AI Use Assessment Framework

Oct 11, 2026 · @Damian Hickey

1. Summary

The AI use assessment establishes, for a representative set of an insurer's arrangements, where AI acts, what it is able to do, and what evidence of its behaviour can be produced at each boundary in the chain. It is the technical half of a joint engagement led by Norton Rose Fulbright (NRF). NRF answers what the insurer can require; this assessment answers what can actually be produced.

Item	Position
Question answered	At each boundary in the chain, what evidence of AI behaviour can be produced, measured against the board's appetite?
Unit of assessment	One arrangement: a material service provider, the platform behind it, or an external AI service the insurer's own staff use
Instruments	The AI use register, the seven questions and the evidence ladder
Duration	Three to four weeks
Inputs	The insurer's AI inventory and the agreement for each arrangement in scope
Not required	Policyholder data, provider involvement, production access
Outputs	The register, one finding per arrangement, a technical schedule and a technical annex in Triodian's name
Independent check	The independent reviewer, on behalf of the CRO
Measure	APRA's standards and expectations. The assessment is solution-agnostic and presents no product

Status of this document. The method, the seven questions, the ladder, the roles and the outputs come from the NRF planning deck (October 2026) and the AI Counsel Briefing (12 September 2026). The field-level detail, codes, rating rules, weekly plan and templates are drafted here to make the method executable. Section 12 lists each drafted element that needs confirmation.

Standing rule. Every legal determination is the insurer's and its counsel's. Triodian supplies the framework, the facts and the technical classification. Nothing in the assessment or its working papers states a legal conclusion.

2. Scope and boundaries

2.1 The chain and its boundaries

The assessment follows one chain and tests three kinds of boundary. Each register entry and each finding is tagged with one of them.

Code	Boundary	Parties	What passes across it
B1	Insurer to third party	The insurer and a material service provider: claims manager, underwriting agency or broker acting under delegated authority	Agreement terms outward; aggregate evidence back
B2	Third party to fourth party	The provider and the platform behind it: the technology, pricing or AI model the provider relies on	Flow-down terms outward; what the provider can obtain back
B3	Insurer's own use	The insurer's legal, claims and compliance staff and an external AI service	The insurer's own rules outward; a signed record back

2.2 The representative set

The assessment covers a representative set of arrangements agreed with the General Counsel, not the whole estate. The default is four arrangements, the number used in the September counsel briefing.

Selection criteria, applied in this order:

1. A live trigger. A regulator request, a remediation or a major provider renewal that sets a date.
2. AI inside a provider. Business written through agencies or outside claims managers, where the agreement holds the audit rights.
3. The insurer is itself asked for evidence. For example a group life insurer answering to superannuation trustees.
4. Coverage of boundary types. At least one B1, one B2 reached through it, and one B3.
5. Spread of authority. At least one arrangement where AI decides or transacts, and one where it only assists.

2.3 In scope

• AI used in or supporting a service to the insurer, wherever in the chain it sits.
• AI used by the insurer's own staff through an external service.
• The agreement, and any schedule or side letter, for each arrangement in the set.
• The insurer's existing AI inventory, AI governance framework and material service provider register.

2.4 Out of scope

• Legal advice. Reach, applicability, grading and privilege are NRF's.
• Policyholder or claims data. The assessment works from documents and interviews.
• Provider involvement. No provider is approached. Obtainability is assessed from the insurer's side.
• Production access. No system is connected to, tested or monitored.
• Model testing. No model is evaluated for accuracy, bias or drift.
• Products and gap filling. Closing a gap is a separate decision for the insurer after the findings.
• The whole estate. Arrangements outside the set are not assessed. Uses found outside the inventory are handed to its owner.

3. Framework at a glance

Three instruments are applied in four steps, the same way to every arrangement at every client.

3.1 The three instruments

Instrument	What it is	Section
AI use register	One entry for each place AI acts in the chain, with a fixed set of fields	6
The seven questions	The ordered questions put to every arrangement. Two sit under all of them: what can the insurer require, and what can actually be produced	7
Evidence ladder	A scale of evidence types, matched to what the AI can do	8

3.2 The four steps

1. Confirm the set. A representative set of arrangements, agreed with the GC.
2. Map the AI. Where AI acts, what authority it holds and who stands behind it. Each use becomes a register entry.
3. Classify and test. The consequence of each use, and the evidence obtainable at each boundary.
4. Questions to counsel. Legal questions go to NRF in writing. Counsel sees every draft first. Counsel's determinations are then recorded against each entry and the finding is set.

3.3 The logic of a finding

Each arrangement produces three positions on the ladder. The finding is the relationship between the last two.

Position	Meaning	Set by
Indicative rung	The evidence type that corresponds to what the AI can do	Triodian, by rule (section 8)
Required rung	The evidence the insurer can reasonably require, given reach and rights	The insurer and counsel
Obtainable rung	The highest evidence type that can in fact be produced and independently checked	Triodian, by test (section 8.4)

3.4 Operating principles

• Aggregate only. No individual decisions, no policyholder data, no access to a provider's model.
• Start from what exists. The insurer's own inventory is the first input. Gaps go back to its owner to fix.
• "No reasonable reach" is a conclusion. It is documented as a finding, not treated as a gap.
• Solution-agnostic. The test is APRA's: can the evidence be produced, and can someone independent check it?
• Every stage can stop. The assessment leaves a usable register and findings whether or not anything follows.

4. Roles and governance

4.1 Parties

Party	Role	Engaged by	Where it stops
The insurer	Client. The CRO raises the risk; the GC decides the brief and owns the outcome	n/a	n/a
Norton Rose Fulbright	Leads the engagement, holds the client, engages both firms, owns every legal call	The insurer	n/a
Triodian	Technical party. Does the assessment and the register. Responsible for the result	NRF	Gives no legal advice. Fills gaps only if the insurer asks, as a separate decision
Independent reviewer	Independent check of Triodian's assessment, on behalf of the CRO	Counsel or the CRO, separately from Triodian (to be confirmed)	Does not do the assessment or build any control. Gives no legal advice
KJR	Scaling partner for the register in later engagements	To be agreed	Not part of the first assessment

4.2 Responsibility by activity

R = does the work. A = accountable for it. C = consulted. I = informed.

Activity	GC	Insurer AI and risk teams	NRF	Triodian	Independent reviewer
Decide the brief and the set	A	C	R	C	I
Brief the AI, governance and safety teams	I	C	A	R	I
Supply inventory and agreements	A	R	C	I	I
Build and extend the register	I	C	I	A, R	I
Classify authority and failure consequence	I	C	C	A, R	I
Test evidence obtainability	I	C	C	A, R	I
Determine nexus, rights, reach and required evidence	A	I	R	C	I
Grade requirement, expectation or better view	A	I	R	I	I
Draft findings and technical annex	I	C	C	A, R	I
Check the assessment and its findings	I	I	A	C	R
Board paper and flow-down clause schedule	A	I	R	C	I

4.3 Team

Role	Person	Responsibility in the assessment
Method owner	Damian Hickey	Scope, sequence and the evidence each stage has to produce. Signs the technical annex
Delivery manager	Leon Cosgrove	Plan, document requests, interview schedule, issue log
AI architect	Abhishek Gupta	Technical reading of each system: authority, credentials, where a record could be produced
Validation	Sam Ferraro PhD	Checks the measures without having written them, before anything goes to the independent reviewer

4.4 Rules of engagement

• Counsel first. Every draft goes to NRF before anyone at the insurer sees it.
• Legal questions in writing. Any point needing a legal view is logged as a numbered question to counsel (template in section 9.5) and is not answered in a working paper.
• Privilege. The work is scoped through counsel, on the privilege position NRF determines. Document marking, storage and circulation follow NRF's instruction.
• One technical party. NRF deals with Triodian for all technical matters.
• Independence of the check. The independent reviewer receives the register, worksheets and annex from counsel or the CRO, not from Triodian, and reports on behalf of the CRO.

5. Workplan

The plan below fits the four steps into four weeks. Where the set is small and documents arrive on day one, weeks 2 and 3 compress into one.

5.1 Phases

Phase	When	Activities	Output	Exit gate
0. Mobilise	Before week 1	NRF convenes the legal and technical briefing. GC decides the brief. Privilege position and document handling set by NRF. Document request issued	Signed brief. Document request list. Agreed contacts	GC has confirmed the brief in writing
1. Confirm the set	Week 1, days 1 to 2	Apply the selection criteria in section 2.2. Agree the arrangements and their boundary types with the GC	Set register: arrangement ID, parties, boundary type, reason selected	GC and NRF agree the set
2. Map the AI	Week 1, day 3, to end of week 2	Brief the AI, governance and safety teams. Take the insurer's inventory as the first input. Read each agreement. Interview arrangement owners. Extend the register down the chain	Register with fields 1 to 14 complete (section 6). Interview notes. Source index	Every entry has a source and a date. Uses not on the inventory are logged for hand-back
3. Classify and test	Weeks 2 to 3	Classify authority and failure consequence. Set the indicative rung. Run the obtainability test at each boundary. Internal validation review	Register fields 15 to 22 complete. Classification rationale per entry. Draft worksheets	Validation review passed with no open exceptions
4. Questions to counsel	Week 3	Issue the numbered question set to NRF. Record counsel's determinations on nexus, rights, location and required evidence	Question log with responses. Register fields 23 to 25 complete	Every question answered, or recorded as open with an owner
5. Findings and report	Weeks 3 to 4	Set one finding per arrangement. Draft the technical schedule and annex. Draft to counsel first	Finding worksheets. Technical schedule. Technical annex, draft	NRF accepts the draft for independent check
6. Independent check	Week 4	The pack goes to the independent reviewer through the agreed reporting line. Triodian answers its queries through the same line. Corrections made and logged	The independent reviewer's check, reported on behalf of the CRO. Final annex	The check is confirmed complete through the agreed reporting line
7. Close	End of week 4	Hand uses not on the inventory back to its owner. Hand the register to NRF. Archive working papers as NRF directs	Hand-back list. Final register. Closure note	GC receives NRF's advice with the annex attached

5.2 Document request list

Issued in phase 0 through NRF.

Ref	Document	Used for
D1	Current AI inventory or register, with owner and date	First input to the register
D2	AI governance framework and acceptable-use policy	Authority and the insurer's own rules at B3
D3	Material service provider register	Confirming the set and boundary types
D4	Board risk appetite statement and tolerance levels for the critical operations in the set	The appetite the evidence is measured against
D5	The agreement for each arrangement, with schedules, service descriptions and side letters	Rights, audit and information terms, sub-contracting terms
D6	Delegated authority schedules for each binder or claims agreement	What authority the provider holds, and so what its AI can hold
D7	Any provider disclosure about AI use already received: due diligence responses, attestations, notices	Evidence already obtainable
D8	Licence terms and security documentation for each external AI service staff use	B3 entries
D9	Existing assurance over the arrangements: audit reports, control reports, certifications	What existing assurance already covers
D10	Incident and near-miss records involving AI in the set, last 24 months	Failure consequence

5.3 Interviews

Who	Purpose	Length
Head of AI or AI governance	Agree the approach. Walk through the inventory and its known limits	60 min
Owner of each arrangement in the set	What the provider does, where AI is known or suspected, what has been asked of the provider	45 min each
Procurement or vendor management	How agreements are renewed and what evidence terms exist today	45 min
Operational risk, second line	Critical operations, tolerances and how provider risk is monitored now	45 min
Users of each external AI service at B3	What the service is used for and what it is permitted to do	30 min each

6. AI use register

The register holds one entry for each place AI acts within an arrangement. An arrangement with AI at the provider and at the platform behind it has at least two entries.

6.1 How it is built

1. Brief the AI teams first. Agree the approach with the AI, governance and safety teams.
2. Start from what exists. The insurer's own AI inventory is the first input.
3. Extend it down the chain. Add the AI inside providers and the platforms behind them.
4. Classify each entry. Authority, consequence, rights and evidence.
5. Hand the gaps back. Uses not on the inventory go to its owner to add.

6.2 Fields

The deck names eight things each entry records. The 26 fields below break those eight into items a reviewer can test one at a time.

#	Field	Allowed values or format	Completed by	Phase
1	Entry ID	REG-001 onward	Triodian	2
2	Arrangement ID	ARR-01 onward, from the set register	Triodian	2
3	Boundary type	B1, B2, B3	Triodian	2
4	Where the AI acts	Own staff · provider · platform behind a provider	Triodian	2
5	Business function and critical operation supported	Free text, plus the critical operation name from the insurer's CPS 230 list, or "none"	Triodian with arrangement owner	2
6	On the insurer's inventory	Yes · no · partly	Triodian	2
7	System name and purpose	Free text	Triodian	2
8	Supplier	Legal entity name	Triodian	2
9	Platform and model provider behind it	Legal entity names, or "not known"	Triodian	2
10	Location of each party, and any offshoring	Country per party	Triodian	2
11	What the AI does	Two sentences, plain words, verbs only from the source	Triodian	2
12	Human involvement	A person acts · a person decides · the output is the decision · unattended	Triodian	2
13	Credentials and systems the AI can change	List, or "none"	Triodian	2
14	Agreement terms on information, audit, sub-contracting and notification	Clause references with verbatim extracts	Triodian	2
15	Authority tier	A1 to A6 (section 8.1)	Triodian	3
16	Failure consequence	C1 to C5 (section 8.2)	Triodian	3
17	Indicative rung	E0 to E6 (section 8.3)	Triodian, by rule	3
18	Evidence currently received	Rung and description, or "none"	Triodian	3
19	Evidence the system could produce	Description of the record and where it would be generated	Triodian	3
20	Independently checkable	Yes · no · partly, with reason	Triodian	3
21	Obtainable rung	E0 to E6 (section 8.4)	Triodian, by test	3
22	Confidence in obtainability	High · medium · low	Triodian	3
23	Nexus, relationship and location determination	Counsel's wording, with question reference	NRF	4
24	Rights the insurer holds	Passive · significant influence · contractual · control or board seat	NRF	4
25	Required rung and grading	E0 to E6, graded requirement · expectation · better view	NRF	4
26	Finding and gap owner	F1 to F4 (section 9.1), with named owner	Triodian, after counsel	5

6.3 Completion rules

• Every value has a source. Each field carries a source reference (document and page, or interview and date). A value without a source is recorded as "not known", never estimated.
• Field 11 is descriptive, field 15 is the judgement. The description is written before the tier is chosen, so a reviewer can re-derive the tier from it.
• Fields 23 to 25 are never completed by Triodian. They are transcribed from counsel's written response, with the question number.
• "Not known" is a valid entry and is counted. The annex reports how many fields are not known, by arrangement.
• Confidence (field 22) is low by default at B2. No provider is approached, so a fourth-party rung rests on documents alone unless field 14 or D7 supports it.
• One classifier, one reviewer. Fields 15 to 22 record who classified and who reviewed. They cannot be the same person.

7. The seven questions

The questions are worked in order for every arrangement. The determination at each step is the insurer's and its counsel's. Triodian supplies the framework and the facts each determination needs.

The paragraph references are those cited in the two decks. Counsel confirms them against the current instrument before they appear in any deliverable.

7.1 Question set

#	Question	Answer scale	Determined by	Facts Triodian supplies	Register fields
1	Does a separate CPS 230 nexus exist? Is the party within the insurer's group (para 5)?	Yes · no	Counsel	Legal entity and ownership of each party	8, 9, 23
2	Is there a separate relationship that may engage CPS 230: a service to the insurer, or part of a chain supporting one (para 54)?	Yes · no	Counsel	The service provided, the critical operation it supports, and each link in the chain	5, 8, 9, 23
3	What rights does the insurer or its manager hold?	Passive · significant influence · contractual, including binder and claims agreements · control or board seat	Counsel	Verbatim information, audit, sub-contracting and notification terms	14, 24
4	Where is each party, and does location change what can be obtained? Enforceability, data access and audit rights; offshoring (paras 53(b) and 60(b))	Per party: no effect · limits evidence · prevents evidence	Counsel	Country of each party and of where the AI runs	10, 23
5	What authority has been delegated to the AI, and what can its action produce?	Productivity · analytical · decision · transactional · agentic · actuating	Triodian classifies; counsel determines the authority delegated	Function, human involvement, credentials	11, 12, 13, 15
6	What could failure do?	Immaterial · contract breach · financial loss · policyholder harm · critical-operation disruption	Triodian classifies; insurer confirms	Critical operation link, tolerance, incident history	5, 16
7	What evidence can reasonably be required?	None · disclosure · attestation · event notification · audit · continuous telemetry · deterministic actuation control	Counsel determines; Triodian tests obtainability	Indicative rung, evidence received, evidence producible, checkability	17 to 22, 25

7.2 Tests under each question

Q1 and Q2, nexus and relationship. Triodian does not answer these. It supplies an entity and service map for each arrangement: who contracts with whom, which entity operates the AI, and which critical operation the service supports. Two points from APRA's guidance, as cited in the counsel briefing, are flagged for counsel on each arrangement: a related party providing a service is assessed on the same footing as a third party, and providers relied on for the same critical operation may be assessed as a cohort.

Q3, rights. For each agreement Triodian extracts, without interpretation:

• any term giving the insurer information about how the service is performed
• any audit or access right, and what it reaches
• any term on sub-contractors and whether the insurer must be notified or consent
• any incident or event notification term and its timing
• any term that mentions AI, automation, models or automated decision-making
• the next contracting point: renewal, amendment or review date

Q4, location. Triodian records where each party is incorporated, where the service is performed and where the AI runs. Where any of these is not known from the documents, the entry says so.

Q5, authority. Three tests, each answered from the source documents:

1. Does a person act on the output, or does the output take effect without one?
2. Does the output move money, change a policy or claim record, or commit the insurer?
3. Does the AI hold credentials to act across steps or systems without a person starting each step?

The tier follows from the answers (section 8.1). Whether that authority was validly delegated under the agreement is counsel's question.

Q6, failure. The classification takes the most serious credible outcome if the AI acted wrongly and went undetected for one reporting period. It is anchored to the insurer's own tolerance for the critical operation in field 5 where one exists.

Q7, evidence. Split into its two halves and kept separate throughout:

• What can be required is counsel's determination and becomes the required rung.
• What can be produced is Triodian's test (section 8.4) and becomes the obtainable rung.

7.3 Stop rules

• Q1 to Q3 all negative. If counsel finds no nexus, no relationship and only passive rights, the arrangement is recorded as F3, no reasonable reach. Q5 and Q6 are still classified so the conclusion is documented against what the AI can do.
• Q5 at productivity with Q6 immaterial. Disclosure is the most the ladder indicates. The obtainability test is limited to whether disclosure exists.
• Facts unavailable. Where the documents do not show whether AI is used at a provider or platform, the entry is recorded as F4, not determinable on documents, with the specific question that would resolve it.

8. Classification and the evidence ladder

8.1 Authority tiers (question 5)

Tier definitions are from the counsel briefing. The insurance examples are illustrative and are replaced with the client's own in each engagement.

Code	Tier	Definition	Deciding test	Illustrative insurance example
A1	Productivity	Drafts, summarises, searches. A person still acts	A person writes or sends the final output	A claims handler uses a copilot to summarise a file
A2	Analytical	Scores, ranks, recommends. A person decides	The output is an input to a named person's decision	A fraud score shown to an assessor
A3	Decision	Approves, declines, triages. The output is the decision	No person reviews before the outcome takes effect	Automatic triage of a claim to fast-track or investigation
A4	Transactional	Prices, allocates, pays. Money or positions move	The output commits or moves money	Automated pricing under a binder, or straight-through claim payment
A5	Agentic	Plans and acts across steps and tools, unattended	The AI starts its own next step	An agent that gathers documents, contacts a repairer and books an assessment
A6	Actuation	Holds credentials that change systems, assets or infrastructure	The AI can write to systems of record under its own credentials	An agent with write access to the policy administration system

Where a system fits two tiers, the higher one is recorded and the reason noted.

8.2 Failure consequence (question 6)

Code	Consequence	Meaning for the assessment
C1	Immaterial	No effect outside the team using it
C2	Contract breach	The provider or insurer would be in breach of an agreement term
C3	Financial loss	Loss to the insurer above its reporting threshold for the operation
C4	Policyholder harm	A policyholder is wrongly declined, delayed, charged or paid
C5	Critical-operation disruption	A critical operation moves outside the tolerance the board has set

8.3 The evidence ladder (question 7)

The two decks use slightly different names for three rungs. Both are shown; the first is used in this document.

Code	Rung	Also called	What the insurer holds	Produced by	Frequency
E0	None		Nothing	n/a	n/a
E1	Disclosure		A statement of what AI is used, where and for what	The provider	On change, or annually
E2	Attestation		A signed statement that the AI operated within stated limits over a period	The provider	Per reporting period
E3	Event notification	Trigger reporting	Notice when a defined trigger is crossed	The provider	On event, within an agreed time
E4	Audit		A right to examine the aggregate record, exercised by the insurer or its verifier	The insurer or an independent party	On demand or scheduled
E5	Continuous telemetry	Continuous evidence	A stream of aggregate measures against the declared appetite	The system	Each window
E6	Deterministic actuation control	Deterministic control	A control at the point of action, independent of the AI's reasoning, with a signed record of what was requested, what applied and what was allowed	The system	Each action, reported in aggregate

Indicative rung. The base rule pairs each authority tier with one rung.

Authority tier	A1	A2	A3	A4	A5	A6
Indicative rung	E1	E2	E3	E4	E5	E6

Adjustment for failure consequence (proposed, for confirmation). The decks set the base rule only. This document proposes two adjustments so that question 6 affects the result:

• C1 with A1 or A2: the indicative rung drops by one.
• C4 or C5 with A1 to A4: the indicative rung rises by one.

Every adjustment is recorded with its reason. The indicative rung is a technical starting point. Which rung applies is a determination for the insurer.

8.4 Obtainability test

The test starts at the indicative rung and steps down one rung at a time until all five conditions pass. That rung is the obtainable rung.

#	Condition	Question	Source
T1	Produced	Does a record of this type exist, or could the system generate it without redesign?	D7, D8, D9, supplier documentation, interviews
T2	Aggregate	Can it be given without individual decisions, policyholder data or the model?	Nature of the record
T3	Reachable	Is there a route for it to reach the insurer: a term in force, or a contracting point where one could be added?	Field 14, next renewal date
T4	Checkable	Can someone independent of the party producing it verify it?	Field 20
T5	Timely	Can it arrive in time for the insurer to act and meet its own notification obligations?	Agreement timing terms, insurer's incident process

T3 records whether a term exists and what it says. Whether the term is sufficient or enforceable is a question to counsel.

Confidence. High where the record has been seen. Medium where the agreement or supplier documentation describes it. Low where the rung is inferred from the type of system.

8.5 Worked illustration

Illustrative only, not drawn from any client.

Field	Entry
Arrangement	Outside claims manager under a claims agreement (B1)
What the AI does	Triage model routes each new motor claim to fast-track payment or investigation. No person reviews the routing
Authority tier	A3, decision
Failure consequence	C4, policyholder harm
Indicative rung	E3 by base rule, raised to E4 by the proposed adjustment
Evidence received today	E1: a due diligence response naming the triage tool
Obtainability	E4 fails T3 (audit clause reaches financial records only). E3 fails T1 (no trigger defined). E2 passes all five
Obtainable rung	E2, medium confidence
Question to counsel	Does the audit clause reach an aggregate record of automated triage outcomes? What can the insurer require at renewal?

9. Findings

9.1 Finding types

One finding is recorded per arrangement, after counsel has set the required rung. The deck names the first three outcomes. F4 is proposed here because no provider is approached.

Code	Finding	Condition	What is recorded
F1	Evidence obtainable	Obtainable rung is at or above the required rung	The evidence, who produces it, and how it is checked
F2	Gap	Obtainable rung is below the required rung	The gap in rungs, which test failed, and the gap owner
F3	No reasonable reach	Counsel determines the insurer cannot reasonably require evidence at this boundary	Counsel's determination and its basis, with the technical classification beside it
F4	Not determinable on documents (proposed)	The documents do not show whether AI is used, or what it can do	The specific question that would resolve it, and who could answer it

9.2 Gap rating (proposed, for confirmation)

Gap size is the required rung minus the obtainable rung. Priority combines gap size with failure consequence.

Gap size	C1 or C2	C3	C4 or C5
1 rung	Low	Low	Medium
2 rungs	Low	Medium	High
3 or more rungs	Medium	High	High

A gap at low confidence (field 22) is reported at its rated priority with the confidence shown beside it. It is not downgraded.

9.3 Gap owner and closing route

Each F2 names one owner. The assessment records the routes available; who closes a gap, and with what, is a separate decision for the insurer.

Failed test	Usual owner	Routes available
T1, produced	The provider, or the insurer's AI team at B3	The provider's own solution · a tool the insurer already has · a boundary control and evidence design, if the insurer asks
T2, aggregate	The provider	Redesign of the record so it carries aggregates and references only
T3, reachable	The GC, through counsel	Evidence terms at the next contracting point, with a technical schedule beside the clauses
T4, checkable	The insurer's second or third line	Appointment of an independent verifier; verification terms in the agreement
T5, timely	The arrangement owner	Notification timing aligned to the insurer's own windows

9.4 Finding worksheet

One worksheet per arrangement. It is the working paper the independent reviewer checks and the source for the annex.

Part	Contents
A. Identification	Arrangement ID, parties, boundary types, reason selected, worksheet version, preparer, reviewer
B. Chain map	Each party, what it does, where it is, and each place AI acts, with register entry IDs
C. Facts	Register fields 4 to 14 for each entry, each with its source reference
D. Classification	Authority tier and failure consequence, each with a two-line rationale tied to the facts in part C
E. Ladder	Indicative rung, any adjustment and reason, evidence received today
F. Obtainability	Result of T1 to T5 at each rung tested, the obtainable rung and confidence
G. Questions to counsel	Each question, date sent, response, date received
H. Counsel's determinations	Fields 23 to 25, transcribed with question references
I. Finding	Finding type, gap size and priority, failed test, gap owner, routes available
J. Not known	Every field recorded as not known, and what would resolve it

9.5 Question-to-counsel template

Field	Contents
Reference	QC-01 onward
Arrangement and entry	ARR and REG IDs
Which of the seven questions	1 to 4, or 7
Facts	What the documents show, with source references. No characterisation
Question	One question, answerable with a determination
Why it matters to the assessment	Which register field and which test depend on it
Response	Counsel's wording, unedited
Status	Open · answered · open with owner

10. Deliverables

10.1 What Triodian produces

#	Deliverable	Contents	Accepted when
DL1	AI dependency register, including AI inside providers	All 26 fields for every entry in the set. Source index. Count of not-known fields	Every entry has a finding or is recorded F4. No field is blank
DL2	Evidence obtainable at each boundary	One finding worksheet per arrangement (section 9.4). A one-page summary table: arrangement, tier, consequence, indicative, required, obtainable, finding, priority	Each worksheet is reviewed, and each rung is traceable to a test result
DL3	Control-library shape for the insurer's own AI boundary	For B3 entries only: the categories of rule the insurer would need to state for its own staff's use, and the record each would produce. Shape only, with no product specified	Reviewed by the insurer's AI governance lead for fit with the existing framework
DL4	Technical schedule, tolerance and trigger design	For each F1 and F2: the evidence item, its contents, its frequency, the tolerance it is measured against, the triggers and their timing. Written to sit beside NRF's flow-down clauses	NRF confirms each item maps to a clause or to a stated reason for none
DL5	Technical annex, in Triodian's name	Method, set, summary table, findings, limitations, not-known count, statement of what was not done	The independent reviewer's check is complete and its points are closed or disclosed
DL6	Hand-back list	AI uses found that are not on the insurer's inventory	Accepted by the inventory owner

10.2 What NRF produces

Listed so the interfaces are clear. These are not Triodian deliverables.

Deliverable	Depends on from Triodian
Reach determination for each arrangement type	Entity and service maps, extracted agreement terms
Grading: requirement, expectation or better view	Indicative rungs and the facts behind them
FAR reasonable-steps note and privilege position	Findings summary
Flow-down clause schedule	Technical schedule (DL4)
Board paper, in the firm's name	Technical annex (DL5)

10.3 Technical annex outline

1. Purpose, and the question the assessment answers
2. What was and was not done: no data, no provider involvement, no production access, no legal advice
3. The set, and why each arrangement was selected
4. Method: register, seven questions, ladder, obtainability test
5. Summary table of findings
6. Finding by arrangement
7. Patterns across the set: which test fails most often, and at which boundary
8. Routes available for each gap, without recommendation of any product
9. Limitations and confidence, including the not-known count
10. Appendices: register extract, question log, source index

10.4 Language rules for every deliverable

• State what the documents show and what the classification is. Do not state what the law requires.
• Use "indicative" for Triodian's rung and "required" only for counsel's.
• Use "gap" only where counsel has set a required rung above the obtainable one.
• Name no product, including Triodian's.
• Report aggregates and references only. No decision, policyholder detail or model content appears anywhere.

11. Quality control

11.1 Review points

#	Review	When	Reviewer	Tests
R1	Set review	End of phase 1	NRF and GC	Selection criteria applied; each boundary type represented
R2	Register completeness	End of phase 2	Delivery manager	Fields 1 to 14 complete or not known; every value sourced and dated
R3	Classification review	During phase 3	A second Triodian reviewer, not the classifier	Tier re-derived from fields 11 to 13 without sight of the original; differences resolved and logged
R4	Validation	End of phase 3	Sam Ferraro	Rules in sections 8.3 and 8.4 applied consistently across entries; every adjustment has a reason; every obtainable rung has five test results
R5	Legal-line check	Before any draft leaves Triodian	Method owner	No sentence states what the law requires; "required" and "gap" used only as section 10.4 allows
R6	Counsel review	Phase 5	NRF	Draft seen by counsel first; determinations transcribed accurately
R7	Independent check	Phase 6	Independent reviewer	Is the assessment sound, and are its findings supported?

11.2 Evidence standards

• Traceable. A reader can go from any finding to the test results, to the register fields, to the source document and page.
• Reproducible. A second classifier given the same part C facts reaches the same tier and rung. R3 tests this on every entry, since the set is small.
• Dated. Every source carries the date it was produced and the date it was read.
• Verbatim where it matters. Agreement terms and counsel's determinations are quoted, not paraphrased.
• Disagreements kept. Where classifier and reviewer differed, both positions and the resolution stay on the worksheet.

11.3 Working papers

• Marked, stored and circulated as NRF directs under the privilege position it determines.
• Versioned. Each worksheet shows preparer, reviewer and date for every version.
• No policyholder, claimant or staff personal data is collected. If a document supplied contains it, it is returned and a redacted copy requested.
• No provider confidential material is sought. Documents are supplied by the insurer from what it already holds.
• Retained or destroyed at close as NRF directs.

11.4 Exceptions log

Any departure from this framework is logged with the reason and who approved it: a substituted arrangement, an interview not held, a document not supplied, a review point skipped. The log is an appendix to the annex.

11.5 Measures reported at close

Measure	Purpose
Entries by boundary type and authority tier	Shows what the set covered
Fields not known, by arrangement	Shows how much rests on missing facts
Findings by type and priority	The headline result
Obtainable rungs by confidence level	Shows how firm the result is
Failed tests by type and boundary	Shows where evidence breaks down in the chain
Classifier and reviewer differences	Shows how repeatable the classification was
Questions to counsel: sent, answered, open	Shows what remains undetermined

12. Assumptions, dependencies and open items

12.1 Assumptions

• The set is four arrangements. A larger set extends phases 2 and 3 roughly in proportion.
• The insurer supplies documents D1 to D6 in week 1. Late documents move the plan day for day.
• NRF answers questions to counsel within five working days.
• The insurer has a current AI inventory. If it has none, phase 2 starts from the material service provider register and takes longer.
• The independent reviewer's check fits inside week 4. Its scope and criteria are not defined in either deck and are to be agreed with the CRO.

12.2 Dependencies

Dependency	Owner	Needed by
Signed brief and privilege position	GC and NRF	Before week 1
Agreed set	GC	Week 1, day 2
Documents D1 to D10	Insurer, through NRF	Week 1
Interview access	Insurer AI and risk teams	Weeks 1 to 2
Board appetite and tolerances for the operations in the set	Insurer operational risk	Week 2
Counsel's determinations	NRF	Week 3
Independent check	Independent reviewer, on behalf of the CRO	Week 4

12.3 Drafted here, to be confirmed

These elements are not in either deck. Each was added to make the method executable and should be confirmed before first use.

• The 26 register fields. The deck names eight things an entry records. Confirm the breakdown against the register schema Triodian already holds.
• The consequence adjustment to the indicative rung (section 8.3).
• The five-part obtainability test (section 8.4). The deck states the test as two questions: can the evidence be produced, and can someone independent check it?
• Finding type F4, not determinable on documents (section 9.1).
• The gap priority matrix (section 9.2).
• The week-by-week plan, document list and interview list (section 5).
• Rung definitions (section 8.3). The decks name the rungs but do not define them.
• The five-day turnaround for counsel and the default set of four.

12.4 For counsel

• CPS 230 paragraph references. Paragraphs 5, 53(b), 54 and 60(b) are cited as they appear in the decks. Confirm each against the instrument in force.
• Who determines authority under question 5. The counsel briefing gives authority to counsel and consequence to Triodian. This document has Triodian classify the tier and counsel determine whether the authority was delegated. Confirm the split.
• Whether T3 stays on the technical side of the line. It records that a term exists and what it says. Confirm that this does not stray into interpretation.
• Document marking and circulation under the privilege position.

12.5 Other open items

• Ladder naming. The two decks differ on three rung names. Choose one set.
• Insurer terms. The counsel briefing was written for a superannuation trustee. "Mandate breach" and "member harm" have been changed to "contract breach" and "policyholder harm" to match the NRF deck.
• Control-library shape (DL3). The deck names it without describing its contents. The description in section 10.1 is a minimal reading.
• Fee and effort. Not covered here. The counsel briefing priced the equivalent assessment for a trustee at A$35k fixed for three to four weeks; no figure for insurers appears in the NRF deck.
• KJR's role. Described as a later scaling partner for the register. No task is assigned to it in this framework.

Sources

• NRF planning deck: "What a CRO will ask, and how we answer it together", Damian Hickey, October 2026, 24 slides.
• "Triodian - AI Counsel Briefing", 12 September 2026, 29 slides.

No external source was consulted. Regulatory references are as stated in those two documents.