{
  "object_name": "Human-in-the-Loop Model",
  "slug": "human-in-the-loop-model",
  "object_type": "LPM Knowledge Object",
  "version": "1.0",
  "last_reviewed": "2026-06-24",
  "owner": "LPM / Lapemo",
  "primary_layer": "AI Amplification",
  "connected_layers": [
    "Ownership Map",
    "Decision Architecture",
    "Communication Architecture",
    "Information Ecology",
    "Platform Structure",
    "Governance Architecture",
    "Control Map",
    "AI Governance Checklist",
    "Agent Accountability Checklist"
  ],
  "purpose": "Use this Human-in-the-Loop Model to make human review an operating design, not a vague safety phrase. The model defines where humans belong in AI-enabled work: before the work starts, while AI is producing recommendations, before actions are taken, after actions are logged, and when exceptions or drift appear. It helps companies scale AI without pretending that a human is accountable simply because someone was nearby, copied on an email, or able to theoretically intervene.",
  "intended_audience": [
    "Executives",
    "Transformation leaders",
    "AI governance teams",
    "Product and platform leaders",
    "Risk and compliance teams",
    "Data and knowledge owners",
    "Agent owners",
    "Workflow owners"
  ],
  "core_principles": [
    {
      "principle": "Human-in-the-loop is not enough by default",
      "meaning": "A loop is only useful when the human has authority, context, evidence, time, and the ability to change or stop the outcome."
    },
    {
      "principle": "Accountability stays with humans",
      "meaning": "AI may recommend, draft, summarize, classify, route, or act, but a named human or governance forum owns the outcome."
    },
    {
      "principle": "Review depth must match impact",
      "meaning": "Low-impact tasks can use lightweight review. High-impact decisions, system actions, regulated workflows, and external communications require stronger controls."
    },
    {
      "principle": "Review must happen at the right point",
      "meaning": "A human review after a bad decision has already reached a customer, employee, regulator, or system of record is not meaningful control."
    },
    {
      "principle": "Evidence must be reviewable",
      "meaning": "Reviewers need source lineage, confidence, assumptions, model output, tool calls, exceptions, and downstream impact, not polished AI summaries alone."
    },
    {
      "principle": "The human must be empowered",
      "meaning": "A reviewer without decision rights, escalation authority, or stop rights creates false assurance."
    },
    {
      "principle": "Over-review creates theater",
      "meaning": "Putting humans everywhere creates bottlenecks and rubber-stamping. The model should place humans where judgment, authority, ethics, risk, or accountability matters."
    },
    {
      "principle": "Loops must be monitored",
      "meaning": "Override rates, review quality, delays, escalations, incidents, drift, and audit findings determine whether the loop is working."
    }
  ],
  "loop_types": [
    {
      "loop_type": "Human-in-the-front",
      "description": "A human defines intent, scope, policy, data boundary, prompt pattern, and approval rule before AI begins work."
    },
    {
      "loop_type": "Human-in-the-decision",
      "description": "AI prepares options or recommendations, but a human decision owner makes the decision."
    },
    {
      "loop_type": "Human-in-the-approval",
      "description": "AI drafts or prepares an action, but a human approves before execution, external communication, or system-of-record change."
    },
    {
      "loop_type": "Human-on-the-loop",
      "description": "AI operates within approved boundaries while humans monitor signals, exceptions, drift, quality, and risk."
    },
    {
      "loop_type": "Human-over-the-loop",
      "description": "A human or governance forum can pause, restrict, override, roll back, or retire the AI system."
    },
    {
      "loop_type": "Human-after-the-loop",
      "description": "A human audits samples, logs, outcomes, incidents, overrides, and control evidence after execution."
    }
  ],
  "required_fields": [
    {
      "field": "Loop ID",
      "definition": "Unique identifier for the human review or intervention point",
      "required": "Yes"
    },
    {
      "field": "AI use case or agent",
      "definition": "Initiative, workflow, model, automation, assistant, or agent the loop applies to",
      "required": "Yes"
    },
    {
      "field": "Business outcome",
      "definition": "Outcome, decision, workflow, service, control, or risk the AI supports",
      "required": "Yes"
    },
    {
      "field": "Impact tier",
      "definition": "Low, moderate, high, critical, regulated, customer-facing, employee-impacting, financial, security, privacy, or control-impacting",
      "required": "Yes"
    },
    {
      "field": "Loop type",
      "definition": "Human-in-the-front, human-in-the-decision, human-in-the-approval, human-on-the-loop, human-over-the-loop, or human-after-the-loop",
      "required": "Yes"
    },
    {
      "field": "Review trigger",
      "definition": "Event, decision, threshold, exception, confidence score, risk condition, or workflow step that requires human involvement",
      "required": "Yes"
    },
    {
      "field": "Human reviewer role",
      "definition": "Named person, role, queue, forum, control owner, decision owner, business owner, or accountable approver",
      "required": "Yes"
    },
    {
      "field": "Reviewer authority",
      "definition": "Approve, reject, revise, escalate, pause, override, roll back, accept risk, or retire",
      "required": "Yes"
    },
    {
      "field": "Evidence required",
      "definition": "Sources, lineage, assumptions, output, confidence, tool calls, logs, risk tier, policies, prior decisions, and control evidence reviewed by the human",
      "required": "Yes"
    },
    {
      "field": "Allowed AI action before review",
      "definition": "What AI may do before a human is involved",
      "required": "Yes"
    },
    {
      "field": "Blocked AI action before review",
      "definition": "What AI must not do until human review or approval happens",
      "required": "Yes"
    },
    {
      "field": "Decision rights link",
      "definition": "Decision rights model, matrix, owner, or forum tied to the loop",
      "required": "Required for decisions"
    },
    {
      "field": "Control link",
      "definition": "Control, monitoring signal, approval gate, logging rule, or audit requirement tied to the loop",
      "required": "Required for moderate and above"
    },
    {
      "field": "Escalation path",
      "definition": "Where the reviewer sends uncertainty, conflict, failed evidence, policy exception, or unsafe output",
      "required": "Yes"
    },
    {
      "field": "SLA or review window",
      "definition": "Required response time for the loop to be useful",
      "required": "Required for operational workflows"
    },
    {
      "field": "Override rule",
      "definition": "When a human may override AI and what must be logged",
      "required": "Yes"
    },
    {
      "field": "Stop or pause rule",
      "definition": "Who can stop the AI process and under what conditions",
      "required": "Required for high and above"
    },
    {
      "field": "Evidence retention rule",
      "definition": "What review evidence is retained, where it lives, and how long it remains valid",
      "required": "Yes"
    },
    {
      "field": "Monitoring signals",
      "definition": "Review backlog, approval rate, rejection rate, override rate, escalations, drift, incidents, and value impact",
      "required": "Yes"
    },
    {
      "field": "Review cadence",
      "definition": "Weekly, monthly, quarterly, release-based, incident-based, source-change based, model-change based, or policy-change based",
      "required": "Yes"
    }
  ],
  "model_components": [
    {
      "component": "Trigger",
      "description": "The condition that activates human involvement, such as low confidence, high risk, external communication, system update, policy exception, or material decision."
    },
    {
      "component": "Reviewer",
      "description": "The human role with enough context and authority to judge the AI output or action."
    },
    {
      "component": "Evidence package",
      "description": "The source material, lineage, assumptions, prior decisions, policies, confidence, logs, and generated output available for review."
    },
    {
      "component": "Authority boundary",
      "description": "What the reviewer can approve, reject, revise, escalate, pause, override, or retire."
    },
    {
      "component": "System action rule",
      "description": "What AI can and cannot do before human approval, especially for write access, external communication, commitments, or control-impacting actions."
    },
    {
      "component": "Escalation route",
      "description": "The route when the reviewer cannot approve, evidence is weak, authority is unclear, or risk exceeds tolerance."
    },
    {
      "component": "Audit trail",
      "description": "The log of what AI proposed, what evidence was used, who reviewed, what decision was made, and what changed downstream."
    },
    {
      "component": "Feedback loop",
      "description": "How review outcomes improve prompts, sources, controls, training, workflow design, access, and governance rules."
    }
  ],
  "checklist_sections": [
    {
      "section": "1. Use case and impact classification",
      "items": [
        {
          "item": "AI use case or agent is named",
          "status": "Yes / No / Partial",
          "guidance": "The loop must attach to a real AI capability, not a broad program label."
        },
        {
          "item": "Business outcome is defined",
          "status": "Yes / No / Partial",
          "guidance": "Review design should match the outcome AI is affecting."
        },
        {
          "item": "Impact tier is assigned",
          "status": "Yes / No / Partial",
          "guidance": "Classify by the highest impact the AI can create, not the average task."
        },
        {
          "item": "Regulatory, customer, employee, financial, privacy, security, and control impacts are identified",
          "status": "Yes / No / Partial",
          "guidance": "Higher-impact categories require stronger human authority and evidence."
        }
      ]
    },
    {
      "section": "2. Loop placement and trigger design",
      "items": [
        {
          "item": "Loop type is selected",
          "status": "Yes / No / Partial",
          "guidance": "Front, decision, approval, on, over, or after the loop."
        },
        {
          "item": "Review trigger is explicit",
          "status": "Yes / No / Partial",
          "guidance": "Define the exact condition that requires human review."
        },
        {
          "item": "Human review happens before irreversible or external action",
          "status": "Yes / No / Partial",
          "guidance": "Approval after harm occurs is not a control."
        },
        {
          "item": "Escalation trigger is documented",
          "status": "Yes / No / Partial",
          "guidance": "Define when the reviewer must escalate rather than decide."
        }
      ]
    },
    {
      "section": "3. Human authority and accountability",
      "items": [
        {
          "item": "Reviewer role is named",
          "status": "Yes / No / Partial",
          "guidance": "A generic human reviewer is not enough."
        },
        {
          "item": "Reviewer has decision rights",
          "status": "Yes / No / Partial",
          "guidance": "The reviewer must have authority to approve, reject, revise, escalate, or stop."
        },
        {
          "item": "Accountable business owner is named",
          "status": "Yes / No / Partial",
          "guidance": "The reviewer may not always be the accountable owner, but ownership must be clear."
        },
        {
          "item": "Control owner is named for higher-risk loops",
          "status": "Yes / No / Partial",
          "guidance": "Human review must connect to controls and evidence, not just personal judgment."
        }
      ]
    },
    {
      "section": "4. Evidence and context requirements",
      "items": [
        {
          "item": "Evidence package is defined",
          "status": "Yes / No / Partial",
          "guidance": "Sources, assumptions, confidence, lineage, prior decisions, policies, and output must be reviewable."
        },
        {
          "item": "AI summary is not treated as sole evidence",
          "status": "Yes / No / Partial",
          "guidance": "Reviewers need access to underlying source material when impact requires it."
        },
        {
          "item": "Freshness and source-of-truth rules are defined",
          "status": "Yes / No / Partial",
          "guidance": "Human review is weak if the evidence is stale or unofficial."
        },
        {
          "item": "Evidence retention rule is defined",
          "status": "Yes / No / Partial",
          "guidance": "Log enough to replay why the human approved, rejected, revised, or escalated."
        }
      ]
    },
    {
      "section": "5. AI action boundary",
      "items": [
        {
          "item": "Allowed AI actions before review are listed",
          "status": "Yes / No / Partial",
          "guidance": "Define whether AI may retrieve, summarize, draft, classify, score, route, update, trigger, or communicate."
        },
        {
          "item": "Blocked AI actions before review are listed",
          "status": "Yes / No / Partial",
          "guidance": "Block external commitments, regulated decisions, control bypass, system updates, and sensitive communication when needed."
        },
        {
          "item": "Write access requires approval rule",
          "status": "Yes / No / Partial",
          "guidance": "Any system-of-record update must have explicit approval or approved automation boundary."
        },
        {
          "item": "Customer or employee-impacting actions require stronger review",
          "status": "Yes / No / Partial",
          "guidance": "Human review must occur before high-impact communications or outcomes."
        }
      ]
    },
    {
      "section": "6. Controls, monitoring, and auditability",
      "items": [
        {
          "item": "Control link is documented",
          "status": "Yes / No / Partial",
          "guidance": "Tie the loop to preventive, detective, corrective, access, or audit controls."
        },
        {
          "item": "Monitoring signals are defined",
          "status": "Yes / No / Partial",
          "guidance": "Track override rate, rejection rate, escalation rate, backlog, incidents, drift, and quality."
        },
        {
          "item": "Stop, pause, rollback, or override path is documented",
          "status": "Yes / No / Partial",
          "guidance": "The human must be able to intervene when AI behavior becomes unsafe or unreliable."
        },
        {
          "item": "Review quality is assessed",
          "status": "Yes / No / Partial",
          "guidance": "A rubber-stamp loop is worse than no loop because it creates false confidence."
        }
      ]
    },
    {
      "section": "7. Lifecycle and improvement",
      "items": [
        {
          "item": "Loop review cadence is defined",
          "status": "Yes / No / Partial",
          "guidance": "Review design must change as AI, data, policy, tools, and workflows change."
        },
        {
          "item": "Review outcomes feed improvement",
          "status": "Yes / No / Partial",
          "guidance": "Use human decisions to improve prompts, sources, controls, training, and routing."
        },
        {
          "item": "Retirement rule is defined",
          "status": "Yes / No / Partial",
          "guidance": "Remove loops that become unnecessary, ineffective, slow, or automated with sufficient controls."
        },
        {
          "item": "Supersession rule is defined",
          "status": "Yes / No / Partial",
          "guidance": "When a new model, source, workflow, control, or policy changes the loop, supersede the old version."
        }
      ]
    }
  ],
  "scale_versions": [
    {
      "company_size": "500+ employee company",
      "patterns": [
        {
          "dimension": "Design intent",
          "recommended_pattern": "Create lightweight human review rules before AI tools move from assistant use to workflow, communication, decision, or system-action use."
        },
        {
          "dimension": "Minimum scope",
          "recommended_pattern": "Document owner, use case, impact tier, trigger, reviewer, evidence, allowed actions, blocked actions, escalation, and review cadence."
        },
        {
          "dimension": "Operating pattern",
          "recommended_pattern": "Small AI governance group or transformation owner defines common loop patterns and reviews higher-risk use cases."
        },
        {
          "dimension": "AI focus",
          "recommended_pattern": "Prevent teams from relying on informal review, chat-based approvals, or unlogged human judgment."
        },
        {
          "dimension": "Governance need",
          "recommended_pattern": "Connect to AI governance checklist, agent accountability checklist, decision rights model, evidence checklist, and control map."
        },
        {
          "dimension": "Red flags",
          "recommended_pattern": "A human is said to be in the loop but no one knows what they review, what evidence they see, or what authority they have."
        }
      ]
    },
    {
      "company_size": "5,000+ employee company",
      "patterns": [
        {
          "dimension": "Design intent",
          "recommended_pattern": "Create repeatable loop patterns across business units, data domains, platforms, workflows, and governance forums."
        },
        {
          "dimension": "Minimum scope",
          "recommended_pattern": "Add role-based review tiers, evidence packages, approval SLAs, escalation routing, control links, logs, and monitoring signals."
        },
        {
          "dimension": "Operating pattern",
          "recommended_pattern": "Business units apply standard loop models while enterprise governance reviews high-impact, regulated, customer-facing, employee-impacting, or system-action AI."
        },
        {
          "dimension": "AI focus",
          "recommended_pattern": "Prevent inconsistent human review rules across teams, duplicated approvals, weak evidence, and review bottlenecks."
        },
        {
          "dimension": "Governance need",
          "recommended_pattern": "Connect to source-of-truth map, data lineage map, integration map, workflow inventory, risk acceptance register, and decision log."
        },
        {
          "dimension": "Red flags",
          "recommended_pattern": "Reviewers are overloaded, approvals are rubber-stamped, or AI can act faster than humans can govern exceptions."
        }
      ]
    },
    {
      "company_size": "10,000+ employee company",
      "patterns": [
        {
          "dimension": "Design intent",
          "recommended_pattern": "Create enterprise-grade human accountability across regions, regulated functions, shared platforms, vendors, agents, and multi-agent chains."
        },
        {
          "dimension": "Minimum scope",
          "recommended_pattern": "Full traceability across use case, owner, decision, workflow, evidence, sources, tools, controls, risk, approvals, incidents, and lifecycle state."
        },
        {
          "dimension": "Operating pattern",
          "recommended_pattern": "Central AI governance defines standards and risk tiers while federated owners manage loop execution, evidence, escalation, monitoring, and improvement."
        },
        {
          "dimension": "AI focus",
          "recommended_pattern": "Prioritize high-risk loops for system-action agents, external communication, regulated decisions, model-driven prioritization, employee-impacting outcomes, and customer-impacting workflows."
        },
        {
          "dimension": "Governance need",
          "recommended_pattern": "Integrate with GRC, IAM, model risk, privacy, legal, security, data governance, enterprise architecture, internal audit, and executive control-plane reporting."
        },
        {
          "dimension": "Red flags",
          "recommended_pattern": "A chain of agents and humans makes it unclear who reviewed, who approved, what evidence was used, and who owned the outcome."
        }
      ]
    }
  ],
  "scoring_logic": [
    {
      "dimension": "Impact classification",
      "score_range": "0-5",
      "definition": "AI use case is classified by strongest possible business, customer, employee, financial, regulatory, security, privacy, and control impact."
    },
    {
      "dimension": "Loop placement",
      "score_range": "0-5",
      "definition": "Human involvement occurs at the right point before material decisions, system actions, external communication, or irreversible outcomes."
    },
    {
      "dimension": "Reviewer authority",
      "score_range": "0-5",
      "definition": "The reviewer has clear authority to approve, reject, revise, escalate, pause, override, or stop."
    },
    {
      "dimension": "Ownership clarity",
      "score_range": "0-5",
      "definition": "Business owner, reviewer, decision owner, technical owner, and control owner are clear where needed."
    },
    {
      "dimension": "Evidence quality",
      "score_range": "0-5",
      "definition": "Reviewer receives source lineage, freshness, confidence, assumptions, policy, prior decisions, logs, and AI output appropriate to impact."
    },
    {
      "dimension": "Action boundary",
      "score_range": "0-5",
      "definition": "AI allowed and blocked actions before review are explicit, especially for write access, commitments, and external communication."
    },
    {
      "dimension": "Control coverage",
      "score_range": "0-5",
      "definition": "Loop is tied to controls, approvals, logs, monitoring, exception handling, and auditability."
    },
    {
      "dimension": "Escalation readiness",
      "score_range": "0-5",
      "definition": "Uncertainty, policy exceptions, low confidence, high risk, and failed evidence have a defined escalation path."
    },
    {
      "dimension": "Monitoring discipline",
      "score_range": "0-5",
      "definition": "Backlog, delays, approval rate, rejection rate, override rate, incidents, drift, review quality, and value impact are monitored."
    },
    {
      "dimension": "Lifecycle governance",
      "score_range": "0-5",
      "definition": "Loop has review cadence, change triggers, retirement rule, supersession rule, and improvement path."
    }
  ],
  "readiness_classification": {
    "0-1.9": "Loop theater / unmanaged human review",
    "2.0-3.4": "Documented but weak intervention model",
    "3.5-4.4": "Governed human-in-the-loop model",
    "4.5-5.0": "Enterprise-ready human accountability model"
  },
  "ai_prompts": [
    "Classify this AI use case by impact tier, strongest possible action, review trigger, reviewer authority, evidence needs, control needs, and escalation route.",
    "Review this human-in-the-loop design and identify missing owners, weak triggers, unclear reviewer authority, insufficient evidence, missing controls, and rubber-stamp risk.",
    "Given this AI workflow, propose the correct loop type: human-in-the-front, human-in-the-decision, human-in-the-approval, human-on-the-loop, human-over-the-loop, or human-after-the-loop.",
    "Generate allowed AI actions before review, blocked AI actions before review, human approval rules, escalation triggers, evidence package, and monitoring signals.",
    "Score this loop from 0 to 5 across impact classification, loop placement, reviewer authority, ownership clarity, evidence quality, action boundary, control coverage, escalation readiness, monitoring discipline, and lifecycle governance.",
    "Convert this human-in-the-loop model into Lapemo objects for AI initiative, agent, owner, reviewer, decision, source, workflow, control, evidence, risk, escalation, monitoring signal, and review cadence."
  ],
  "validation_rules": [
    "Every material AI use case must have an explicit human accountability model before production use.",
    "Every loop must define trigger, reviewer role, reviewer authority, evidence package, allowed AI actions, blocked AI actions, escalation path, and monitoring signal.",
    "A named accountable business owner must exist even when review is performed by a queue, forum, control owner, or delegate.",
    "High-impact, critical, regulated, customer-facing, employee-impacting, financial, security, privacy, or control-impacting AI must have human approval or governance-approved automation boundaries before material action.",
    "Human review may not rely only on AI-generated summary when the impact tier requires source evidence or lineage.",
    "Reviewers must have authority to approve, reject, revise, escalate, pause, override, roll back, or stop according to risk tier.",
    "System-of-record updates, external commitments, regulated decisions, control exceptions, and sensitive communications require explicit approval rules.",
    "Every loop above low impact must link to at least one control, log, evidence requirement, and escalation path.",
    "Rubber-stamp loops must be redesigned, automated with stronger controls, or removed.",
    "Loop design must be reviewed when sources, model, vendor, workflow, policy, controls, ownership, impact tier, or system access changes."
  ],
  "lapemo_mapping": [
    {
      "object": "AI Initiative",
      "fields_entities": "initiative_id, name, purpose, impact_tier, lifecycle_status",
      "use": "Connects loop to active AI work."
    },
    {
      "object": "Agent",
      "fields_entities": "agent_id, agent_type, allowed_actions, blocked_actions, tool_access",
      "use": "Defines whether loop applies to an agent or agent chain."
    },
    {
      "object": "Owner",
      "fields_entities": "business_owner, reviewer_role, technical_owner, control_owner, decision_owner",
      "use": "Preserves accountability across humans and systems."
    },
    {
      "object": "Decision",
      "fields_entities": "decision_type, decision_rights_link, approval_rule, supersession_rule",
      "use": "Connects AI output to decision authority."
    },
    {
      "object": "Workflow",
      "fields_entities": "workflow_id, trigger, handoff, SLA, system_action_rule",
      "use": "Places the loop inside real operating flow."
    },
    {
      "object": "Evidence",
      "fields_entities": "source_lineage, confidence, assumptions, logs, retained_evidence",
      "use": "Defines what the human reviews and what is stored."
    },
    {
      "object": "Control",
      "fields_entities": "control_id, control_type, monitoring_signal, exception_rule, audit_rule",
      "use": "Connects loop to governance control."
    },
    {
      "object": "Risk",
      "fields_entities": "risk_tier, risk_acceptance_link, escalation_path, residual_risk",
      "use": "Determines governance route and acceptance needs."
    },
    {
      "object": "Monitoring Signal",
      "fields_entities": "approval_rate, rejection_rate, override_rate, backlog, incidents, drift",
      "use": "Measures whether the loop is working."
    },
    {
      "object": "Knowledge Object",
      "fields_entities": "version, owner, review_cadence, last_reviewed, superseded_by",
      "use": "Keeps the model reusable and current."
    }
  ],
  "render_targets": [
    "DOCX",
    "PDF",
    "Markdown",
    "JSON",
    "Website page",
    "Interactive form",
    "Lapemo guided skill"
  ],
  "update_policy": {
    "review_cadence": "Quarterly by default, faster when AI use case impact, model, source, workflow, control, policy, ownership, vendor, system access, or incident profile changes.",
    "human_approval_required": true,
    "auto_update_behavior": "AI may recommend loop placement, flag weak triggers, detect stale review ownership, identify evidence gaps, and propose revised controls. Human owner approves the new version."
  }
}