Implementing AI judgment boundaries for public-sector and administrative decisions

Assessment finding

Administrative decision areas are not explicit.

Service B implementation

Create AI-excluded decision lists and administrative decision-type boundaries.

Assessment finding

Review conditions for external impact, legal interpretation, and information management are insufficient.

Service B implementation

Define review conditions, approvers, and thresholds by decision type.

Assessment finding

AI-specific review criteria are missing.

Service B implementation

Create a public-sector AI Human Judgment Review procedure for source, version, jurisdiction, and rationale checks.

Assessment finding

Differences between AI drafts and final documents are not recorded.

Service B implementation

Introduce a public-sector Decision Log for drafts, edits, approvals, and final documents.

Assessment finding

No process reviews boundaries when new apps or models are introduced.

Service B implementation

Define Boundary Governance for additions, changes, approvals, and periodic reviews.

Workflow

Parliamentary or council response preparation

AI use

Past response summarization and draft response creation

Design focus

Define when a draft becomes an organizational position and who must confirm it.

Workflow

Legal, notification, and guideline search

AI use

Search and summarization of relevant provisions and policy documents

Design focus

Require source text, version, and responsible-office confirmation for formal use.

Workflow

Administrative documents and approval materials

AI use

Drafting, summarization, and issue structuring

Design focus

Separate AI drafts from approval decisions and define evidence records.

Workflow

Citizen and business inquiry responses

AI use

Inquiry summary and response drafting

Design focus

Trigger review for external impact, rights and obligations, and personal information.

Deliverable

AI-excluded decision list

Description

Defines areas such as legal interpretation, administrative disposition, and organizational position where AI must not make decisions.

Primary users

Policy teams, legal and document management, information systems

Deliverable

Administrative Decision Boundary™ design

Description

Separates AI support and human judgment across responses, legal search, document drafting, and approvals.

Primary users

Administrative teams, reviewers, management functions

Deliverable

Review trigger and threshold list

Description

Defines review conditions for external impact, legal interpretation, rights and obligations, and information management.

Primary users

Managers, reviewers, document management

Deliverable

Public-sector AI Human Judgment Review procedure

Description

Standardizes source, version, responsible-office, rationale, and unpublished-information checks.

Primary users

Staff, managers, reviewers

Deliverable

Public-sector Decision Log template

Description

Records AI drafts, human edits, approvals, final documents, and approvers.

Primary users

Administrative teams, audit and records management

Deliverable

Boundary Governance rules

Description

Defines review and approval for new tools, model updates, and workflow expansion.

Primary users

AI adoption teams, information systems, control functions

Phase

1. Kickoff and scope definition

Duration

1 week

Work

Confirm target workflows, existing rules, internal AI policy, and assessment findings.

Output

Implementation scope

Phase

2. Decision type mapping

Duration

1-2 weeks

Work

Classify organizational positions, legal interpretation, external responses, information management, and approvals.

Output

Administrative decision type inventory

Phase

3. Decision Boundary™ design

Duration

2-3 weeks

Work

Define AI support, human review, human final judgment, and AI-excluded areas.

Output

Decision Boundary™ design document

Phase

4. Responsibility and review design

Duration

2 weeks

Work

Define reviewers and approval conditions for external impact, legal interpretation, and information management.

Output

Review trigger and responsibility matrix

Phase

5. Evidence design

Duration

1-2 weeks

Work

Define record fields for AI drafts, edits, approvals, and final documents.

Output

Public-sector Decision Log template

Phase

6. Governance design

Duration

1 week

Work

Define approval flows for new apps, model updates, and new use cases.

Output

Boundary Governance rules

Phase

7. Pilot and adoption

Duration

1-2 weeks

Work

Pilot with target teams and verify operational burden and accountability.

Output

Final deliverables and training materials

Decision type

Turning a draft response into an organizational position

AI role

Drafting only

Review condition

External impact, policy judgment, or rights and obligations are involved

Final owner

Responsible office and manager

Log requirement

Record AI draft, edit rationale, approver, and final document

Decision type

Formal legal or policy interpretation

AI role

Search and summarization only

Review condition

Used for a formal decision or external response

Final owner

Responsible office or legal review function

Log requirement

Record source text, version, responsible-office confirmation, and rationale

Decision type

Input of personal or unpublished information into AI

AI role

Generally excluded

Review condition

Exception approval only

Final owner

Information management owner

Log requirement

Record approval rationale, input scope, and purpose

Before

It was unclear when an AI draft became an organizational position.

After

Conditions and reviewers for organizational-position use are explicit.

Before

AI summaries were starting to substitute for source-text review.

After

Source text, version, and responsible-office confirmation are required for formal decisions.

Before

It was difficult to trace who reviewed external responses.

After

Decision Logs record drafts, edits, approvals, and final responses.

Before

Input decisions for personal and unpublished information were left to individual staff.

After

Prohibited, restricted, and exception-approved inputs are defined.

Before

New AI applications did not trigger boundary review.

After

Boundary Governance requires review when new tools or workflows are added.