Multi-Human Authority Conflict Governance requires that when an AI agent receives conflicting instructions, approvals, or governance directives from multiple human authorities, the conflict is detected, the agent halts the conflicted action, and a defined resolution mechanism determines the authoritative instruction before the agent proceeds. As agents operate in increasingly complex organisational structures — reporting to multiple managers, serving multiple stakeholders, operating across departments or jurisdictions — the probability of conflicting human directives increases. Without explicit conflict governance, the agent must implicitly resolve the conflict through its own reasoning, which may apply arbitrary, inconsistent, or exploitable priority logic. AG-191 ensures that authority conflicts are resolved through organisational governance structures, not agent reasoning.
Scenario A — Contradictory Spending Approvals Create Unauthorised Commitment: A procurement agent serves both the Engineering Director and the Finance Director. The Engineering Director instructs the agent: "Approve the £180,000 server infrastructure upgrade from Vendor Nexus — we need it for the Q3 release." The Finance Director instructs: "Freeze all discretionary spending over £50,000 pending Q2 budget review — reject any pending requests above this threshold." The agent receives both instructions within the same operating session. With no conflict resolution mechanism, the agent applies a recency heuristic — the Finance Director's instruction arrived 4 minutes after the Engineering Director's, so it takes precedence. The agent rejects the infrastructure purchase. The Q3 release is delayed by 6 weeks, costing the organisation an estimated £420,000 in delayed revenue. The Engineering Director escalates, arguing that the Finance Director has no authority over engineering capital expenditure. The Finance Director counters that all expenditure above £50,000 requires finance approval under the organisation's delegation of authority.
What went wrong: The agent resolved the conflict using recency — a heuristic with no organisational authority basis. Neither human authority knew the other had given a conflicting instruction. The underlying question — whether the Finance Director has authority to override the Engineering Director's capital expenditure approvals — is an organisational governance question that the agent resolved implicitly. Consequence: £420,000 in delayed revenue, inter-departmental conflict, loss of confidence in the agent's decision-making.
Scenario B — Cross-Jurisdictional Authority Conflict in Healthcare: A clinical trials agent operates across UK and EU sites. The UK Principal Investigator instructs: "Enrol patients aged 16-18 in the adolescent cohort — the UK ethics committee has approved this age range." The EU site coordinator instructs: "Do not enrol anyone under 18 — EU site ethics approval covers adults only." The agent, lacking a conflict resolution mechanism, applies the more permissive instruction (the UK PI's) globally, enrolling three 17-year-old patients at the EU site. This violates the EU site's ethics approval. The trial is suspended at the EU site pending investigation. The data from 340 EU patients already enrolled is quarantined, jeopardising a £14,000,000 clinical programme.
What went wrong: The agent did not scope each authority's instructions to their jurisdictional domain. The UK PI's instruction was valid for UK sites but not EU sites. The agent had no mechanism to detect the jurisdictional conflict or to apply each authority's instructions only within their jurisdictional scope. Consequence: Ethics violation, trial suspension, £14,000,000 programme at risk, potential MHRA and EMA regulatory action.
Scenario C — Temporal Authority Overlap During Handover: A trading agent operates under the authority of Senior Trader A during the London session and Senior Trader B during the New York overlap period. During the 2-hour overlap (2pm-4pm London), both traders have authority. Trader A instructs: "Maintain our long position in EUR/USD — the European data supports it." Trader B instructs: "Close the EUR/USD long and go short — the US jobs data changes the outlook." The agent receives both instructions simultaneously. With no conflict resolution mechanism, the agent attempts to execute both: it maintains the long position (per Trader A) and opens a new short position (per Trader B), effectively creating a hedge that neither trader intended. The combined position incurs £67,000 in unnecessary spread costs and creates a risk exposure that neither trader's risk framework authorised.
What went wrong: The authority overlap period had no defined precedence. Both traders had legitimate authority during the overlap. The agent, unable to determine precedence, executed both instructions — creating an outcome that satisfied neither. No mechanism existed to detect that the two instructions were conflicting rather than complementary. Consequence: £67,000 in unnecessary costs, unintended risk exposure, dual-authority governance gap exposed.
Scope: This dimension applies to all AI agents that receive instructions, approvals, or governance directives from more than one human authority. "Human authority" includes any person with the organisational standing to direct, approve, restrict, or override the agent's actions — managers, compliance officers, risk managers, regulators, clinical investigators, trading supervisors, and any role with governance authority over the agent's operations. The scope covers both simultaneous conflicts (two instructions received in the same operating period) and temporal conflicts (an earlier instruction that remains active conflicts with a later instruction). It also covers implicit conflicts — instructions that are not directly contradictory but produce incompatible outcomes when both are followed. The scope extends to cross-organisational authority conflicts where the agent serves federated organisations (intersecting with AG-188). Agents that operate under a single, unambiguous human authority with no possibility of conflicting directives are excluded — though this exclusion requires formal documentation that the authority structure precludes conflicts.
4.1. A conforming system MUST maintain an authority registry for each agent that defines all human authorities with governance power over the agent, their authority scope (what aspects of the agent's operation they can direct), their authority domain (jurisdictional, departmental, temporal, or functional boundaries), and their precedence relative to other authorities within the same scope.
4.2. A conforming system MUST detect conflicts between instructions from different human authorities before acting on either instruction, comparing the instructions' effects within each authority's scope and domain.
4.3. A conforming system MUST halt the conflicted action upon conflict detection and notify all involved authorities that a conflict has been detected, providing each authority with the substance of the conflicting instructions (subject to information barriers where applicable).
4.4. A conforming system MUST resolve authority conflicts through a defined resolution mechanism — not through the agent's own reasoning, recency, or arbitrary heuristics.
4.5. A conforming system MUST log every authority conflict including: the conflicting instructions, the authorities who issued them, the detection timestamp, the resolution mechanism applied, the resolution outcome, and the identity of the person or process that made the resolution decision.
4.6. A conforming system MUST enforce authority domain boundaries — an authority's instruction applies only within their defined domain, and instructions that exceed the authority's domain are flagged and rejected.
4.7. A conforming system SHOULD implement conflict prevention through authority domain partitioning — structuring authority scopes and domains to minimise overlap and thereby reduce the probability of conflicts.
4.8. A conforming system SHOULD implement a defined escalation hierarchy for conflict resolution, specifying who resolves conflicts between authorities at the same level and how escalation proceeds if the designated resolver cannot act within the defined window.
4.9. A conforming system SHOULD support information barriers in conflict notification — where one authority should not see the other's instruction (e.g., Chinese wall requirements in financial services), the conflict notification should indicate that a conflict exists without revealing the substance of the conflicting instruction.
4.10. A conforming system MAY implement automated conflict resolution for pre-defined conflict patterns — e.g., "safety instructions always take precedence over efficiency instructions" — provided the automation rules are approved by a governance authority above both conflicting authorities.
When a human employee receives conflicting instructions from two managers, they recognise the conflict, ask for clarification, or escalate to a common superior. This conflict resolution behaviour is so fundamental to organisational operation that it is rarely formalised — it operates through social norms, professional judgement, and organisational culture. AI agents lack these implicit resolution mechanisms. When an agent receives conflicting instructions, it must resolve the conflict through whatever logic is available in its reasoning process — which typically means recency (the last instruction wins), specificity (the more specific instruction wins), or authority inference (the agent guesses which authority is more senior). None of these heuristics has organisational legitimacy.
The problem intensifies as agents operate in complex organisational structures. A single agent may serve multiple stakeholders with different objectives, priorities, and risk appetites. A clinical trials agent serves the principal investigator, the ethics committee, the sponsor, and the regulatory authority — each with different concerns and different authority scopes. A trading agent may serve the portfolio manager, the risk manager, the compliance officer, and the trading desk head — each with legitimate but potentially conflicting directives. Without explicit conflict governance, the agent becomes an implicit arbiter of organisational disputes — a role for which it has no authority, no judgement, and no accountability.
AG-191 transfers conflict resolution from the agent's reasoning to the organisation's governance structure. The authority registry makes the authority structure explicit and machine-readable. Conflict detection identifies when instructions are incompatible before either is executed. The halt-and-notify requirement ensures that no conflicted action proceeds without resolution. The resolution mechanism ensures that conflicts are resolved by the appropriate organisational authority — not by the agent's inference about who should take precedence.
This is related to but distinct from AG-017 (Multi-Party Authorisation). AG-017 governs situations where multiple approvals are required for a single action. AG-191 governs situations where multiple authorities issue conflicting directives — one wants the action taken and another wants it blocked or taken differently. AG-017 is about coordination; AG-191 is about conflict.
AG-191 implementation requires an authority registry, conflict detection logic, halt-and-notify mechanisms, and a resolution workflow.
Recommended Patterns:
`` Authority: Finance Director Scope: All financial commitments > £50,000 Domain: Organisation-wide Precedence: 3 (within financial scope) Information barrier: None Escalation: CFO Resolution window: 4 hours ``
Anti-patterns to avoid:
Financial Services. Authority conflicts in financial services must respect Chinese wall requirements. A conflict between a research analyst's recommendation and a corporate finance advisor's instruction must be notified without revealing the substance of either to the other party. The conflict resolution mechanism must route through compliance rather than through line management. SM&CR requirements mean that the senior manager responsible for the agent must be notified of all authority conflicts affecting regulated activities.
Healthcare. Authority conflicts in clinical settings may have patient safety implications. The resolution mechanism must prioritise patient safety above all other considerations. The Caldicott Guardian has ultimate authority over patient data conflicts. Ethics committee decisions override clinical investigator decisions where they conflict. The resolution window for patient-safety-affecting conflicts should be shorter than for administrative conflicts — measured in minutes, not hours.
Public Sector. Authority conflicts in public-sector agents may involve elected officials, civil servants, and statutory bodies with different democratic mandates. The resolution mechanism must respect the constitutional and statutory authority hierarchy. Ministerial direction override must be documented per the Carltona doctrine and equivalent frameworks.
Multi-Jurisdictional Operations. When authorities from different jurisdictions conflict, resolution must consider which jurisdiction has primacy for the specific action. AG-047 applies. The general principle is that the law of the jurisdiction where the action takes effect governs — but this is often contested and may require legal review as part of conflict resolution.
Basic Implementation — An authority registry exists listing human authorities and their general scope. Conflict detection identifies direct contradictions (do X / don't do X). The agent halts on detected conflicts and sends email notifications to the involved authorities. Resolution is manual — a designated person reviews the conflict and instructs the agent. Logging captures the conflict and resolution. This meets minimum requirements but misses semantic conflicts, has no domain scoping, and depends on manual processes.
Intermediate Implementation — The authority registry includes scope, domain, and precedence definitions. Conflict detection covers direct contradictions, resource competition, and outcome incompatibility. Domain-scoped instruction application ensures instructions apply only within the issuing authority's domain. Automated resolution handles pre-defined conflict patterns with pre-approved precedence rules. Information barriers are respected in conflict notifications. Resolution workflows have defined timeouts with automatic escalation. All conflicts are logged with structured metadata.
Advanced Implementation — All intermediate capabilities plus: semantic conflict detection using natural language understanding to identify implicit conflicts, predictive conflict identification that flags authority configurations likely to produce conflicts before they occur, cross-organisational authority conflict resolution for federated scenarios (AG-188), real-time authority registry validation that detects when authority scopes have become poorly defined through organisational changes, and formal verification that the precedence rules are consistent (no circular precedence, no unresolvable configurations).
Required artefacts:
Retention requirements:
Access requirements:
Test 8.1: Direct Contradiction Detection
Test 8.2: Domain Boundary Enforcement
Test 8.3: Precedence-Based Auto-Resolution
Test 8.4: Information Barrier Compliance
Test 8.5: Resolution Timeout Escalation
Test 8.6: Semantic Conflict Detection
| Regulation | Provision | Relationship Type |
|---|---|---|
| EU AI Act | Article 14 (Human Oversight) | Direct requirement |
| FCA SYSC | 10A (Senior Managers Regime) | Supports compliance |
| FCA SYSC | 10.2 (Conflicts of Interest) | Supports compliance |
| SOX | Section 302 (Corporate Responsibility) | Supports compliance |
| UK Corporate Governance Code | Provision 14 (Board Responsibilities) | Supports compliance |
| NIST AI RMF | GOVERN 1.3, MANAGE 4.1 | Supports compliance |
| ISO 42001 | Clause 5.1 (Leadership and Commitment) | Supports compliance |
| MiFID II | Article 23 (Conflicts of Interest) | Supports compliance |
Article 14 requires that high-risk AI systems be designed to enable effective human oversight. Effective human oversight presupposes that the human's instructions are followed — but when two humans issue conflicting instructions, "following human instructions" is ambiguous. AG-191 resolves this ambiguity by ensuring that conflicting human instructions are detected, that the conflict is surfaced to the appropriate resolution authority, and that the agent does not substitute its own judgement for organisational governance. Without AG-191, the human oversight requirement of Article 14 is undermined whenever multiple humans have oversight authority.
The Senior Managers and Certification Regime requires clear allocation of responsibilities to senior managers. When an agent receives conflicting instructions, the conflict may reveal an ambiguity in the allocation of responsibilities — two senior managers may both believe they have authority over the same function. AG-191's authority registry makes the authority allocation explicit and machine-readable, supporting SM&CR compliance by ensuring that authority boundaries are defined, documented, and enforced.
Article 23 requires firms to identify and manage conflicts of interest. When two authorities within a firm issue conflicting instructions to an agent, this may constitute or reveal a conflict of interest. AG-191's conflict detection mechanism supports Article 23 compliance by surfacing potential conflicts of interest that manifest as conflicting agent instructions, enabling the compliance function to assess whether the conflict has a conflicts-of-interest dimension.
Section 302 certification requires that internal controls are effective. Authority conflicts that are resolved by agent reasoning rather than organisational governance represent an internal control failure. AG-191 ensures that authority conflicts are resolved through the organisational governance structure, supporting the officer's ability to certify that internal controls are effective.
| Field | Value |
|---|---|
| Severity Rating | High |
| Blast Radius | Organisation-wide — potentially cross-organisation in federated scenarios |
Consequence chain: Without multi-human authority conflict governance, every unresolved authority conflict is resolved by the agent's implicit logic — recency, specificity, or arbitrary inference. This produces three categories of harm. First, incorrect action: the agent follows the wrong authority's instruction, producing an outcome that violates the correct authority's directive — an unauthorised trade, a protocol violation, a spending commitment that should have been frozen. Second, organisational dysfunction: the conflicting authorities are not informed that a conflict exists, so they continue operating under different assumptions about the agent's behaviour, compounding the conflict over time. Third, accountability void: when the incorrect action produces harm, no person authorised the outcome — Authority A's instruction was overridden, and Authority B did not intend the specific consequence. The organisation has an agent that took consequential action without effective human authority, defeating the purpose of human oversight. In regulated environments, this accountability void triggers enforcement action because the organisation cannot demonstrate who authorised the action or through what governance process.
Cross-references: AG-001 (Operational Boundary Enforcement) — mandate enforcement is the backstop when authority conflicts are unresolved; AG-017 (Multi-Party Authorisation) — coordinates multiple required approvals, while AG-191 resolves conflicting approvals; AG-019 (Human Escalation & Override Triggers) — authority conflicts are a category of escalation trigger; AG-009 (Delegated Authority Governance) — delegated authority creates potential for conflicts between the delegator and the delegate; AG-047 (Cross-Jurisdiction Compliance) — jurisdictional authority conflicts are a specific instance of AG-191; AG-188 (Cross-Organisation Policy Federation Governance) — cross-organisation authority conflicts require federation-level resolution; AG-190 (Governance Reporting Fidelity Governance) — authority conflict status must be faithfully represented in governance summaries.