AIFC-021: AI as External Expert Capacity

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-010 Knowledge Structure
• AIFC-011 Operational DNA
• AIFC-020 Human-Managed AI

Purpose of this document: Define the principle that AI should be governed in an AIFC community as external expert capacity: with a clear purpose, scope, confidentiality boundary, budget, accountable owner, audit, rules for returning know-how to the source of truth, and exit strategy.

1. Purpose of this document

This document describes how an AIFC community should understand and govern AI as external expert capacity.

AI is not only ordinary software.

AI may:

• read internal know-how,
• synthesize information,
• propose changes,
• formulate strategy,
• analyze risks,
• create documentation,
• propose workflows,
• support decision-making,
• work with Operational DNA,
• act as a team member,
• and accelerate community change.

In this sense, AI resembles external consulting or expert capacity.

A community would normally not give an external consulting firm only a login to a system. It would define the purpose, scope of work, confidentiality, budget, accountability, expected output, knowledge transfer, and ability to end the engagement.

AIFC requires the same level of conscious governance for AI.

2. Core principle

The core principle of this document is:

AI may act as external expert capacity only within a defined purpose, boundary, budget and ownership model.

AI may help a community significantly.

But it must not enter the community’s know-how, decision-making, and operations without rules.

AIFC therefore says:

Do not let external intelligence enter the community without a contract-like boundary.

3. Why this analogy matters

AIFC uses the analogy of an external consulting firm because it is understandable for humans.

A community usually understands that an external consultant needs:

• clear assignment,
• scope,
• NDA,
• access rights,
• budget,
• accountable internal owner,
• expected output,
• deadline,
• quality control,
• rules for working with internal know-how,
• and termination of cooperation.

With AI, these principles are often skipped because AI looks like a tool.

But if AI reads internal documents, proposes changes, generates decision support, or works with Operational DNA, it behaves more like external intelligence than like an ordinary text editor.

Minimum requirement

Significant AI use over non-public know-how must be governed at least as consciously as the involvement of external expert capacity.

4. AI is not just a tool

An ordinary tool performs a predefined function.

AI, however, may interpret, combine, complete, and propose.

For example:

• a text editor saves text,
• AI proposes what the text should mean;
• a search tool finds a document,
• AI summarizes what follows from it;
• a workflow tool starts a process,
• AI proposes a change to the process;
• a reporting tool displays data,
• AI proposes a strategic decision.

AI therefore enters the layer of understanding, interpretation, and decision support.

Minimum requirement

If AI only technically processes a public or low-risk input, it may be governed as an ordinary tool.

If AI interprets non-public know-how, proposes changes, or works with decision support material, it must be governed as external expert capacity.

5. AI engagement model

AIFC recommends governing significant AI use as an engagement.

An AI engagement is a specific involvement of AI for a defined purpose.

Example:

Engagement:
Use AI to analyze existing Confluence documentation and propose AIFC knowledge structure.

Purpose:
Transform scattered documentation into structured source of truth.

Boundary:
AI may read selected internal documentation.
AI must not access restricted customer data.
AI output is proposal, not approved knowledge.

Owner:
Knowledge transformation owner.

Output:
Draft structure, gaps, duplicate detection, change proposals.

Review:
Human approval required before source of truth update.

Minimum requirement

A significant AI engagement must have:

• purpose,
• scope,
• owner,
• allowed data,
• forbidden data,
• expected output,
• review rules,
• rules for writing into the source of truth,
• cost boundary,
• exit or fallback.

6. Purpose and scope

Every significant AI use must have a clear purpose.

Weak:

Use AI to improve documentation.

Better:

Use AI to identify outdated, duplicated and ownerless documents in the internal knowledge base and create maintenance change proposals.

Purpose must be specific enough to make it possible to say:

• what AI may do,
• what AI must not do,
• what data it needs,
• who will review the output,
• when the engagement ends,
• how value will be recognized.

Scope protects the community from AI involvement expanding silently into additional areas.

Minimum requirement

An AI engagement without clear purpose and scope must not receive access to non-public know-how.

7. Internal owner

Every AI engagement must have an internal owner.

The owner is accountable for:

• purpose,
• scope,
• data approval,
• risk assessment,
• output review,
• recording relevant know-how in the source of truth,
• cost control,
• closing the engagement,
• communication with affected communities,
• and deciding whether the AI output will be used.

AI cannot be the owner of its own engagement.

Minimum requirement

Significant AI use without a human or community owner is not AIFC-compatible.

8. Confidentiality boundary

AI as external expert capacity needs a confidentiality boundary.

AIFC describes this boundary in detail in:

AIFC-022: AI-NDA Boundary

The basic principle in this document is:

AI must not automatically see everything.

The community must determine:

• what data AI may read,
• what data AI must not read,
• where data is processed,
• whether it is stored,
• whether it may be used for training,
• who sees prompts and outputs,
• how access is logged,
• how access is revoked,
• how incidents are handled.

Minimum requirement

An AI engagement over non-public or sensitive data must have a defined AI-NDA Boundary.

9. Data access

AI access to data must be governed according to the principles:

least privilege
need to know
purpose limitation
auditability
revocation

AI should receive only the data it needs for the approved purpose.

Not all data that might be useful.

Examples:

• AI for cleanup of public documentation does not need customer data.
• AI for drafting UX text does not need financial data.
• AI for incident analysis does not need the company’s entire Operational DNA.
• AI for a workflow change does not need to see restricted HR documents.

Minimum requirement

An AI engagement must describe its data scope and forbidden data areas.

10. Work output

The output of an AI engagement must be clearly defined.

It may be, for example:

• analysis,
• proposal,
• change proposal,
• decision support,
• document draft,
• cleanup report,
• risk assessment,
• skill update proposal,
• workflow proposal,
• code suggestion,
• test cases,
• migration map.

AI output must not be automatically treated as approved know-how.

AI output is usually:

draft
proposal
interpretation
recommendation

Only after review may it become:

approved
active
source of truth

Minimum requirement

An AI engagement must define what type of output AI creates and how it may become approved knowledge.

11. Knowledge return obligation

An external consultant is not only expected to “do something”. They are also expected to transfer know-how so that it does not remain outside the community.

The same principle applies to AI.

If AI creates or clarifies know-how during work, that know-how must be assessed for inclusion in the source of truth.

For example:

• new rule,
• better definition of a term,
• decision pattern,
• workflow,
• anti-pattern,
• checklist,
• skill update,
• risk,
• maintenance need,
• change proposal.

If AI output remains only in chat, the community does not fully own it.

Minimum requirement

Every significant AI engagement must have a rule for how relevant know-how is returned to the source of truth.

12. Cost boundary

AI as external expert capacity consumes resources.

Not only money, but also:

• tokens,
• compute,
• human review time,
• governance capacity,
• attention,
• security capacity,
• risk capacity.

Every significant AI engagement must have a cost boundary appropriate to its risk and value.

A cost boundary may define:

• maximum budget,
• maximum number of runs,
• maximum number of agents,
• maximum review capacity,
• rules when 80 %, 90 %, or 100 % of budget is reached,
• automatic reduction of AI Intensity,
• switch to Conservative mode.

Minimum requirement

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

13. Value measurement

An AI engagement must be evaluated by value, not only by activity.

It is not enough to measure:

• how many prompts were run,
• how many documents were processed,
• how much text AI produced.

It is necessary to measure:

• what improved,
• what was decided,
• what debt was reduced,
• what risk was detected,
• what skill was created,
• what workflow was improved,
• what know-how returned to the source of truth,
• how much human attention was saved,
• whether AI dependency emerged.

Minimum requirement

A significant AI engagement must be evaluated at least briefly: value, cost, risk, created know-how, next step.

14. Auditability

An AI engagement must be auditable in proportion to its risk.

Audit may include:

• who approved the AI engagement,
• what the purpose was,
• what data was used,
• what model or vendor was used,
• what outputs were created,
• who reviewed them,
• what was accepted,
• what was rejected,
• what was recorded in the source of truth,
• whether an incident occurred,
• what costs were incurred.

Minimum requirement

An AI engagement over restricted data, Operational DNA, or critical decision-making must have an audit trail.

15. Exit strategy

It must be possible to end an AI engagement.

An exit strategy says:

• what happens after the AI tool is ended,
• where the know-how remains,
• how skills are exported,
• how the vendor is replaced,
• how the workflow is restored without AI,
• how data is deleted or closed,
• who verifies that the community has not lost capability.

AI should not become a non-terminable consultant.

Minimum requirement

Critical AI engagements must have an exit strategy or an approved risk of not having one.

16. Vendor and model dependency

An AI engagement may create dependency on:

• a specific vendor,
• a specific model,
• a specific agent memory,
• a specific prompt workflow,
• a specific skill store,
• a specific UI,
• a specific integration.

This dependency is risky if know-how or critical capability moves into it.

Minimum requirement

Critical AI workflows must not depend on a single vendor, model, or proprietary skill store without an exit strategy.

17. Engagement lifecycle

AIFC recommends governing an AI engagement through a lifecycle:

identified
|
proposed
|
risk assessed
|
approved
|
active
|
reviewed
|
closed
|
knowledge returned
|
retrospective

Lifecycle states

Recommended states:

draft
proposed
approved
active
paused
completed
closed
rejected
terminated

Minimum requirement

Significant AI engagements must have a status and owner.

18. AI engagement types

AIFC distinguishes different types of AI engagements.

18.1 Exploration engagement

AI helps understand a problem, possibilities, or direction.

Risk: AI may sound persuasive without sufficient data.

18.2 Knowledge transformation engagement

AI helps convert documentation chaos into a structured source of truth.

Risk: AI interpretation may be mistaken for approved knowledge.

18.3 Operational support engagement

AI helps with normal operations, support, or maintenance.

Risk: gradual AI dependency.

18.4 Decision support engagement

AI prepares material for a decision.

Risk: AI becomes a hidden decision-maker.

18.5 Agentic execution engagement

AI agents execute steps in systems.

Risk: impact without a sufficient approval boundary.

18.6 Security or compliance engagement

AI helps review rules, risks, or compliance.

Risk: work with sensitive data and a false sense of safety.

Minimum requirement

The type of AI engagement must be determined because different types require different governance.

19. Risk assessment

Before a significant AI engagement, risk must be assessed.

Risk may include:

• data sensitivity,
• Operational DNA exposure,
• legal impact,
• security impact,
• reputational impact,
• impact on other communities,
• AI dependency,
• AI lock-in,
• human capability degradation,
• purpose drift,
• values conflict,
• financial impact.

Minimum requirement

An AI engagement with access to restricted data, Operational DNA, or critical decision-making must have a risk assessment.

20. Relationship with AI-NDA Boundary

The AI-NDA Boundary defines the confidentiality boundary.

An AI engagement defines the specific involvement of AI.

Relationship:

AI engagement
-> why and for what we use AI

AI-NDA Boundary
-> what data and know-how AI may see and under what conditions

An AI engagement without an AI-NDA Boundary may be acceptable only for public or low-risk data.

Minimum requirement

If an AI engagement works with non-public data, it must reference an approved AI-NDA Boundary.

21. Relationship with Human Capability Reserve

AI as external expert capacity must not degrade the community’s human capabilities.

An external consultant may help, but if the organization cannot continue after the consultant leaves, the engagement has failed.

The same is true for AI.

An AI engagement must also be assessed by whether it:

• strengthened human capabilities,
• created a human skill,
• captured know-how,
• reduced dependency,
• or instead moved capability outside the community.

Minimum requirement

A critical AI engagement must evaluate its impact on the Human Capability Reserve.

22. Relationship with Operational DNA

An AI engagement over Operational DNA is highly sensitive.

Operational DNA describes how the community actually works.

If AI gains uncontrolled access to Operational DNA, the community may lose control over its most valuable know-how.

Minimum requirement

An AI engagement with Operational DNA must have:

• explicit approval,
• limited scope,
• AI-NDA Boundary,
• audit,
• human owner,
• output review,
• exit strategy,
• rule for knowledge return.

23. Relationship with Human Cockpit Layer

The Human Cockpit Layer must allow people to see significant AI engagements.

It should display:

• engagement purpose,
• owner,
• status,
• data boundary,
• cost status,
• risk,
• outputs,
• pending review,
• created change proposals,
• knowledge return status,
• possible dependency risk.

Without human visibility, an AI engagement may become a hidden operational mechanism.

Minimum requirement

Significant AI engagements must be human-visible to the accountable members of the community.

24. AI engagement record

AIFC recommends using an AI engagement record for significant AI engagements.

Example metadata:

ai_engagement:
  id:
  title:
  status: draft | proposed | approved | active | paused | completed | closed | rejected | terminated
  engagement_type:
    - exploration
    - knowledge_transformation
    - operational_support
    - decision_support
    - agentic_execution
    - security_compliance
  purpose:
  scope:
  owner:
  approved_by:
  data_boundary:
  ai_nda_boundary:
  allowed_data:
  forbidden_data:
  tools_or_models:
  output_type:
  human_review_required: true | false
  approval_boundary:
  cost_limit:
  risk_level:
  operational_dna_access: true | false
  fallback:
  exit_strategy:
  knowledge_return_required: true | false
  source_of_truth_targets:
  retrospective_required: true | false

This structure is illustrative.

The final schema should be defined in the agent-actionable layer of the standard.

25. Anti-patterns

AIFC rejects the following anti-patterns.

25.1 AI without purpose

AI is used because it is available, not because it has a clear purpose.

25.2 AI without scope

An AI engagement gradually expands beyond its original area without governance.

25.3 AI without owner

Nobody is accountable for outputs, costs, risks, or knowledge return.

25.4 AI without confidentiality boundary

AI works with non-public know-how without an AI-NDA Boundary.

25.5 AI as invisible consultant

AI creates significant material and proposals, but its role is not visible or auditable.

25.6 AI output as accepted truth

AI output is treated as approved knowledge without review.

25.7 AI engagement without knowledge return

AI helps solve a problem, but know-how remains in chat and does not return to the source of truth.

25.8 AI engagement without exit

The community begins depending on an AI tool without a termination or replacement plan.

25.9 AI engagement creating human degradation

AI speeds up work, but people lose the ability to understand or perform the work without AI.

25.10 AI vendor as hidden memory

An external AI tool becomes the informal memory of the community.

26. Minimal requirements

In the area of AI as External Expert Capacity, an AIFC community must at minimum:

1. Understand significant AI use as an AI engagement.
2. Give every significant AI engagement a clear purpose.
3. Give every significant AI engagement a scope.
4. Give every significant AI engagement a human or community owner.
5. Give AI engagement over non-public data an AI-NDA Boundary.
6. Define allowed and forbidden data for the AI engagement.
7. Define the output type of the AI engagement.
8. Ensure AI output is not automatically treated as approved know-how.
9. Assess significant know-how created by the AI engagement for inclusion in the source of truth.
10. Give significant AI use cost visibility.
11. Give higher-risk AI engagements a risk assessment.
12. Give AI engagement over Operational DNA explicit approval and audit.
13. Give critical AI engagements an exit strategy or approved risk of not having one.
14. Give the AI engagement a lifecycle status.
15. Make significant AI engagements visible in the Human Cockpit Layer.
16. Evaluate the impact of critical AI engagements on the Human Capability Reserve.
17. Prevent AI engagement from creating uncontrolled AI lock-in.
18. Prevent AI engagement from turning AI into external memory of the community without rules.

27. Summary

AI can be an extraordinarily powerful external expert capacity for a community.

It can bring speed, synthesis, proposals, new perspectives, and the ability to work with large amounts of know-how.

But precisely because of that, it must be governed.

External expert capacity without purpose, boundary, budget, owner, and exit strategy can become an uncontrolled influence inside the community.

AIFC therefore says:

Invite AI like an expert.
Govern AI like a consultant.
Capture its contribution like community knowledge.
Exit it like a vendor.
Never let it own the purpose.

AI as External Expert Capacity turns AI usage into governed engagement.