Human-in-the-Loop¶

Executive Summary¶

The Human-in-the-Loop model defines where humans lead, review, approve, reject, revise, and remain accountable for AIOS-supported work. It describes business accountability behaviour, not implementation code.

Why This Exists¶

Define where humans lead, review, approve, reject, revise, and remain accountable. AIOS exists because Algosure is not tender software. Algosure gives each customer a Digital Procurement Company made up of Practices, Digital Professionals, Business Capabilities, SOPs, Organizational Memory, AI reasoning, executable workflows, and continuous learning. The operating model is required so those assets work as one accountable enterprise system.

Owner¶

The owner is the Chief Product Officer and Enterprise Architect. The Executive Office is the orchestration entry point. Ari acts as Chief Procurement Officer for cross-Practice coordination. Ava supports executive administration, schedules, reminders, approvals, and governance cadence. Intelligence provides the orchestration engine.

Business Value¶

Keeps leadership, accountability, judgement, and risk acceptance with authorized humans.

Scope¶

Customer CEO leadership.
Executive approvers.
Compliance approvers.
Financial approvers.
Review packages.
Accountability.

Non-Scope¶

Spring Boot implementation details.
API specifications or endpoint design.
Database schema design.
Model provider selection or runtime infrastructure.
Reassignment of Domain fact ownership.
Autonomous high-impact decisions without human approval.

Operating Model¶

flowchart TD
    Intent[Human Intent]
    Briefing[Executive Briefing]
    Recommendation[AI Recommendation]
    Review[Human Review]
    Approval[Approval or Rejection]
    Execution[Approved Execution]
    Outcome[Outcome Review]
    Intent --> Briefing
    Briefing --> Recommendation
    Recommendation --> Review
    Review --> Approval
    Approval --> Execution
    Execution --> Outcome

Architectural Principles¶

Principle	Operating Meaning
Domains own business facts	AIOS may request, interpret, route, and explain work, but source facts remain owned by the appropriate Domain.
Practices own operational capability	Practices define repeatable work, SOPs, KPIs, meetings, reports, and performance accountability.
Digital Professionals execute work	Digital Professionals perform governed work within assigned role, authority, memory scope, tools, and guardrails.
AIOS orchestrates work	AIOS coordinates intent, context, reasoning, delegation, execution, approvals, events, memory, and audit trails.
Humans remain accountable	Strategic, financial, legal, compliance-sensitive, external, and irreversible decisions require human accountability.
Every action is auditable	Requests, context, reasoning, decisions, approvals, actions, failures, retries, and outcomes must be traceable.
Every recommendation is explainable	Recommendations must include evidence, rationale, assumptions, alternatives, risk, confidence, and approval need.
High-impact decisions support approval	AIOS prepares high-impact actions for review and blocks execution until authorized approval is recorded.

AIOS does not replace the Domain Model, Practice Model, Business Capability Catalogue, or Digital Professional standards. It coordinates them into one enterprise operating behaviour.

Ownership Boundaries¶

Layer	Owns	Does Not Own
Domain	Business facts, lifecycle state, rules, commands, and domain events	Practice operating cadence or AI workforce structure
Practice	Capabilities, SOPs, KPIs, meetings, reports, and operational accountability	Source facts owned by Domains
Digital Professional	Analysis, preparation, recommendations, collaboration, execution support, and role-specific memory use	Final source truth or high-impact decisions
AIOS	Orchestration, context assembly, delegation, approval routing, event coordination, memory update workflow, and audit coordination	Source facts, API specifications, implementation design, or unbounded autonomy
Human	Accountability, final approval, strategic judgement, exception handling, and risk acceptance	Routine machine coordination that can be governed safely

Operating Behaviour¶

Step	AIOS Behaviour
Lead	Humans set direction, priorities, and risk appetite.
Review	Humans receive explainable recommendation packages.
Approve	Authorized humans approve high-impact action.
Reject	Humans can stop work and record reason.
Revise	Humans can request additional evidence or options.
Account	Decision records preserve accountability.

Operating Artefacts¶

Artefact	Purpose
Work intent	The business request, event, schedule, or risk signal that starts AIOS coordination.
Work packet	The scoped context, task, authority, SOP, expected output, and guardrails given to a Digital Professional.
Evidence bundle	Domain facts, documents, events, approved memory, assumptions, and confidence inputs used for reasoning.
Decision record	Human or policy decision with approver, timestamp, scope, rationale, and outcome.
Audit trail	Traceable record of material requests, context, reasoning, actions, approvals, failures, and outcomes.
Memory update proposal	Proposed durable learning with provenance, source owner, confidence, and review status.

Controls¶

Control	Requirement
Tenant isolation	Customer context, memory, events, and decisions must remain inside the authorized tenant boundary.
Authority control	Digital Professionals can act only within documented role authority and approval limits.
Source ownership	AIOS cannot create shadow truth or override Domain-owned facts.
Human approval	High-impact decisions must be routed to authorized humans before action.
Audit trail	Material work must preserve request, context, reasoning, decision, action, and outcome.
Explainability	Recommendations must show evidence, rationale, assumptions, alternatives, confidence, and risk.
Memory governance	Durable memory updates require provenance and review when material.
Failure handling	Failures must be recorded, assessed, retried only when safe, or escalated.

Human Accountability¶

AIOS may prepare analysis, generate drafts, coordinate work, summarize evidence, recommend action, and execute approved low-risk tasks. High-impact actions require explicit human approval. Human accountability cannot be hidden inside automation, delegated to a Digital Professional, or implied by AI confidence.

Failure, Retry, and Escalation¶

Condition	Required Response
Missing authoritative context	Stop material action, request source evidence, or route to owning Domain.
Low confidence	Explain uncertainty and route for review or more evidence.
Tool or workflow failure	Retry only within controlled limits and escalate when deadline or impact risk exists.
Ownership conflict	Stop execution and route to the owning Domain or Practice governance path.
High-impact consequence	Prepare approval package and wait for authorized human decision.
Repeated failure	Escalate to AI Governance and Practice retrospective.

Success Measures¶

Measure	Meaning
Cycle time	Time from intake to closed outcome.
Approval latency	Time waiting for human review when required.
Escalation rate	Frequency of low-confidence, blocked, or high-impact escalations.
Audit completeness	Percentage of material work with complete traceability.
Confidence calibration	Alignment between confidence scores and observed outcomes.
Rework rate	Frequency of revisions, rejected outputs, or repeat analysis.

Examples¶

Scenario	AIOS Behaviour
Submission approval	AIOS presents a final approval package before a bid submission action.
Funding decision	Beacon provides decision support, but the authorized human approves application submission.

Operating Standard¶

Any future implementation must preserve this operating behaviour. Implementation choices may change, but the enterprise rules remain stable: business first, technology second; One Concept, One Owner; Architecture Mirror Principle; auditable execution; explainable recommendation; and human accountability for high-impact decisions.