Skip to content

AI Applicability Rationale

Determining applicability across a newly adopted framework means writing a justification for every obligation you exclude. For a 93-clause standard with a dozen exclusions, that is a dozen pieces of careful writing.

AI can draft them. It cannot decide them.


What it does

Given an applicability record, the AI reads the requirement text and the scope context and proposes:

  • Draft justification wording
  • A suggested applicability status

Both are advisory. Neither is applied.


The authority boundary

AI doesAI does not
Read the requirement and its framework contextSet the applicability status
Read the declared scope characteristicsSubmit the decision
Draft justification wordingApprove anything
Suggest a likely determinationMake the record authoritative

A suggested status is not a decision

The AI may suggest Not Applicable. The record stays where it is until a person sets the status, and stays non-authoritative until an independent approver approves it.

There is no path by which an AI suggestion becomes an approved applicability decision.

The permission split enforces this: generating a draft needs only compliance.read plus the AI entitlement. Recording the decision needs compliance.manage or obligations.update. Approving it needs compliance.validate or workflow.approve, and cannot be done by the person who recorded it.


What the model can and cannot know

This distinction is the whole reason the human gate exists here.

The model can readThe model cannot know
The requirement textWhether your organisation actually operates that business line
The framework and clause structureWhich entity in the group holds which licence
The scope name, type and purposeWhat was agreed with your supervisor last year
Scope membership characteristicsWhether an obligation is discharged elsewhere under a different scope
Related requirements in the frameworkYour legal counsel's view of an ambiguous provision

An AI draft that says "the institution does not process cardholder data" is repeating a plausible pattern, not asserting a fact about you. Verifying it is the work.


Reviewing a draft

Check every factual claim. A justification is a factual assertion about your organisation. If the draft says you do not operate a trading book, confirm that with someone who would know.

Check the reasoning, not the fluency. AI drafts read well. Reading well is not the same as being right, and a fluent wrong justification is more dangerous than a clumsy correct one because it invites less scrutiny.

Add what the model could not know. The strongest justifications cite a named person, a date and a source: "Confirmed with the Head of Payments, March 2026." No model can produce that line.

Reject drafts for statutory obligations. Where a requirement is statutorily mandatory, a Not Applicable justification is expected to cite the lawful waiver or exemption relied on. That is a legal position, and it should come from someone qualified to take it.


Where a draft genuinely helps

SituationWhy a draft helps
A clearly out-of-scope technical clauseThe reasoning is mechanical; the draft saves the typing
A long exclusion list in one framework domainConsistency of structure and tone across a dozen justifications
An analyst new to the frameworkThe draft surfaces what a justification should address
Restating a decision already made for another scopeThe context is already established; the draft is a reformulation

Where it does not

SituationWhy
Anything jurisdictionalThe model does not know which regime binds you
Anything depending on group structureIt does not know your legal entity map
Anything an assessor is likely to challengeThese need a human argument, not a generated one
Statutorily mandatory obligationsThese need a legal position

Permissions

ActionPermission
Request an AI rationale draftcompliance.read plus the AI entitlement
Record the applicability decisioncompliance.manage or obligations.update
Approve or reject the decisioncompliance.validate or workflow.approve

Example

Requirement: a clause on secure development environment separation.

Scope: SCP-2026-0007 — Payments Platform Production, 63 members, all production systems.

AI draft: "This requirement addresses the separation of development, testing and production environments. The declared scope comprises production systems only. Development activity does not occur within this boundary."

What the analyst changed:

"No development activity occurs within this scope. All change is delivered by the group engineering function under separate scope SCP-2026-0009, where this clause is Applicable and addressed by CTL-2026-0052. Confirmed with the Head of Platform Engineering, 14 March 2026."

The difference: the AI draft states that development does not happen here. The analyst's version states where it does happen, that the obligation is met there, under which scope, by which control, and who confirmed it.

An assessor reading the first version asks a follow-up question. Reading the second, they do not.


Troubleshooting

"AI rationale is unavailable." Requires the AI entitlement. Manual justification is unaffected and always available.

"The draft states something untrue about our organisation." Expected, and the reason for review. Edit or discard it.

"The suggested status was applied automatically." It cannot be. Check the record — if the status changed, a person changed it.

"Drafts across a framework read identically." They will. If a dozen exclusions share one generated structure, an assessor will notice. Vary them by writing what is actually specific to each.


OrviQ Enterprise Governance, Risk & Compliance Platform