2026 Outlook Enterprise Legal · Risk · Compliance AI Transformation

The function that
both uses AI and
must police it.

Legal, Risk, and Compliance face a tension no other corporate function carries: they are simultaneously adopting AI across their own work and serving as the enterprise's control authority over everyone else's AI. The winners treat this dual mandate as the design principle — redesigning workflows, the operating model, and the three lines of defense — not as a compliance afterthought.

75%

of UK financial-services firms
already use AI (BoE/FCA, 2024)

of FS AI use cases are
fully autonomous (BoE/FCA)

34%

claim “complete understanding”
of the AI they use (BoE/FCA)

1,547

court decisions on AI-hallucinated
content (Charlotin, Jun 2026)

12 reference tables · expandable below

Every analytical table from the underlying research report — value chain, workflows, financial crime, risk intelligence, operating model, roles, model risk, legal, red lines, cases, governance, and governing enterprise AI — is available below as a click-to-expand data view. Closed by default to keep the briefing scannable.

01AI Use Cases 02Workflow Redesign 03Financial Crime 04Risk Intelligence 05Operating Model 06Roles & Skills 07Model Risk & Bias 08Legal & Reg 09Red Lines 10Cases 11Governance 12Governing Enterprise AI

01 The Thesis

Three positions this report defends.

LRC functions treating AI as a productivity overlay capture some efficiency but no durable control advantage. The functions pulling ahead recognize that the same discipline that makes them effective adopters — governance, evidence, defensibility — is exactly what the enterprise needs from them as AI's control authority.

The dual mandate is the design principle, not a conflict to manage.

LRC both adopts AI and governs the enterprise's use of it. Treating these as one capability — built on inventory, validation, audit trails, and human-in-the-loop thresholds — is what separates leaders from firms that bolt governance on after the fact.

ii.

The model-risk ground shifted in April 2026.

The interagency Revised Guidance (OCC 2026-13; Fed SR 26-2; FDIC FIL-15-2026) superseded SR 11-7 and SR 21-8 — and expressly placed generative and agentic AI outside its scope. The tools LRC worries about most now sit in a governance gap LRC must close with its own frameworks.

iii.

The professional-responsibility and disclosure floor is rising fast.

ABA Opinion 512, a court record of 1,547 hallucination decisions, and SEC AI-washing enforcement scoped to registrants and RIAs have redrawn the red lines. Verification, privilege preservation, and disclosure-committee review are now non-negotiable.

02 Adoption vs. Understanding

75% are using AI. 34% understand it. 2% run it autonomously.

The most important numbers in the Bank of England/FCA survey aren't the adoption rates — they're the gaps. Most financial-services firms have deployed AI faster than they can explain it, and control-function adoption deliberately lags the business side because the same function setting the guardrails is cautious about deploying the riskiest tools on itself.

The FS AI understanding gap

BoE/FCA · n=118 firms · Nov 2024

Using AI in some formup from 58% in 2022

75%

Use cases with some automated decision-makingbut a human remains in or over the loop

55%

Firms with “complete understanding” of their AIa further 46% report only “partial” understanding

34%

Use cases that are fully autonomousthe frontier remains overwhelmingly human-supervised

Adopter signal

88%

of organizations use AI in at least one function (McKinsey State of AI 2025, n=1,993) — but only 7% report AI fully scaled across the enterprise (Exhibit 1). LRC adoption is real but early, and control functions trail the business.

Governor signal

1,547

court decisions worldwide now address AI-hallucinated content (Charlotin database, Jun 6 2026) — up from 712 when Bloomberg Law reported it in December 2025. The trajectory, not the count, is the board-relevant point: it more than doubled in six months.

03 Workflow Redesign

Four levels of AI in legal, risk & compliance work.

The four are not interchangeable. Most enterprises stall at level two — adding AI to existing review workflows — and never restructure the work itself, the decision rights, or the lines of defense around it. The most valuable and riskiest shift is reinvention: moving from “review everything” to risk-sampled, near-comprehensive AI triage.

Task automation

A discrete step gets faster. AI extracts contract clauses or summarizes a regulation; the workflow is unchanged.

Workflow augmentation

Human + AI in the same process. Reviewers triage alerts, draft memos, and redline contracts with AI assistance.

III

Workflow reinvention

The process logic changes: sampling becomes near-comprehensive coverage; downstream review becomes upstream guardrail. Sequence, handoffs, and decision rights all shift.

Autonomous LRC

Agents perform defined activities — first-level alert disposition, triage — within control boundaries, with exception handling, audit trails, and approval thresholds. New roles (AI control owner, audit-of-AI specialist) appear that did not exist before.

04 Enterprise Cases

Four reference points — what scaled deployment actually looks like.

Outcome metrics below are company-, vendor-, or court-reported and not independently audited. Treat them as directional evidence rather than benchmarks for your own business case. Two are adoption stories; two are cautionary tales about the governor's side of the mandate.

HSBC

Banking

“Dynamic Risk Assessment” AML AI built with Google Cloud · monitors >1bn transactions/month · Celent Model Risk Manager of the Year 2023

2–4×

more suspicious activity identified vs. prior rules-based system

60%+

reduction in alert volumes — less time chasing false leads

~4×

more financial crime found in retail banking; ~2× in commercial

Company-reported (HSBC/Google), not independently audited · alert-volume reduction is only beneficial if false negatives do not rise · explainability and human disposition thresholds remain essential

JPMorgan Chase

Banking

COiN (Contract Intelligence), live since June 2016 · automates review of commercial credit agreements · LLM Suite rolled out to 200,000+ staff

12,000

commercial credit agreements reviewed per year

360,000

lawyer/loan-officer hours/yr the manual process previously consumed

200k+

staff with access to the firmwide LLM Suite assistant

12,000/360,000 figures from Bloomberg (Hugh Son, Feb 27 2017), company-reported & unaudited · the widely repeated “~150 attributes/contract” detail traces to secondary write-ups (HBS case study), not the Bloomberg primary source

Delphia & Global Predictions

Asset/Wealth · RIAs

SEC's first AI-washing enforcement actions (Mar 18, 2024) · settled charges for false/misleading statements about AI use · charged under Advisers Act §§206(2)/206(4), Marketing Rule, Compliance Rule

$225k

civil penalty — Delphia (USA) Inc.

$175k

civil penalty — Global Predictions, Inc.

$400k

total — “say what you do, do what you say”

Primary source (SEC PR 2024-36) · scoped to registered investment advisers, not universal coverage · the LRC lesson: disclosure-committee review of any external AI claim

Arup

Engineering

Victim of a deepfake-CFO video-conference fraud, Hong Kong, January 2024 · an employee was induced to make 15 wire transfers after a call populated with AI-cloned executives

~$25.6M

(HK$200M) lost across 15 transfers to five accounts

fraudulent transfers triggered by one deepfake call

Media-reported (CNN/Fortune/WEF, 2024) · cross-industry warning · the threat LRC must govern: voice cloning and deepfakes for payment approvals and executive instructions are now an explicit red line

05 Maturity Model

Where most LRC functions actually sit.

Most large financial-services enterprises sit between Level 2 and Level 3 for both LRC adoption and their governance of enterprise AI — pilots running, an AI inventory emerging, the operating model and three-lines design largely untouched. The BoE/FCA finding that only 2% of FS AI use cases are fully autonomous confirms Level 5 is still aspirational.

Level 01

Ad-hoc experimentation

Individuals use public AI tools. No enterprise AI policy. High shadow-AI risk.

Level 02

Functional pilots

Teams pilot contract review, legal research, monitoring, reg-change tracking. First-draft AI policy. Training nascent.

Level 03 · Most enterprises

Scaled use cases with control

AI embedded in repeatable workflows. Approved tools, model governance, audit trails, measurable gains. Enterprise AI inventory and risk-tiering exist.

Level 04

Workflow & operating-model redesign

AI reshapes monitoring, contracting, investigations, decision rights, controls, and the three-lines model. LRC actively governs enterprise-wide AI.

Level 05

Autonomous & decision-intelligent LRC

Agents, controls, data products, and professionals work together for continuous monitoring, near-real-time risk sensing, and defensible automated dispositions within bounds — with regulator-ready evidence.

Self-diagnostic: if your organization cannot point to a workflow that has been structurally rebuilt around AI — not just augmented — and to a complete AI inventory that captures the generative and agentic tools now outside the scope of the revised model-risk guidance, you are at Level 2–3 regardless of tool count. The governor's side of the mandate matures more slowly than the adopter's, and that gap is itself a supervisory finding waiting to happen.

06 AI Red Lines

Four tiers of LRC AI permission.

A defensible enterprise LRC AI policy distinguishes between routine, enhanced, approval-gated, and prohibited use. The prohibited tier is where the function's professional accountability is non-delegable — and where 2023-era policies are most often silent.

01Normal review

Allowed with standard QA

Summarizing public documents
Internal research first-drafts
First-pass alert triage with human disposition

02Enhanced review

Allowed with heightened scrutiny

Contract drafting & redlining
Regulatory-change mapping
Reg-reporting QA

03Approval-gated

Only with GC / CRO / CCO / privacy / audit sign-off

Anything touching privileged / MNPI / customer data
Training on internal data
Customer-impacting models

04Prohibited

No use, no exception

Final legal advice without attorney sign-off
Filing AI citations unverified
Autonomous AML/sanctions alert closure
Autonomous customer-impacting decisions (closure, denial, de-risking)
SAR/STR filings without human accountability
Voice cloning/deepfakes for approvals
Same model performing & testing a control

07 Regulatory Floor

The 2024–2026 record that redrew the floor.

The governing frameworks moved faster than most LRC policies. The single most consequential change for financial-services risk teams arrived in April 2026 — and it leaves the enterprise's most novel AI without a default prudential model-risk regime.

Feb 2025EU

EU AI Act

Prohibited practices & AI-literacy obligation enter application

Article 5 (banned uses) and Article 4 (staff AI literacy) became applicable on 2 February 2025. GPAI model obligations followed on 2 August 2025.

Jul 2024US

ABA

Formal Opinion 512 on generative AI

Maps existing Model Rules to GenAI: competence, confidentiality, candor to tribunals, supervision (including over third-party AI), and reasonable fees. Lawyers may not abdicate professional judgment to AI.

Mar 2024US

SEC

First AI-washing enforcement

Settled charges against two registered investment advisers (Delphia, Global Predictions) for misleading AI claims; $400k in total penalties. Scoped to registrants/RIAs under the Advisers Act.

Apr 17 2026US

OCC · Fed · FDIC

Revised Guidance on Model Risk Management supersedes SR 11-7 & SR 21-8

OCC Bulletin 2026-13 / Fed SR 26-2 / FDIC FIL-15-2026. Principles-based, most relevant above $30B assets, non-compliance not itself a supervisory criticism. Critically, it expressly excludes generative and agentic AI from scope — those tools must be governed under existing frameworks pending a forthcoming interagency RFI.

May 7 2026EU

EU Digital Omnibus

Provisional agreement to defer high-risk obligations

A provisional political agreement would move Annex III high-risk obligations to 2 Dec 2027 and Annex I to 2 Aug 2028. Not yet formally adopted — until Official Journal publication, 2 August 2026 remains the binding date. Plan two parallel tracks.

08 12–24 Month Plan

From pilots to governed transformation.

A practical sequence for a large financial-services LRC function moving from AI pilots to AI-enabled workflow, operating-model, control, and enterprise decision transformation — covering both its own adoption and its stewardship of enterprise-wide AI governance. The discipline is govern, inventory, assess, design, implement, monitor.

Months 0–6

Govern & Inventory

Charter an enterprise AI governance committee. Clear GC / CRO / CCO / CPO / CISO / CDO ownership. Publish an AI acceptable-use policy and approved-tools allowlist. Mandate AI-literacy training (EU AI Act Art. 4 baseline).

Build a complete AI/model inventory. Include LLMs, agentic tools, and the generative/agentic systems now outside the scope of the revised interagency model-risk guidance. Close the gap examiners are finding on un-inventoried LLMs.

Issue the red-lines policy and a citation-verification mandate. Brief the board risk and audit committees on the dual mandate and the model-risk scope change.

Months 6–12

Assess & Design

Risk-tier every use case. Against EU AI Act Annex III, the revised model-risk materiality logic, and NIST AI RMF. Prioritize high-risk: credit/insurance models, AML, customer-impacting decisions. Explicitly extend model-risk frameworks to GenAI/agentic tools the guidance leaves out of scope.

Design the controls. Human-approval thresholds, segregation-of-duties rules preventing the same model from performing and testing a control, audit logging, exception management.

Pilot 2–3 well-bounded use cases. Contract extraction, reg-change mapping, AML alert triage — each with defined metrics and human-in-the-loop.

Months 12–18

Implement & Monitor

Scale validated pilots into production. With model monitoring (drift/bias/accuracy), evidence retention, and regulator-ready documentation.

Stand up model-risk-for-AI and audit-of-AI capabilities. Ensure third-line independence: where internal audit uses AI, independent assurance of that use comes from elsewhere.

Implement disclosure-committee review of external AI claims. The AI-washing defense — say what you do, do what you say.

Months 18–24

Optimize & Operating Model

Redesign the operating model where evidence supports it. Surveillance hubs, legal/compliance product teams, human-AI exception management. Address junior-talent development explicitly.

Establish continuous-control and assurance automation. Measure OUTCOMES and DEFENSIBILITY, not just usage and throughput.

Watch the triggers. If false-negative rates rise, tighten thresholds and pause expansion. If the Digital Omnibus is adopted, lock in the deferred dates. If the agencies issue the AI/GenAI RFI, re-open the model-risk section.

12 Reference Tables · Expandable

The full evidence, on demand.

Every analytical table from the underlying research is below as a click-to-expand data view. The Roles & skills table is open by default — it's the most relevant evidence for individual contributors on how AI augments specific legal, risk, and compliance roles. The other eleven are closed to keep the briefing scannable; open one, several, or all at once.

Sixteen use cases across Legal, Risk, and Compliance, mapped by AI type, workflow impact, maturity, value, risk, and example deployment. The most mature are contract review, legal research, document discovery, and transaction/communications surveillance; the most experimental are autonomous regulatory interpretation and agentic investigations.

Domain	Use case	AI type	Workflow impact	Maturity	Business value	Risk	Control requirement	Example/source
Legal	Contract review & data extraction	Generative/NLP	Augment→reinvent	Scaled	Cycle-time, risk visibility	Med	Human sign-off on judgment terms	JPMorgan COiN [company]
Legal	Legal research & memo drafting	Generative (RAG)	Augment	Scaled	Hours saved	Med-High	Citation verification	TR CoCounsel; Magesh et al. 2025
Legal	eDiscovery / TAR	Predictive ML	Reinvent	Mature	Cost, defensibility	Med	TAR protocol, sampling	Da Silva Moore (2012)
Legal	Matter intake & triage	Generative	Augment	Pilot-scaled	Routing speed	Low-Med	Routing audit log	Gartner top-6
Legal	M&A due diligence / entity mgmt	Generative/agentic	Augment	Pilot	Deal speed	Med	Human review of findings	PwC/Harvey [vendor]
Risk	Credit/counterparty memos, early warning	Generative	Augment	Pilot	Analyst time	High	Independent validation (Revised MRM Guidance)	McKinsey/IACPM
Risk	Model risk mgmt for AI	Analytics	Reinvent	Emerging	Coverage	High	GenAI/agentic outside MRM scope	Revised Guidance 2026 [primary]
Risk	Emerging-risk horizon scanning	Generative	Augment	Pilot	Foresight	Med	Source traceability	Gartner
Compliance	AML transaction monitoring	Predictive ML	Reinvent	Production	FP reduction, detection	High	Human disposition thresholds, explainability	HSBC/Google DRA [company]
Compliance	Sanctions / name screening	ML	Augment	Production	FP reduction	High	Bias testing, human review	FATF guidance [primary]
Compliance	Trade / comms surveillance	ML/NLP	Reinvent	Production	Coverage	High	Lexicon governance, QA review	BoE/FCA
Compliance	SAR/STR narrative drafting	Generative	Augment	Pilot	Drafting time	High	Human accountability sign-off	FinCEN [primary]
Compliance	Regulatory change management	Generative	Augment	Pilot-scaled	Lead time	Med	Lawyer review of interpretation	Gartner
Compliance	Policy drafting & attestation	Generative	Augment	Pilot	Throughput	Med	Qualified review	McKinsey RegTech
Privacy	DSAR handling, ROPA/PIA	Generative/NLP	Augment	Pilot	Turnaround	Med	PII handling controls	IAPP
Cross	Third-party / vendor risk extraction	Generative	Augment	Pilot-scaled	Throughput	Med	Concentration-risk review	BoE/FCA

How AI changes representative LRC workflows — not just individual tasks. The consistent pattern: AI shifts effort from production (drafting, searching, first-pass review) to verification and judgment (approval, exception handling, defensibility). New approval gates and audit trails are the price of scaling.

Workflow	Current-state pain	AI-enabled redesign	Human role	AI role	New controls	Defensibility/audit evidence	Metrics
Contract review	Manual clause-by-clause	Auto-extract attributes, flag deviations	Approve non-standard terms, negotiate	Extract, classify, redline-suggest	Playbook governance, approval gate	Extraction logs, version history	Turnaround time, % auto-handled
Legal research	Hours of manual search	RAG-grounded answers with citations	Verify every citation, exercise judgment	Draft, summarize, cite	Mandatory citation check	Verification checklist	Hours saved, error rate
eDiscovery	Linear manual review	TAR/CAL prioritization	Train, sample, certify	Predictive coding	TAR protocol, QC sampling	Seed-set logs, recall/precision	Recall, cost/GB
AML monitoring	Rules engine, ~90% FP	ML risk-scoring + segmentation	Disposition above threshold, SAR decision	Score, triage, cluster	Disposition thresholds, explainability	Model docs, alert audit trail	FP rate, SAR conversion, detection
Reg change mgmt	Manual horizon scan	AI summarizes/maps to obligations	Validate interpretation, decide actions	Scan, summarize, map	Lawyer review of interpretation	Mapping audit, sign-off	Implementation lead time
SAR/STR drafting	Manual narrative	AI drafts from case data	Review, attest, file	Draft narrative	Human accountability sign-off	Draft-vs-final diff, reviewer ID	Quality score, cycle time

Financial crime is where AI in LRC has the most mature, quantified production deployments — and the highest accountability stakes. Reducing alert volume is only beneficial if false negatives do not rise; this is where supervisory scrutiny and human-in-the-loop thresholds concentrate.

Process stage	Pre-AI process	AI-enabled process	Human judgment required	Regulatory/supervisory implication	Defensibility requirement	Metrics
Transaction monitoring	Rules engine, high FP	ML risk-scoring/segmentation	Disposition above threshold	Revised MRM Guidance 2026 (GenAI out of scope); FinCEN/FATF	Model docs, drift testing, audit trail	FP rate, SAR conversion, detection
Sanctions screening	Fuzzy match, high FP	ML name-matching	All true-hit decisions	OFAC strict liability	Bias testing, match logs	FP rate, hit accuracy
KYC/CDD	Periodic review	Perpetual KYC	Risk-rating changes	CDD rule	Decision audit	Refresh cycle time
Trade/comms surveillance	Lexicon sampling	NLP near-full coverage	Escalation review	Market-abuse rules	Lexicon governance, QA	Coverage %, escalation precision
SAR/STR	Manual narrative	AI draft	Review, attest, file	FinCEN accountability	Reviewer ID, draft-vs-final	Quality, cycle time

How AI moves Risk from periodic reporting toward continuous, decision-grade intelligence. Predictive analytics is long-standing in FS; the new layers are generative and agentic AI for sensing, synthesis, and narrative. Because GenAI/agentic tools sit outside the revised model-risk guidance, GenAI-based risk tooling does not inherit a validated regime by default.

Risk activity	Pre-AI process	AI-enabled process	Business decision supported	Data requirements	Explainability requirement	Metrics
Emerging-risk scanning	Manual research	AI synthesis of signals	Risk-appetite adjustments	External + internal feeds	Source traceability	Lead time to detection
Early-warning credit	Periodic review	Continuous AI surveillance	Rating review trigger	Financials, market data	Feature attribution	False-alarm rate, lead time
Risk reporting	Manual decks	NL-generated narrative	Board/regulator reporting	Governed metrics	Metric lineage	Cycle time, accuracy
Scenario / stress	Manual design	AI-assisted scenario gen	Capital/limit decisions	Historical scenarios	Assumption transparency	Coverage, plausibility

The traditional model is being supplemented by an AI-enabled one. The central challenge is the three-lines stress test: AI now sits inside first-line processes and inside the second-line tools that test them, risking the second line “marking its own homework.” In FS, regulators require clear first/second-line delineation, so AI agents must not circumvent lines-of-defense separation.

Traditional component	AI-era equivalent	What changes	Capabilities required	Control / lines-of-defense implication	Risks
Legal ops	Legal product/data teams	Workflows productized	Product mgmt, data	First-line process ownership	Shadow AI
Compliance monitoring	AI surveillance hub	Sampling→near-full coverage	ML ops, QA	2nd line may test AI it also uses	Marking own homework
Model risk (stat models)	Model-risk-for-AI	LLM/agent governance (GenAI outside MRM scope)	Explainability, drift	Independent challenge; gap for GenAI/agentic	Validation backlog; un-scoped tools
Internal audit	Audit-of-AI	Tests AI governance	AI assurance skills	3rd-line independence	Auditing own AI tools
Enterprise (none)	AI governance office/committee	New authority	Cross-functional	Owns enterprise AI oversight	Unclear accountability

Open by default — the most relevant evidence for individual contributors on how AI changes specific LRC roles. Deconstructing jobs into tasks reveals which are automated, augmented, retained, or newly created. Compression of entry-level scaled work raises average seniority but threatens the development pipeline for senior judgment.

Role family	Tasks likely automated	Tasks likely augmented	New tasks created	Capabilities increasing in value	Reskilling	Career-path implication
In-house counsel	First-pass review, research	Drafting, analysis	AI output verification, playbook design	Judgment, negotiation	AI fluency, prompting	Higher seniority, fewer juniors
Compliance analyst	Alert triage, screening	Investigation, disposition	AI QA review	Risk judgment	ML literacy	Shift to oversight
Risk analyst	Data gathering, reporting	Early-warning, scenario	Model-risk-for-AI	Interpretation, challenge	Data/model literacy	Business-partner shift
Legal ops	Manual workflow steps	Process design	Legal data product mgmt	Product thinking	Tech/data skills	New product roles
Internal audit	Sample testing	Continuous assurance	Audit-of-AI	Independent skepticism	AI assurance	Specialist tracks

LRC-specific decision risks, anchored in COSO and the NIST AI RMF. The key distinction for control design: false positives (cost, customer friction, alert fatigue) versus false negatives (undetected financial crime, conduct, or legal exposure). Even purpose-built legal AI requires verification — Magesh et al. (2025) found Westlaw and Lexis tools hallucinated ~33% and ~17% of answers in May 2024 testing.

Use case	Risk type	Affected stakeholders	Legal/regulatory/defensibility risk	Detection method	Mitigation control	Monitoring metric
AML monitoring	False negative / over-flagging	Customers, firm	BSA/AML deficiency; disparate impact	Outcome testing, fairness analysis	Human disposition thresholds	FP/FN rate by segment
Sanctions screening	FP with discriminatory effect	Customers	OFAC + fair-treatment	Bias testing	Human review of hits	Match accuracy by name type
Credit/insurance models	Disparate impact	Applicants	Fair lending; EU AI Act Annex III	Disparate-impact testing	Explainability, human review	Approval rates by protected class
Legal research/filings	Hallucination	Court, client	Sanctions; candor (MR 3.3)	Citation verification	Mandatory human check	Error/hallucination rate
Surveillance	Language/style bias	Employees	Conduct, employment	Lexicon review	QA sampling	Escalation precision by group

The frameworks that bind LRC, with financial-services scope stated precisely. EU AI Act high-risk obligations attach to specific Annex III categories (credit scoring, insurance pricing, employment, biometrics) — not all AI. AI-washing enforcement to date is scoped to registrants and RIAs. The April 2026 Revised MRM Guidance excludes generative and agentic AI from scope.

Regulation/standard	LRC implication	High-risk use cases	Required controls	Owner	Evidence source
EU AI Act	Tier + Annex III obligations	Credit scoring, insurance pricing, employment	Risk mgmt, human oversight, docs	GC/AI office	Regulation 2024/1689 [primary]
Revised MRM Guidance 2026	Model governance for non-GenAI AI/ML; GenAI/agentic out of scope	AML, credit, trading	Independent validation, inventory; extend frameworks to GenAI	CRO/model risk	OCC 2026-13; Fed SR 26-2; FDIC FIL-15-2026 [primary]
GDPR Art. 22	Restrict solely-automated decisions	Customer-impacting decisions	Human intervention right	CPO	EDPB; CJEU SCHUFA [primary]
ABA Op. 512	Verify, preserve privilege	Filings, advice	Citation check, confidentiality	GC	ABA [primary]
SEC AI-washing	Accurate AI disclosures	Marketing AI claims	Disclosure-committee review	GC/CCO	SEC PR 2024-36 [primary]
FATF / FinCEN	Explainable AML AI	Transaction monitoring	Explainability, governance	CCO/FCU	FATF 2021; 2018 Joint Statement [primary]
ISO/IEC 42001	Certifiable AIMS	All AI	Management-system controls	AI office	ISO [primary]

Classification into Allowed with normal review, Allowed with enhanced review, Allowed only with GC/CRO/CCO/privacy/audit approval, and Prohibited. The prohibited tier captures the non-delegable and deceptive uses — where the function's professional accountability cannot be transferred to a model.

Use case	Risk type	Why prohibited/restricted	Applicable law/standard	Required policy language	Exception process
Final legal advice / reg interpretation without qualified sign-off	Professional responsibility	Lawyer judgment non-delegable	ABA Op. 512 (MR 1.1)	“No AI output is legal advice without attorney sign-off”	None — always human
Filing submissions with unverified AI citations	Candor, sanctions	Hallucination record	MR 3.3; Mata, Park v. Kim	“All citations independently verified before filing”	None
Autonomous closing of AML/sanctions/fraud/surveillance alerts	False negative, BSA/AML	Regulatory accountability	BSA; FinCEN	“Human disposition required above [threshold]”	Below-threshold auto-clear with sampling
Autonomous customer-impacting decisions (closure, denial, de-risking)	Discrimination, fairness	Disparate impact	GDPR Art. 22; fair lending; EU AI Act Annex III	“Human decision-maker required”	Human-in-the-loop with contest right
Autonomous approval of contracts/waivers/obligations beyond thresholds	Authority, financial	Exceeds delegated authority	Internal authority matrix	“AI may draft, not approve, above [threshold]”	Approval gate
SAR/STR narratives or filings without human accountability	Accountability	Human must attest	FinCEN	“Named human reviewer attests”	None
Drafting/modifying policies or risk appetite without qualified review	Governance	Requires expert judgment	COSO; internal	“Qualified review required”	None
Concealing AI involvement where reliance material	Candor, fraud	Misrepresentation	MR 8.4(c); SEC	“Disclose AI use where material”	None
Privileged/MNPI/customer/employee data in unapproved public tools	Privilege, confidentiality	Waiver/leakage	MR 1.6; GDPR; MNPI rules	“Only approved tools for sensitive data”	Approved-tool allowlist
Training external models on privileged/confidential data without approval	Confidentiality	Irreversible exposure	MR 1.6; data law	“No training on confidential data without sign-off”	Legal/security/governance approval
Voice cloning/deepfakes for approvals/communications	Fraud	Impersonation	Fraud law	“Prohibited”	None
Same model performing AND independently testing a control	SoD collapse	Lines-of-defense	IIA Three Lines; Revised MRM Guidance	“Independent model/validator required”	None
AI agents overriding case-mgmt/recordkeeping/access controls	Recordkeeping	Audit integrity	SEC/FINRA recordkeeping	“Agents cannot override controls”	None

Eight enterprise reference points with consistent fields. Outcome figures are labeled by evidence type — company-reported, vendor-reported, primary (enforcement/court), or media. None of the productivity figures are independently audited; treat them as directional.

Company	Industry	AI application	LRC role	Workflow/org redesign	Outcomes	Risks/caveats	Sources	Evidence
HSBC	Banking	AML “Dynamic Risk Assessment” (Google Cloud)	Compliance/FCU adopter	Monitoring reinvented; investigators focus on genuine cases	2–4× more suspicious activity; 60% fewer alerts	Company-reported, not independently audited	Google Cloud/PR Newswire 2023; Celent	[company]
JPMorgan Chase	Banking	COiN contract intelligence; LLM Suite	Legal adopter	Contract review automated; firmwide assistant	~12,000 agreements/yr; ~360,000 hrs displaced; 200k+ staff on LLM Suite	12k/360k from Bloomberg, company-reported; “150 attributes” is secondary	Bloomberg 2017 (primary); HBS (secondary)	[company]
HSBC	Banking	Ayasdi AML (earlier)	Compliance adopter	FP reduction	>20% FP/investigation reduction	Vendor-reported; distinct from DRA	SymphonyAI 2019	[vendor]
Danske Bank	Banking	Teradata fraud ML	Risk/fraud adopter	Real-time fraud scoring	60% FP reduction; +50% detection (80% projected)	Vendor-reported; fraud not AML	Teradata 2017	[vendor]
PwC	Professional services	Harvey AI for legal-entity transformation	Legal-adjacent adopter	Entity lifecycle automated	Faster, more consistent delivery	Vendor alliance source	PwC/Harvey 2025	[vendor]
Arup	Engineering	Victim — deepfake CFO fraud	Cyber/fraud cautionary	—	~$25.6M (HK$200M) lost over 15 transfers, Jan 2024	Cross-industry warning; the threat LRC must govern	CNN/Fortune/WEF 2024	[media]
Delphia / Global Predictions	Asset/wealth (RIAs)	Enforcement — AI-washing	Disclosure cautionary	—	$225k + $175k = $400k penalties	Scoped to RIAs under Advisers Act	SEC PR 2024-36	[primary]
Westlaw / Lexis legal AI	Legal tech	RAG legal research	Legal adopter (caution)	Research augmented	~33% / ~17% hallucination (May 2024 testing)	Independent study; vendors dispute methodology; products since updated	Magesh et al. 2025	[academic]

A finance-grade control checklist adapted for LRC. Ownership spans GC, CRO, CCO, CPO, Head of Financial Crime, Model Risk, Internal/External Audit, CIO, CISO, CDO, Chief AI Officer, HR, Procurement, and the board risk/audit committees. Most FS firms start centralized for gen-AI risk and shift toward federated as capabilities mature.

Risk area	Why it matters in LRC	Required control	Control owner	Evidence/regulation	Monitoring metric
Acceptable use	Prevent shadow AI	AI acceptable-use policy + approved-tools list	CCO/CISO	EU AI Act Art. 4	Policy attestation rate
Data classification	Protect privilege/MNPI	Prompting rules by data class	CPO/CISO	GDPR; MR 1.6	Leakage incidents
AI inventory	Know your models incl. out-of-scope GenAI/agentic	Enterprise AI/model inventory	CDO/model risk	Revised MRM Guidance 2026	% inventory complete
Risk tiering	Proportionate control	Risk assessment per use case	AI risk officer	EU AI Act; NIST	% tiered
SoD / lines separation	Avoid marking own homework	Independent validator/model	CRO/audit	IIA; Revised MRM Guidance	SoD exceptions
Human thresholds	Accountability	Approval gates by materiality	Process owners	FinCEN; GDPR Art. 22	% above-threshold human-reviewed
Audit logging	Defensibility	Prompt/output logs, version history	CISO	Recordkeeping rules	Log completeness
Model monitoring	Drift/bias	Drift/bias/accuracy testing	Model validation	Revised MRM Guidance; NIST	Drift, fairness metrics
Disclosure review	AI-washing	Disclosure-committee review of AI claims	GC/CCO	SEC	Claims reviewed
Incident escalation	Containment	AI-incident response plan	CISO/CCO	NIST Manage	Time-to-escalate
Third/fourth-party	Supply-chain risk	Vendor + sub-processor governance	Procurement/TPRM	EU AI Act GPAI	Concentration metrics
Recertification	Continued fitness	Periodic recert + decommission	Model risk	Revised MRM Guidance	% recertified on time

The table with no analog in a finance dashboard. Because LRC owns enterprise AI governance, this maps the lifecycle — govern, inventory, assess, design, implement, monitor, decommission — to LRC responsibility, owning role, required artifact, and regulator-facing consideration. This is the “policing” half of the dual mandate made operational.

AI lifecycle stage	LRC responsibility	Owning role	Required artifact / evidence	Regulator-facing consideration
Govern	Set AI policy, acceptable use, risk appetite	GC/CCO/CRO + AI committee	AI policy, decision rights	“Convincing narrative” (FCA)
Inventory	Maintain AI/model inventory	Model risk/CDO	Complete inventory incl. LLMs and out-of-scope GenAI/agentic	Revised MRM Guidance inventory expectation
Assess	Risk-tier each use case	AI risk officer	Risk assessment, EU AI Act tiering	Annex III classification
Design	Controls, approval gates, SoD	Control owners	Control design docs	Lines-of-defense separation
Implement	Approved tools, access mgmt	CISO/CIO + LRC	Approved-tools list	Data-leakage controls
Monitor	Drift/bias/accuracy testing	Model validation	Monitoring metrics, logs	Explainability evidence
Decommission	Retire models/agents	Model risk	Decommission record	Audit trail retention

09 · So What

What the GC, CRO & CCO should do differently now.

Stop measuring AI maturity by how many tools the function has adopted. Start measuring it by the question that defines the dual mandate: can you produce, on demand, a complete inventory of the enterprise's AI — including the generative and agentic systems your regulator just placed outside the model-risk perimeter — and the evidence that each is governed?

By 2027, the difference between LRC functions that strengthened the enterprise through the AI transition and those that exposed it will not be access to models — every function will have the same frontier tools. The difference will be defensibility.

That difference rests on three things: (1) whether workflows, decision rights, and the three lines of defense were genuinely redesigned rather than merely augmented; (2) whether the function closed the governance gap the April 2026 model-risk guidance opened, instead of assuming the old rules still cover the new AI; and (3) whether LRC treated its role as the enterprise's AI control authority as a source of advantage rather than a burden to minimize.

The GCs, CROs, and CCOs who will lead are the ones who recognize that the discipline making them credible adopters — inventory, validation, audit trails, human-in-the-loop, and the refusal to delegate professional judgment — is exactly the discipline the enterprise needs from them as AI's governor. Using AI well and policing it well are not two jobs. They are one.

End of briefing · The Dual Mandate · v2 (corrected)

Your future is our focus.™

The function thatboth uses AI andmust police it.