeds: —

SYNThe synergy condition

Q(H + OrgAI) > max{Q(H), Q(AI), Q(H + AI), Q(H + K), Q(Team)}

Synergy may be claimed only when performance exceeds relevant comparators on the same task and declared thresholds for traceability, autonomy, security and coordination cost are satisfied.

• Synergistic Org^AI is an outcome label, not an M6 level.

Depends on: FRT Performance, frontier and homogenisation
Feeds: P1–P8 The framework’s propositions

1Situated execution

actor + delimited HumanAI configuration

An actor performs a task within a delimited Human^AI configuration.

Depends on: —
Feeds: 2 Documented output

2Documented output

sources · configuration · interventions · confidence

The output is accompanied by sources, configuration, interventions and confidence level.

Depends on: 1 Situated execution
Feeds: 3 Validation by a competent role

3Validation by a competent role

accept / reject in proportion to risk

A competent role validates or rejects the contribution in proportion to risk.

Depends on: 2 Documented output
Feeds: Δepi Epistemic decoupling

4Attribution and integration

attributed decision → controlled process

The decision is attributed and integrated into a controlled organisational process.

Depends on: —
Feeds: 5 Memory update through adjudication

5Memory update through adjudication

separate adjudication workflow

Accepted conclusions may update memory only through a separate adjudication workflow.

Depends on: L Validated organisational learning, 4 Attribution and integration
Feeds: 6 Monitoring and withdrawal

6Monitoring and withdrawal

outcomes and errors monitored · access withdrawn

Outcomes and errors are monitored, and access or rules can be withdrawn or corrected.

Depends on: 5 Memory update through adjudication
Feeds: —

ΔepiEpistemic decoupling

Δepi(T;X,t) = A(T;X,t) − V(T;X,t)

The gap between the range of claims, interpretations or options that AI makes accessible and the demonstrated capacity of the accountable actor to validate them. Across a declared task set T, A is the proportion of tasks for which AI provides a usable response, and V the proportion for which the accountable actor can produce a verified justification.

• It is calculated within the same domain and ranges from −1 to 1.
• A high positive value shows that access exceeds validation.

Depends on: Vᵢ Validation and trust calibration, 3 Validation by a competent role
Feeds: Ωprop Propagation exposure

A ≠ auth.Access is not authority

generation · validation · authorisation

Access to a claim does not automatically produce justified knowledge, and technical validation does not automatically confer decision authority. The framework separates three questions: what the system can generate or retrieve, what the competent actor can validate, and who is authorised to decide or execute.

Depends on: —
Feeds: —

ΩpropPropagation exposure

Ωprop = max(0, Δepi) × ρ × κ

ρ is the rate at which an output is reused or diffused and κ the criticality of consequence. Without local calibration, the formula does not estimate real-world probabilities: it shows how a small error reused frequently in a critical process can become a major organisational risk.

• A conceptual ordering indicator, not a probability.

Depends on: Δepi Epistemic decoupling
Feeds: MET O Org^AI metrics

SIGObservable signals

6 signals of decoupling

Signs that access has outrun validation.

• Responses are accepted faster than their sources can be checked.
• Citations, explanations or calculations cannot be reproduced independently.
• The same output is reused across processes without reassessing context.
• Reported confidence rises while error-detection rates decline.
• Organisational memory is updated with unadjudicated outputs.
• Authority is formally assigned to a person, but practical constraints make contestation impossible.

Depends on: INF The informational environment
Feeds: Mitigation measures

MITMitigation measures

7 measures

Measures that raise validation without necessarily reducing access.

• Limit the domain and state unknowns.
• Cite and verify primary sources.
• Require independent validation for high-impact outputs.
• Separate generation from approval and execution.
• Test trust calibration and error detection.
• Control reuse and memory updates.
• Preserve rights to contest, stop and return to the unassisted process.

Depends on: SIG Observable signals
Feeds: PRINC Governance principles

P1–P8The framework’s propositions

3 conceptual · 5 empirical

The propositions make the framework open to criticism and verification. P2, P3 and P4 are conceptual commitments that follow from the definitions; P1, P5, P6, P7 and P8 are empirical and can be confirmed or refuted by observation.

• The full list is in the Propositions section below.

Depends on: EPI The status of claims, SYN The synergy condition
Feeds: —

M0–M5The maturity model

M0 → M1 → M2 → M3 → M4 → M5 (per domain)

Describes observable states of integration, not an obligation to reach the highest level. A level is assigned to a bounded domain and process, not to an organisation in the abstract.

• An organisation may operate at M1 in one domain, at M4 in another and prohibit AI in a third.
• For some tasks, managed Human^AI is more appropriate than cumulative Org^AI.

Depends on: Iorg The constitutive Org^AI threshold, L Validated organisational learning
Feeds: —

PRINCMeasurement principles

6 principles

Rules that precede any measurement.

• Define the unit, task, period and configuration before measurement.
• Select the comparator according to the real decision, not to favour the AI result.
• Report quality and risk together with time and coordination cost.
• Separate test data from data used for configuration or training.
• Retain negative outcomes, abstentions and not-applicable cases.
• Produce a profile and an uncertainty interval rather than false precision.

Depends on: —
Feeds: —

CALCThe computational companion

The framework calculator · versioned separately

The Human^AI and Org^AI calculator operationalises six relationships: epistemic decoupling, propagation exposure, the constitutive threshold, material contribution, synergy and maturity classification. It implements the formal rules without changing their meaning; the whitepaper remains the canonical source.

• An unknown state never counts as demonstrated.
• Exported values remain user-entered claims until linked to evidence.

Depends on: —
Feeds: —

MET HHuman^AI metrics

8 dimensions

Quality, time, error, detection, calibration, autonomy, learning and vulnerability, each with an indicator and evidence.

• The full table is in the Evaluation section below.

Depends on: dₓ ≥ θₓ The material contribution condition
Feeds: —

MET OOrg^AI metrics

9 dimensions

Knowledge, coordination, validation, traceability, security, resilience, diversity, learning and propagation.

Depends on: Ωprop Propagation exposure
Feeds: —

PROTMinimum evaluation protocol

8 steps

From delimiting the task to reporting residual risk.

• Define the task, population, period and consequence of error.
• Freeze the evaluated configuration: model, version, tools, sources, interface and rules.
• Set comparators, materiality threshold and success criteria before testing.
• Build a representative task set containing normal, boundary and adversarial cases.
• Measure quality, time, errors, detection, autonomy and coordination cost.
• Test permissions, data separation, prompt injection, unavailability and uncertain responses.
• Adjudicate results through competent roles and preserve disagreements.
• Report configuration, evidence, limitations, applicability interval and residual risk.

Depends on: —
Feeds: STAGES The implementation pathway

✕ SCOREWarning on scores

dimensional profile = the default form

A composite score can hide faster execution achieved by reducing validation, or greater consistency achieved by reducing diversity.

Depends on: —
Feeds: —

PRINCGovernance principles

6 principles

What must remain true throughout the workflow.

• Authority remains explicit and attributable.
• Validation is independent in proportion to impact and cannot be replaced by nominal approval.
• Contestability includes time, competence, evidence access and a real ability to stop or correct.
• Only authorised knowledge with ownership, provenance, validity and access rules is reusable.
• Configuration changes are versioned, tested, approved and reversible.
• Negative outcomes and disagreements remain visible for learning and audit.

Depends on: Mitigation measures
Feeds: —

ROLERoles and separation of duties

7 roles · incompatible shortcuts

Each role has a responsibility and a shortcut that cancels it.

• The full table is in the Governance section below.

Depends on: Roles, rights and accountability
Feeds: —

THRThe threat model

mechanism → minimum controls

Prompt injection, unauthorised disclosure, privilege confusion, hallucinated authority, memory contamination, automation bias and common-mode failure.

Depends on: Security and data protection
Feeds: —

AUDEvidence and audit trail

7 categories of evidence

What must be retained so the decision can be reconstructed.

• Identity, role and authorisation of the person and the service.
• Model, version, system instructions, tools and configuration state.
• Source identifiers, provenance, access decision and validity.
• Material prompts or interventions, without unnecessary sensitive content.
• Generated output, uncertainty, citations and validation status.
• Approval, decision, execution, exception and rollback events.
• Memory updates, adjudication, withdrawal and affected downstream uses.

Depends on: T Configuration and decision traceability
Feeds: —

STAGESThe implementation pathway

delimitation → … → monitoring

Eight stages, each with an exit criterion.

• The full table is in the Implementation section below.

Depends on: PROT Minimum evaluation protocol
Feeds: Acceptance criteria

ACCAcceptance criteria

9 acceptance criteria

The conditions to meet before an implementation is accepted.

• The configuration and domain are declared.
• Comparators and thresholds were established before testing.
• Sources are authorised, versioned and withdrawn in a controlled manner.
• Permissions were tested for unauthorised users and cross-domain access.
• Critical outputs have independent validation and an identifiable decision-maker.
• Actions are controlled, logged and reversible.
• Feedback does not update production memory directly.
• A safe fallback, an incident procedure and suspension criteria exist.
• The performance and risk profile reports negative outcomes rather than hiding them.

Depends on: STAGES The implementation pathway
Feeds: —

ANTIAnti-patterns and incorrect uses

8 anti-patterns · 6 incorrect uses

The confusions that make the framework look applied without being applied.

• The full table is in the Anti-patterns section below.

Depends on: —
Feeds: —

STOPStop conditions

stop or return to the previous state

Implementation must stop or return to the previous state on data exposure, loss of decision attribution, degradation in error detection, unauthorised access, memory contamination, inability to reproduce critical results, or an unvalidated configuration change.

Depends on: —
Feeds: —

✕ LAWThe regulatory boundary

analysis and control design ≠ legal compliance

The framework supports analysis and control design but does not establish legal compliance. Systems with material effects on rights, safety, employment, finance, health, security or public services require domain-specific legal assessment.

• Human^AI or Org^AI labels do not replace applicable obligations.

Depends on: —
Feeds: —