Practical guide

How to apply the ASEAN Guide on AI Governance and Ethics to knowledge-worker AI programmes

The ASEAN Guide on AI Governance and Ethics is voluntary guidance, not a certification scheme or a substitute for law. A knowledge-worker AI programme can use it as an organizing framework for accountable ownership, proportionate human involvement, operational controls, and stakeholder communication. Implementation should begin with a defined use case and evidence, not a generic principle.

By Two Prune

The ASEAN Guide on AI Governance and Ethics is voluntary guidance, not a certification scheme or a substitute for law. A knowledge-worker AI programme can use it as an organizing framework for accountable ownership, proportionate human involvement, operational controls, and stakeholder communication. Implementation should begin with a defined use case and evidence, not a generic principle statement. This operating method uses ASEAN: applicability, structure, evaluation, action controls, and notice and review.

Source basis, status, and applicability

The ASEAN Secretariat's 2024 Guide focuses on traditional AI in commercial, non-military or non-dual-use applications. Its organizational framework covers internal governance structures, human involvement in AI-augmented decisions, operations management, and stakeholder interaction and communication. The Expanded Guide supplements it with generative-AI considerations, including accountability across the technology stack, data traceability, dynamic testing, incident reporting, content provenance, security, and capability development. The 2025-2030 Responsible AI Roadmap provides regional policy direction, not organization-level obligations. Applicable national laws, sector rules, contracts, and organizational policies remain controlling. ASEAN Guide Expanded Guide Responsible AI Roadmap

Sources: [1] [2] [3]

A - Confirm applicability and classify statements

Create a source map before implementing controls.

| Statement | Source and section | Classification | Local applicability | Owner | Evidence needed | | --- | --- | --- | --- | --- | --- | | | | Direct guide text / local inference / optional practice / external requirement | | | |

Ask:

Do not restate voluntary guidance as a legal requirement. Route legal interpretation to qualified counsel.

  • Is the programme using traditional AI, generative AI, or both?
  • Is the use commercial and outside military or dual-use contexts?
  • Which ASEAN member state, sector, data, and contractual rules apply?
  • Is the organization a developer, deployer, user, purchaser, or several of these?
  • Which decisions affect rights, safety, access, finances, or other material interests?

S - Establish structure and responsibility

Use existing governance where it can carry the work; create a new body only when the responsibility, expertise, or independence is missing.

| Decision or control | Accountable owner | Contributors | Authority limit | Escalation route | Retained evidence | | --- | --- | --- | --- | --- | --- | | Use-case approval | | | | | | | Data and tool approval | | | | | | | Human-involvement level | | | | | | | Testing and release | | | | | | | Incident response | | | | | | | Stakeholder communication | | | | | |

Make responsibility explicit across vendor, internal technology, programme owner, users, reviewers, and affected stakeholders. A supplier contract does not transfer all accountability away from the deployer.

E - Evaluate the use case and human involvement

Build a use-case record:

| Field | Question | | --- | --- | | Purpose and benefit | What problem and authorized outcome are being pursued? | | People affected | Who uses, reviews, or experiences the output? | | Inputs and data | What enters the system, from where, under what permission? | | Output and decision | What does AI produce, and who decides what happens next? | | Failure modes | What can be wrong, harmful, misleading, inaccessible, or insecure? | | Human involvement | Who reviews, can override, can stop, and is accountable? | | Evidence | What testing, monitoring, feedback, and incidents are retained? |

Set the level of human involvement from context and consequence. “Human in the loop” is incomplete unless the person has time, competence, evidence, authority, and a usable intervention route.

A - Implement action controls across the lifecycle

### Data and provenance

### Testing and operations

### Knowledge-worker capability

  • inventory approved data types and sources;
  • record permission, sensitivity, quality concerns, transformations, and retention;
  • preserve source links or provenance where feasible;
  • block restricted inputs from unapproved services;
  • verify whether outputs reproduce sensitive or unsupported material.
  • define expected behavior and unacceptable outcomes before release;
  • test representative and difficult cases, including local languages and accessibility where relevant;
  • compare output with a non-AI or prior process when that comparison is meaningful;
  • monitor changing models, prompts, data, integrations, and user behavior;
  • record incidents, near misses, overrides, and unresolved limitations;
  • maintain rollback, suspension, correction, and exit routes.
  • train people on the task, evidence, limits, and escalation route, not only tool features;
  • give reviewers access to original sources and decision criteria;
  • test whether controls operate in realistic work;
  • refresh guidance when systems, risks, policies, or work change.

N - Notify stakeholders and review evidence

Communication should be proportionate to the relationship and effect. Document when people need to know that AI is used, what role it plays, what information is processed, how to question or correct an outcome, and which limitations matter. Avoid broad transparency claims that expose security or proprietary information without helping the affected person.

At each review gate, ask:

Retain the decision and evidence. A policy, training completion record, or dashboard alone does not show that controls operated.

  • Does the use case still match the authorized purpose?
  • Are ownership and human intervention operating as designed?
  • What do testing, overrides, incidents, and stakeholder feedback show?
  • Have data, vendors, models, integrations, laws, or context changed?
  • Which control failed, and what repair or restriction follows?
  • Should the programme continue, narrow, pause, or stop?

Worked example: fictional research-summary assistant

A regional professional-services team wants a generative-AI assistant to summarize public market research for internal proposals. The initial plan says a manager will review every answer and the vendor is responsible for model safety.

ASEAN begins by classifying the assistant as a generative-AI deployment used in commercial knowledge work. The organization remains responsible for its purpose, inputs, workflow, and downstream use. The use-case record limits inputs to approved public sources, prohibits client-confidential material, and states that the output is a research aid rather than evidence on its own.

The human-involvement design names the proposal lead as decision owner and requires source-level review for claims used externally. Reviewers receive links and excerpts, not only a fluent summary. Tests include stale sources, fabricated citations, conflicting reports, a non-English source, and a request containing restricted client data. The team records unsupported-claim rate by test set, reviewer overrides, blocked inputs, and incidents, while preserving the limits of each measure.

The first test finds that managers approve quickly when citations look plausible. The programme pauses external use, adds a claim-to-source ledger, requires passage-level checking for material claims, and gives reviewers more time. Stakeholder communication tells staff what data is prohibited and how to report an incident. This is evidence of a repaired local control, not a claim of legal conformity or risk elimination.

Implementation evidence pack

applicability and source-classification map;

  • use-case and data record;
  • responsibility and authority table;
  • human-involvement rationale;
  • test plan, cases, results, and limitations;
  • incident, override, and correction log;
  • stakeholder notices and feedback route;
  • review decisions, conditions, and version history.

Two Prune boundary

Two Prune can assess whether a participant applies supplied governance guidance, navigates evidence, uses AI thoughtfully, verifies outputs, identifies authority limits, and communicates a controlled recommendation in a realistic simulation. It does not provide legal advice, certify conformity with ASEAN guidance, operate an organization's AI governance programme, eliminate AI risk, or make employment decisions.

Sources

ASEAN Secretariat: ASEAN Guide on AI Governance and Ethics

  • ASEAN Secretariat: Expanded ASEAN Guide on AI Governance and Ethics - Generative AI
  • ASEAN Secretariat: ASEAN Responsible AI Roadmap 2025-2030

Sources: [1] [2] [3]

Sources

  1. 1.ASEAN Guide · asean.org
  2. 2.Expanded Guide · asean.org
  3. 3.Responsible AI Roadmap · asean.org