AIFC-020: Human-Managed AI

Status: Draft 0.1 Standard: AI-First Community Standard Abbreviation: AIFC Builds on:

• AIFC-000 Manifesto of an AI-first community
• AIFC-001 Core Concepts
• AIFC-002 Community Model
• AIFC-003 Values and Purpose
• AIFC-004 Feedback and Change Proposals
• AIFC-010 Knowledge Structure
• AIFC-011 Operational DNA
• AIFC-012 Metadata and Markdown
• AIFC-013 Human and AI Readable Content

Purpose of this document: Define the principle of Human-Managed AI: how AI can accelerate, support, and extend a community’s capabilities without taking ownership of its purpose, values, accountability, critical decisions, or operational capability.

1. Purpose of this document

This document defines the basic rules for using AI in an AIFC community.

AIFC is not an anti-AI standard.

On the contrary, it assumes that AI can help communities significantly:

• understand their own know-how,
• synthesize information,
• detect risks,
• propose changes,
• accelerate work,
• create skills,
• clean the knowledge base,
• support decision-making,
• improve workflows,
• find contradictions,
• protect against degradation,
• and help the community learn.

At the same time, AIFC rejects a model in which a community gradually becomes so dependent on AI that, without it, it loses the ability to decide, work, understand its own system, or continue its purpose.

This document therefore answers the following questions:

• What does Human-Managed AI mean?
• What is the difference between AI-first and AI-dependent?
• What may AI own and what must it not own?
• Who owns purpose, values, and decisions?
• How are AI-generated proposals handled?
• When is human approval required?
• How is AI autonomy governed?
• How is AI dependency prevented?
• What is the role of the Human Cockpit Layer?
• How does the community remain capable of operating without AI?

2. Core principle

The core principle of this document is:

AI may accelerate the community.
AI must not own the community.

An AIFC community may be AI-first.

It must not be AI-dependent.

AI may:

• analyze,
• propose,
• synthesize,
• generate,
• flag issues,
• validate,
• compare,
• prepare supporting material,
• execute approved low-risk actions,
• help maintain the knowledge base,
• and support decision-making.

AI must not own, without accountable governance:

• purpose,
• values,
• critical decisions,
• accountability,
• operational capability,
• confidentiality of know-how,
• or the direction of the community.

AI may propose paths. The community holds the direction.

3. AI-first vs AI-dependent

AIFC distinguishes two fundamentally different states.

AI-first community

An AI-first community has its knowledge, values, work, decision-making, and interfaces structured so that AI can safely read, improve, and accelerate them.

An AI-first community:

• has a source of truth,
• has values and purpose,
• has human ownership of accountability,
• has AI governance,
• has an AI-NDA Boundary,
• has fallback for critical workflows,
• returns AI-generated know-how to the source of truth,
• monitors AI dependency,
• maintains a Human Capability Reserve,
• uses AI as an accelerator.

AI-dependent community

An AI-dependent community loses the ability to operate without AI.

An AI-dependent community:

• does not know where AI is being used,
• has no fallback,
• has no exit strategy,
• leaves know-how inside AI tools,
• loses human skills,
• cannot validate AI outputs,
• automates critical decisions without governance,
• transfers operational capability into an external tool,
• stops when AI, tokens, a model, or a vendor becomes unavailable.

Minimum requirement

An AIFC community must regularly distinguish whether its use of AI represents:

AI acceleration

or

AI dependency

AI acceleration is desirable. AI dependency must be managed as a risk.

4. AI as accelerator, not owner

AI has the role of an accelerator in an AIFC community.

This means that it helps the community more quickly or more effectively:

• understand,
• formulate,
• propose,
• decide with supporting material,
• work,
• learn,
• maintain knowledge,
• reveal contradictions,
• create alternatives,
• improve the system.

AI, however, is not the owner.

The owner is a human, role, team, governance body, or community.

Example

AI may propose a workflow change.

AI must not decide by itself that the new workflow is approved if the change is significant.

Correct model:

AI detects issue
-> AI drafts change proposal
-> human/process owner reviews
-> decision is recorded
-> approved change updates source of truth

Incorrect model:

AI detects issue
-> AI updates active workflow
-> no review
-> no decision record
-> no owner

Minimum requirement

An AI-generated proposal for a significant change must be marked as a proposal and must pass through accountable governance.

5. Human ownership of purpose

The purpose of the community must be owned by a human or by the community.

AI may help:

• summarize an existing discussion,
• name recurring themes,
• compare formulations,
• identify contradictions,
• propose questions,
• formulate variants,
• translate purpose into strategy.

AI must not be the final owner of purpose.

Required distinction

AIFC must distinguish:

AI-generated purpose proposal

and

community-approved purpose

An AI-generated purpose proposal is a proposal. Community-approved purpose is a decision.

Minimum requirement

An approved purpose must be traceably accepted by a human or by community governance.

6. Human ownership of values

Values are the highest governance layer of the community.

AI may help values be:

• formulated,
• explained,
• compared,
• tested on examples,
• checked for contradictions between values and practice,
• updated in their interpretation.

AI must not decide by itself which values the community adopts or abandons.

Values define what the community does not want to sacrifice, even under pressure for performance, speed, efficiency, profit, or AI intensity.

Minimum requirement

A change to values or to a significant interpretation of values must pass through a higher human or community decision level.

7. Human ownership of decisions

AI may support decision-making.

It may prepare:

• summaries,
• options,
• arguments,
• risks,
• impacts,
• values conflicts,
• cost scenarios,
• a proposed decision level,
• recommendations.

But significant decisions must have an accountable human or community owner.

Decision distinction

AIFC must distinguish:

analysis
recommendation
proposal
decision
approved change
implemented change

AI may create analysis, recommendation, and proposal.

Decision must belong to accountable governance.

Minimum requirement

Critical decisions must not be executed solely on the basis of AI output without traceable human or community approval.

8. Human ownership of accountability

Accountability cannot be delegated to AI.

AI may perform an action. AI may propose a decision. AI may generate an output.

But accountability remains with a human, team, organization, or community.

This applies especially to:

• customer impacts,
• legal impacts,
• security impacts,
• financial impacts,
• impacts on employees,
• impacts on other communities,
• environmental impacts,
• impacts on Operational DNA,
• impacts on values and purpose.

Minimum requirement

Every significant AI workflow must have a defined human owner or community owner of accountability.

9. AI-generated proposals

An AI-generated proposal is a proposal created by an AI agent or AI tool.

It may concern, for example:

• workflow changes,
• skill updates,
• knowledge base cleanup,
• changes in AI Capacity,
• detection of AI dependency,
• reduction of AI Intensity,
• changes to the AI-NDA Boundary,
• a new fallback,
• strategy adjustment,
• detection of purpose drift,
• resolution of a values conflict.

An AI-generated proposal may be very valuable.

But it must not be silently substituted for an approved decision.

Minimum requirement

AI-generated proposals must be:

• marked as AI-generated,
• linked to an observed signal or reason,
• classified,
• assigned to a decision level,
• reviewed by a human or by accountable governance,
• recorded in the source of truth if significant.

10. AI approval boundaries

An AI approval boundary defines where AI may act independently and where it must request human approval.

This boundary depends on:

• risk,
• data sensitivity,
• impact,
• decision level,
• reversibility of the change,
• AI-NDA Boundary,
• values impact,
• impact on other communities,
• impact on Operational DNA.

Low-risk actions

AI may have higher autonomy for low-risk and reversible actions.

For example:

• proposing a text edit,
• classifying a document,
• proposing tags,
• detecting missing metadata,
• preparing a draft,
• summarizing,
• creating a maintenance proposal.

High-risk actions

AI must have limited autonomy for high-risk actions.

For example:

• changing an active workflow,
• changing the AI-NDA Boundary,
• accessing Operational DNA,
• publishing external content,
• making a decision with customer impact,
• making a financial decision,
• making a security change,
• changing the interpretation of values,
• deleting critical knowledge.

Minimum requirement

Every significant AI workflow must have defined approval boundaries.

11. Human review

Human review is the process by which a human or accountable governance evaluates an AI output.

Human review must not be a purely formal click.

It must be proportionate to the risk.

Review depth

Low risk may require light review.

High risk may require:

• source checking,
• comparison with values,
• security review,
• legal review,
• owner approval,
• decision record,
• fallback test,
• verification of impact on other communities.

Minimum requirement

Human review must be defined for AI workflows that affect critical decisions, customers, security, Operational DNA, or values.

12. Human override

An AIFC community must be able to stop, override, or bypass AI.

Human override means that a human or accountable governance can:

• pause an AI workflow,
• cancel an AI proposal,
• reduce AI autonomy,
• switch into AI-off mode,
• remove an agent’s permissions,
• block AI access to data,
• mark an AI output as rejected,
• return a workflow to a non-AI fallback.

Human override is not a failure of the AI-first approach.

It is a safety and governance mechanism.

Minimum requirement

Critical AI workflows must have a defined human override mechanism.

13. AI autonomy levels

AI autonomy is the degree to which AI may act without continuous human confirmation.

AIFC recommends governing autonomy as a scale.

0 %  - no AI
25 % - AI proposes only
50 % - AI executes drafts with human approval
75 % - AI executes approved low-risk actions with review gates
100 % - AI operates autonomously within strict pre-approved boundaries

100 % autonomy does not mean the absence of governance.

It only means that AI may act independently inside a pre-approved, audited, and limited area.

Minimum requirement

Every AI agent or AI workflow must have a defined autonomy level.

14. AI operating modes

An AI operating mode is a named mode of AI involvement.

Examples:

Conservative
Balanced
Aggressive
Mission Mode
Emergency AI-Off Mode

Conservative

AI mainly proposes; humans decide.

Suitable for:

• new workflows,
• sensitive data,
• low AI maturity,
• high risk.

Balanced

AI helps actively, but significant steps are approved by a human.

Suitable for normal operation.

Aggressive

AI has higher involvement and faster action inside a pre-approved scope.

Suitable for:

• migrations,
• cleanup,
• urgent transformation,
• high work volume with acceptable risk.

Mission Mode

Temporarily increased AI Intensity for a specific objective.

It must have:

• clear purpose,
• time limit,
• budget,
• owner,
• risk boundary,
• exit.

Emergency AI-Off Mode

A mode in which the community turns off or significantly limits AI.

It is used during:

• an incident,
• data leakage,
• vendor problem,
• loss of trust,
• budget limit,
• legal risk,
• critical model failure.

Minimum requirement

A critical AI-first community must have at least a normal AI mode and an AI-off fallback mode.

15. AI dependency risk

AI dependency risk is the risk that a community loses the ability to perform work, make decisions, or restore operations without AI.

AI dependency may emerge slowly and quietly.

Examples:

• a developer cannot finish routine code without AI,
• a team cannot write a customer response without AI,
• the knowledge base is updated only through AI and nobody understands the structure,
• decision material is created only by an agent,
• AI skills contain know-how that has no human skill variant,
• a workflow has no non-AI fallback,
• agent memory contains critical context outside the source of truth.

Minimum requirement

An AIFC community must regularly monitor AI dependency risk and create measures to reduce it.

16. Human Capability Reserve

Human Capability Reserve is the deliberately maintained ability of people to understand, perform, review, or recover work without AI.

AI should extend the community’s capabilities, not replace them in a way that leaves the community impaired without AI.

If a token outage stops simple routine work, the organization has not gained intelligence. It has lost resilience.

Human Capability Reserve may include:

• regular AI-free work,
• preservation of junior tasks,
• fallback checklists,
• training in basic skills,
• human review skills,
• disaster recovery tests without AI,
• documented human skills,
• non-AI operating mode.

Minimum requirement

Critical capabilities of the community must have a human-understandable and recoverable variant.

17. AI-off fallback

AI-off fallback is the ability to continue a critical workflow without AI.

It does not need to mean the same speed.

It must mean acceptable operational capability.

AI-off fallback may be:

• manual checklist,
• simplified workflow,
• reduced service mode,
• alternate vendor,
• local tool,
• human review board,
• static template,
• documented procedure.

Minimum requirement

Critical AI workflows must have an AI-off fallback or an explicitly approved risk of not having one.

18. Role of Source of Truth

The source of truth is the authoritative memory of the community.

AI must not be the system’s only memory.

AIFC requires that:

• AI-generated know-how is returned to the source of truth,
• critical decisions are stored as decision records,
• AI skills are versioned and exportable,
• Operational DNA is not locked in agent memory,
• AI proposals are traceable,
• approved changes are recorded.

AI may help maintain the source of truth. It must not replace it.

Minimum requirement

Significant know-how created with AI must be assessed for inclusion in the source of truth.

19. Role of Human Cockpit Layer

The Human Cockpit Layer is the human access layer to the source of truth.

It has a central role in Human-Managed AI because it makes visible:

• where AI is used,
• what autonomy it has,
• what data it uses,
• who the owner is,
• what proposals AI has created,
• what is waiting for approval,
• where AI dependency is emerging,
• where the budget has been exhausted,
• where AI waste exists,
• where fallback is needed,
• where AI detected risk,
• where human override is needed.

Without the Human Cockpit Layer, AI governance may be formally written down but not human-operable.

Minimum requirement

The community must have a human-accessible way to see and govern significant AI involvement.

20. AI and Operational DNA

Operational DNA is a critical capability of the community.

AI may help Operational DNA be:

• read,
• analyzed,
• cleaned,
• synthesized,
• updated,
• protected,
• validated,
• converted into skills,
• checked for contradictions,
• improved through proposals.

But AI access to Operational DNA must be governed.

Operational DNA must not be handed over uncontrolled to external intelligence.

Minimum requirement

AI access to Operational DNA must be limited, auditable, revocable, and governed by the AI-NDA Boundary.

21. AI and feedback loop

AI may be an important part of the feedback loop.

It may detect:

• recurring problems,
• values conflicts,
• purpose drift,
• AI waste,
• AI dependency,
• missing fallback,
• outdated knowledge,
• missing owner,
• cross-community impact,
• security risk.

AI may prepare a change proposal.

But a change proposal is not a decision.

Minimum requirement

AI-generated feedback must be marked and processed through the same governance mechanism as other significant change proposals.

22. AI and maintenance

AI may help significantly with maintenance.

It may look for:

• outdated content,
• duplicates,
• missing metadata,
• artefacts without an owner,
• deprecated content,
• broken links,
• conflicting rules,
• AI workflows without fallback,
• Operational DNA without review,
• knowledge trapped in chat.

Maintenance is not secondary work.

What a community does not maintain tends to degrade or create debt.

AI can speed up maintenance, but it must not replace ownership of care.

Minimum requirement

AI maintenance proposals must have an owner and lifecycle if they are meant to change the source of truth.

23. AI and skills

AI may help develop human skills and AI skills.

It may:

• extract rules from good outputs,
• propose new skills,
• update checklists,
• detect anti-patterns,
• convert experience into learning material,
• create AI agent instructions.

AIFC requires that an AI skill is not the only place where know-how is stored.

Critical know-how must have a human-understandable form.

Minimum requirement

Critical AI skills must be connected to human-readable knowledge or a human skill.

24. AI and cost control

Human-Managed AI also includes cost control.

AI consumes:

• money,
• tokens,
• compute,
• review time,
• attention,
• governance capacity,
• security capacity.

AI use must be measurable and plannable.

A budget limit may automatically reduce AI Intensity or switch to a restricted mode.

Minimum requirement

Significant AI use must have cost visibility and rules for exceeding budget.

25. AI and risk control

AI risk is not only a technological risk.

It may include:

• security risk,
• legal risk,
• reputational risk,
• AI dependency,
• AI lock-in,
• knowledge leakage,
• values conflict,
• purpose drift,
• incorrect automation,
• loss of human capability,
• uncontrolled decision-making,
• impact on other communities.

Minimum requirement

Significant AI workflows must have a risk assessment proportionate to their impact.

26. AI as external expert capacity

AI may be understood as external expert capacity.

Like a consulting firm, it can bring know-how, speed, and a new perspective.

But, like an external consulting firm, it needs:

• purpose,
• scope,
• NDA boundary,
• budget,
• owner,
• rules for working with know-how,
• audit,
• exit strategy.

This principle is described in detail in:

AIFC-021: AI as External Expert Capacity

Minimum requirement

Significant AI use over non-public know-how must be governed as external expert capacity, not as an ordinary internal tool without boundaries.

27. AI-NDA boundary

The AI-NDA Boundary defines what data AI may see, for what purpose, where it is processed, and how it is protected.

Without an AI-NDA Boundary, AI may function as an uncontrolled external memory of the community.

This principle is described in detail in:

AIFC-022: AI-NDA Boundary

Minimum requirement

AI must not work with non-public or sensitive know-how without an approved AI-NDA Boundary.

28. AI as team member

An AI agent may function as a governed team member.

It must have:

• role,
• scope,
• owner,
• permissions,
• forbidden actions,
• value measurement,
• audit,
• approval rules,
• ability to be turned off.

This principle is described in detail in:

AIFC-023: AI as Team Member

Minimum requirement

An AI agent with tools or access to a non-public knowledge base must have a defined role, permissions, and human owner.

29. Human capability reserve

Human Capability Reserve is described in detail in:

AIFC-024: Human Capability Reserve

In this document, the important foundational principle is:

AI should increase community capability.

It must not remove the ability of people to understand, perform, validate, or recover critical work.

Minimum requirement

The community must monitor whether AI is moving critical capability out of people and the source of truth and into an external model, vendor, or agent memory.

30. Anti-patterns

AIFC rejects the following anti-patterns.

30.1 AI as owner of purpose

AI formulates purpose and the community accepts it without a real decision.

30.2 AI as hidden decision maker

AI outputs become decisions in practice, even though AI was formally supposed only to recommend.

30.3 AI-generated truth without review

AI-generated content is stored as active source of truth without review.

30.4 AI dependency disguised as productivity

A team appears productive, but without AI it loses the ability to perform basic work.

30.5 No AI-off fallback

A critical workflow works only with AI.

30.6 No human owner

An AI workflow has no human or community accountable for the output.

30.7 No approval boundary

It is unclear what AI may execute on its own and what requires approval.

30.8 No cost control

AI consumption grows without planning, measurement, and prioritization.

30.9 No AI-NDA boundary

AI works with internal or sensitive know-how without a clear confidentiality boundary.

30.10 AI memory as source of truth

Agent memory or chat history becomes the informal authoritative memory of the community.

30.11 AI skills without human skills

Critical know-how is available to agents but not to humans.

30.12 Human Cockpit without AI visibility

The human interface shows work, but does not make AI involvement, risks, proposals, and approvals visible.

31. Minimal requirements

In the area of Human-Managed AI, an AIFC community must at minimum:

1. Distinguish between AI-first and AI-dependent states.
2. Have a human or community owner of purpose.
3. Have a human or community owner of values.
4. Have a human or community owner of critical decisions.
5. Give every significant AI workflow a human or community owner.
6. Mark AI-generated proposals as proposals.
7. Ensure that AI-generated decision support material is not automatically treated as a decision.
8. Review significant AI proposals.
9. Define AI approval boundaries.
10. Give critical AI workflows a human override.
11. Govern AI autonomy according to risk.
12. Give critical AI workflows an AI-off fallback or approved risk of not having one.
13. Regularly monitor AI dependency risk.
14. Give critical capabilities a human-readable form.
15. Assess significant AI-generated know-how for inclusion in the source of truth.
16. Govern AI access to Operational DNA through the AI-NDA Boundary.
17. Give significant AI use cost visibility.
18. Give significant AI workflows a risk assessment.
19. Make AI use, proposals, approvals, and risks visible through the Human Cockpit Layer.
20. Give the community the ability to switch critical areas into AI-off or restricted AI mode.

32. Summary

Human-Managed AI is central to AIFC.

AI can bring enormous speed, understanding, and ability to act to a community.

But speed without purpose accelerates chaos. Automation without values increases risk. Agentic work without ownership blurs accountability. AI without fallback reduces resilience. Know-how stored only in an AI tool weakens the community.

AIFC therefore says:

Use AI deeply.
Manage it consciously.
Keep purpose human-owned.
Keep decisions accountable.
Keep knowledge in the source of truth.
Keep the community capable without AI.

An AI-first community is not a community governed by AI.

It is a community structured so well that AI can safely accelerate it.

Human-Managed AI turns artificial intelligence into governed community capacity.