A better model does not automatically deserve broader permissions
Capability, behavior, and delegated authority answer different questions. Permission changes need evidence about the proposed action and environment.
A FIELD GUIDE TO AI GOVERNANCEPUBLIC SECTOR / TECHNICAL REFERENCE
Make the
control explicit.
What an AI system can do. Who can authorize it. What it costs. And what the evidence actually supports.
Explore the reference libraryMechanisms, not assurances.
PROPOSED ACTIONTRUST BOUNDARY
Model
proposesENFORCEMENTValidate
AuthorizeExecutor
acts
proposesENFORCEMENTValidate
AuthorizeExecutor
acts
DENY / NO EXECUTION
RECORD · OBSERVE · INTERVENE
Who can authorize an action?
Connect permission, resource limits and human intervention to the point where work is executed.
A policy decision cannot stop an action unless the executor applies it.
THE REFERENCE LIBRARY
All entries 15Distinctions that matter.
START HERE / COST & ACCOUNTABILITY
AI cost per decision: the denominator determines the claim
Cost per request, per automated assessment and per completed case measure different units of an AI service. Human handling and case mix determine what the comparison means.
$16,000TOTAL PROCESS COST1,000COMPLETED CASES
=$16PER COMPLETED CASE
CONSTRUCTED EXAMPLE / ONE COST BOUNDARY
2 MIN READ
An AI trace needs a record of the external effect
A successful tool-call span does not by itself establish what changed outside the agent. Link requests, authorization, and resulting state.
Changing embedding models is an index migration
Equal vector dimensions do not establish compatible meaning. Evaluate a new query encoder together with the document representations it searches.
Context, execution state, and memory need separate controls
The information a model sees, the state a workflow resumes, and the records an application retains have different lifecycles.
Provenance explains an answer’s history, not its truth
A traceable chain can preserve an error perfectly. Keep derivation, factual support, and responsibility separate when reviewing AI output.
How an AI budget cap stops spending
A runtime budget cap must account for work before admitting it. Delayed billing, concurrent requests and bypass routes determine whether the limit holds.
When AI time savings become budget savings
Released staff time, avoided future costs and reduced expenditure are different benefits. A cashable AI saving needs a spending change and a defined baseline.
Decommissioning AI without losing the decision record
Stopping an AI service, transferring its records and authorizing disposal are separate operations. A usable exit preserves required evidence and assigns custody after shutdown.
Human oversight requires time and authority to intervene
A reviewer can prevent an AI action only while the action remains controllable. Approval timing, workload and error detection determine what oversight achieves.
Valid JSON is not permission to act
Schema validation checks declared constraints. Authorization connects an authenticated caller to a permitted action on the actual resource.
Open weights do not make an AI service portable
Model parameters are one part of an AI service. Portability also depends on usable rights, runtime dependencies, retrieval indexes and tested behaviour after migration.
Privacy impact assessment and algorithmic impact assessment
A privacy assessment examines information handling; an algorithmic assessment examines automated decision impacts. Shared project evidence does not make their conclusions interchangeable.
What does a public AI register actually count?
A register counts records within a defined publication scope. Coverage, detail and currency are separate properties of the resulting inventory.
Showback and chargeback for shared AI services
Showback reports assigned costs; chargeback posts them to accountable budgets. Both depend on an explicit rule for distributing shared platform costs.
A claim is only as useful
as the boundary around it.
Each entry separates the mechanism from its limits. Sources, related concepts and questions stay with the text.
An assessment is not certification. A policy is not enforcement. A model request is not a completed decision.
Read the control mapKEEP THE DISTINCTIONS CLOSEThe control vs. alignment
The control vs. alignment
taxonomy reference.
Four questions for your next system review.
Get the reference and register for future updates.