It is a Wednesday afternoon in late Q3 2026. Maya Okafor, Chief AI Risk Officer at Acme, has nine printed memos on her conference-room table, each making a different recommendation for the structural question her board AI subcommittee has asked her to resolve by November 15. The question is short. Where does Acme's AI risk capability sit? Centralized in Maya's AI Risk Office, embedded inside the seven business units, or hybrid? The CEO argues for centralized, independence preservation matters more than speed; central concentration is the only way to recruit. The CTO argues for embedded, the BUs are building the AI; central oversight will become a bottleneck. The seven BU heads each have a different opinion. Manufacturing wants embedded. Customer Success wants centralized. HR wants hybrid with HR-specific Champions reporting through HR. The General Counsel argues for centralized Internal AI Audit and hybrid for everything else. The CFO argues for hybrid with a charge-back model so AI risk is not a free tax on the BUs. Maya reads each memo twice and starts working through a five-dimension scorecard she has been carrying since her first read of the IIA Three Lines Model 2020 update. This lesson is what she walks through, why operating-model design is the Year-1 structural question every CAIRO inherits, why hybrid (federated) is the 2026 mainstream pattern, the seven elements of the hybrid pattern in detail, the migration paths from extremes to hybrid, the 3LoD and RACI cross-walks that determine whether a hybrid can survive ISO 42001 Stage 2, and the seven common 2026 operating-model failures that surface in regulator engagement and audit findings.

Three Operating-Model Archetypes - Centralized, Embedded, Hybrid

The structural question every CAIRO must answer in Year 1, and that every board AI subcommittee should expect on the Q3 or Q4 agenda of the function's first year, is where AI risk capability sits relative to the business. The answer determines who hires, who fires, who reports to whom, who carries cost, who controls policy, who controls tooling, who has the conversation with the deployer team when an Annex IV gap surfaces three days before launch, and who signs the Article 47 declaration. The 2026 industry pattern has converged on three archetypes, with hybrid (federated) as the mainstream choice for enterprises with five or more business units running AI in production.

Centralized - All Capability in the CAIRO's Office

The centralized archetype concentrates all AI risk capability in a single function reporting to the CAIRO. The AI Risk Office (per lessons 074, 087, 099) houses policy, standards, FRIA execution, Independent Model Validation (lesson 068), the Red Team (lessons 081-084), the AI Governance Committee secretariat, the AIMS evidence vault custodianship, and the Article 47 signing workflow. Business units consume services from the central function. They file intake forms (lesson 019), they receive review feedback, they integrate central FRIA findings into deployment, they cannot bypass the gate. The pattern's strength is independence preservation: the AI Risk Office is structurally separated from the build teams, the 2L / 1L distinction (lesson 073) is sharp on the org chart, the function's reporting line runs directly to the board AI subcommittee without business-unit-leader filtration. The pattern's strength is also talent leverage, the scarce 2026 AI risk talent pool (lesson 107) is concentrated where it can be deployed across the portfolio rather than diluted across BUs each running below critical mass.

The pattern's weakness is speed-to-business. Every AI deployment routes through a single central function. Bottlenecks form when the function does not scale linearly with the number of AI systems in flight. The function develops the reputation as the "department of no" rather than the partner that helps deployments ship safely. BU leaders begin shadow-AI workarounds (lesson 021) precisely because central cannot keep up. The 2026 failure pattern is not that centralized governance is wrong in theory. It is that centralized governance without sufficient scaling tips into bottleneck and then into evasion. Centralized is the right pattern at small scale (under 25 AI systems in production, under 1,000 employees) and at maximum-independence scale (financial-sector firms where SR 11-7 independence drives the choice). It is the wrong pattern at enterprise scale unless the central function itself scales to 30+ FTE, which most 2026 organizations cannot fund or hire fast enough.

Embedded - Capability Inside Business Units

The embedded archetype places AI risk specialists inside each business unit, reporting to the business-unit leader. The central function, where it exists, is a small policy / standards / coordination team. Capability sits close to the use cases. The pattern's strength is speed-to-business and cultural fit: the AI risk specialist understands the deployment context, the build team treats them as a partner rather than a gate, the policy interpretations are grounded in business reality, the throughput rises sharply because there is no single bottleneck. The pattern's strength is also business-unit ownership of risk, the 1L (business unit as model owner under lesson 073's Three Lines framing) carries the AI risk explicitly and visibly.

The pattern's structural weakness is independence preservation. An AI risk specialist whose performance review flows through the BU leader cannot easily reject the BU leader's pet project. The 2L / 1L distinction collapses on the org chart. Reporting findings up to central becomes the BU leader's filtered view of what to escalate. Internal AI Audit (lesson 093) cannot legitimately sit in BUs at all, 3L independence is a structural absolute under the IIA Three Lines Model 2020 update; an embedded 3L is a 3L breach. The 2026 failure pattern with pure embedded is that aggregation breaks down: each BU runs its own AI risk practice with its own policies, tooling, definitions of "high risk," FRIA template, bias-eval scorecard. The CAIRO cannot answer the board's question "what is our enterprise AI risk posture" because there is no enterprise view. There are seven incommensurate BU views. ISO 42001 Stage 2 surfaces the gap as a structural finding in Clause 7 support and A.4 organizational structure. The market surveillance authority surfaces it the first time an Article 73 incident has to be aggregated across BUs and the aggregation does not exist.

Hybrid (Federated) - Central Spine + Embedded Champions

The hybrid (federated) archetype is the 2026 mainstream pattern. Capability is split deliberately between a centralized spine that owns policy, standards, governance, the Red Team, Independent Model Validation, Internal AI Audit (with the 3L exception that audit reports to the Audit Committee for independence), and the AIMS evidence vault: and embedded AI Champions that sit inside each business unit at 10-25% time allocation, executing under the central spine's standards, escalating into the spine for non-routine matters, and acting as the day-to-day operational presence of AI risk inside the BU. The central spine preserves independence at the structural level. The embedded layer preserves speed and cultural fit at the operational level. The pattern is designed for enterprises with five or more business units running AI in production at scale where neither pure centralization nor pure embedding can hold.

The 2026 hybrid pattern is not "split everything down the middle." The split is deliberate by function. Policy, standards, FRIA methodology, IMV, Red Team, Internal AI Audit, AIMS vault, AIGC secretariat, central. Use-case-level execution, intake-form facilitation, BU-specific risk register maintenance, deployer-side instructions-for-use review, day-to-day deployment partnership, embedded. Decision rights are tiered. Tooling is shared. The model only works if rotation is real. Anything else is the same dysfunction as embedded with extra coordination overhead.

The Five-Dimension Scorecard for Choosing the Archetype

The 2026 reference scorecard Maya walks through with her board AI subcommittee is five dimensions, each scored 1-5 for the three archetypes against the specific enterprise context. The dimensions are not weighted equally, independence preservation carries the highest weight in financial-sector and Annex III §5 critical-infrastructure contexts; speed-to-business carries higher weight in technology and consumer-product contexts; cultural fit always carries weight because operating models that fight the culture fail regardless of their abstract elegance.

1. Independence preservation. Centralized = 5. Embedded = 1. Hybrid = 4 (the central spine preserves independence; the embedded champions sit in 1L by design, not as 2L breach). The dimension reflects the 2L / 3L structural integrity requirement under the IIA Three Lines Model and the SR 11-7 / EU AI Act QMS independence expectations. Financial-sector firms regulated under SR 11-7 typically score this dimension at the top of their weighting.
2. Speed-to-business. Centralized = 2 (bottleneck-prone). Embedded = 5. Hybrid = 4 (embedded champions execute the day-to-day; central spine handles non-routine escalations). The dimension reflects throughput, how many AI systems can move from intake to deployment per quarter without the function becoming the constraint. Consumer-product and SaaS organizations whose competitive position depends on AI feature velocity weight this dimension highly.
3. Cost efficiency. Centralized = 4 at small scale; 2 at large scale (the central function cannot keep up and shadow-AI proliferation costs more than the function saved). Embedded = 2 at small scale (each BU below critical mass); 3 at large scale (BU-level scale efficiencies). Hybrid = 3 at small scale (overhead of the spine + embedded); 5 at large scale (spine economies + embedded efficiency). The cost-efficiency dimension is the one most often miscalculated in the Year-1 design, the apparent FTE efficiency of pure centralized turns into the bottleneck-then-evasion cost spiral by Year 2.
4. Talent leverage. Centralized = 5 (scarce talent concentrated). Embedded = 2 (talent diluted across BUs, each below critical mass for community of practice). Hybrid = 4 (spine concentrates the scarcest specialties: Red Team, IMV, FRIA methodology; embedded layer is generalist AI Champions who can be hired more broadly). The 2026 AI risk talent market (lesson 107) makes this dimension structurally important, Red Teamers and IMV leads cannot be hired in every BU, period. The hybrid model puts them where they can be deployed.
5. Cultural fit. Highly enterprise-specific. Engineering-led, BU-autonomous cultures (Big Tech, product companies) score embedded = 5, centralized = 2, hybrid = 4. Compliance-led, central-services cultures (banking, insurance, regulated utilities) score centralized = 5, embedded = 1, hybrid = 4. Federated holding-company structures with distinct BU P&Ls score hybrid = 5 reliably. The cultural-fit dimension is what often tips a borderline decision; designing against the culture is a known failure mode regardless of any other scorecard outcome.

The scorecard is not a deterministic algorithm. It is a structured conversation with the board AI subcommittee. The five dimensions are scored and weighted on the specific context; the resulting recommendation is brought as a memo to the board. The scorecard's value is that it surfaces the trade-offs explicitly rather than letting the decision drift on personality. Maya's Acme scorecard scored hybrid 22, centralized 17, embedded 12, with cultural fit weighted highest because Acme is a federated holding-company structure where pure centralized would be culturally fought.

The Hybrid Pattern in Detail - Seven Elements

The hybrid pattern that the 2026 mainstream has converged on is not "central plus some embedded people." It is a deliberate seven-element design where each element supports the others. Skipping any of the seven elements is a known 2026 failure mode. Maya's Acme implementation walks all seven; the board AI subcommittee resolution naming her design records each.

Element 1 - Centralized Functions (CAIRO + AI Risk Office + Red Team + IMV + Internal AI Audit)

The centralized spine houses the functions where independence, scarce talent, and aggregate enterprise view matter most. CAIRO and the AI Risk Office sit central (lessons 074, 087, 099). The Red Team sits central (lessons 081-084) - Red Team capability is too scarce in the 2026 talent market to dilute across BUs, and Red Team independence from the build teams is structural. Independent Model Validation sits central (lesson 068) - IMV is by definition independent from the model-development team. Internal AI Audit (lesson 093) sits central with the critical exception that audit's reporting line runs to the board Audit Committee, not to the CAIRO, preserving 3L independence under the IIA Three Lines Model 2020 update. The AIGC secretariat sits central. The AIMS evidence vault and the master policy / standards repository sit central. Acme's central spine is 18 FTE in 2026 Q3: CAIRO (1), AI Risk Office leadership (3), FRIA / risk-assessment specialists (3), IMV (3), Red Team (3), AI Audit (2: dotted-line to CAIRO, solid-line to Audit Committee), AIGC secretariat / AIMS custodian (2), Policy + Standards (1).

Element 2 - Embedded AI Champions (10-25% Time in Each Business Unit)

Each BU has one or more AI Champions, sitting inside the BU at 10-25% time allocation, reporting solid-line to the BU leader and dotted-line to central. The AI Champion is not a 2L specialist embedded into 1L. They are a 1L role with the responsibility of being the central spine's day-to-day operational interface inside the BU. They run intake-form facilitation. They maintain the BU's AI use inventory (lesson 020). They liaise with central on FRIA execution. They review deployer-side instructions-for-use before publication. They escalate into central for any matter outside the routine. They participate in the community of practice and consume central tooling. Acme has 12 AI Champions across 7 BUs in 2026 Q3 - Manufacturing 2, Customer Success 2, HR 2, Finance 1, Sales 2, R&D 2, IT 1. Each is at 25% time allocation. Under 10% becomes nominal; over 25% loses the BU-context grounding by detaching the Champion from BU work.

Element 3 - Community of Practice (Monthly Practitioner Gathering)

Central spine + embedded champions meet as a community of practice on a fixed monthly cadence. The agenda is reference patterns (one BU presents a recent FRIA, IMV finding, or Red Team result), tooling updates (eval pipeline changes, ATLAS technique additions), policy / standards refresh items (any change surfaces here before going to AIGC), and emerging-issue discussion (jailbreak techniques surfacing in the public literature, regulator-engagement updates, vendor incidents). The community of practice is the operational glue. It is what prevents the AI Champions from becoming siloed inside their BUs and what gives central the visibility into how standards are landing in the field. The 2026 best practice is monthly cadence, two-hour session, recorded for AIMS-vault retention, attendance mandatory for AI Champions and Red Team / IMV / AI Risk Office leads. Acme's runs the first Thursday of each month, 14:00-16:00 CET.

Element 4 - Federated Decision Rights (Three Tiers)

Federated decision rights make the hybrid pattern functional. Without explicit decision rights, every decision drifts toward the spine (paralysis) or toward the BU (independence breach). The 2026 reference pattern is a three-tier matrix:

• Tier 1, central decision. Enterprise policy, enterprise standards, FRIA methodology, IMV pass / fail disposition, Red Team scope and findings, Article 47 declaration signing, AIGC charter changes, AIMS-vault retention rules, vendor-AI master-list additions, prohibited-practice (Article 5) determinations, substantial-modification (Article 43(4)) dispositions. The central spine decides; the BU is informed.
• Tier 2, central decision with BU consultation. Annex III categorization for a specific BU system, FRIA scope and findings for a specific BU deployment, instructions-for-use approval, deployer-side disclosure decisions (Article 50), AI risk tier assignment, exception requests against enterprise policy. The central spine decides after consulting the BU; the BU's input is on record but not binding.
• Tier 3 - BU decision with central audit. Day-to-day deployment within an approved system's envelope, fine-tuning within the pre-determined-change carve-out (Article 43(4)), routine model retraining within the IMV-approved cadence, BU-internal AI literacy training delivery (Article 4), BU-level vendor evaluation (against the enterprise vendor framework, lesson 070). The BU decides; the central spine audits sample disposition through Internal AI Audit annual planning.

The decision-rights matrix is written, retained in the AIMS evidence vault, reviewed annually by AIGC, and surfaced explicitly in any RACI design (see cross-walk below). Ambiguous decision rights are the single most-cited operating-model failure in the 2026 ISO 42001 Stage 2 audit findings literature.

Element 5 - Shared Services (Eval Pipelines, Tooling, Registries, Repos)

Central operates shared services that the embedded champions consume rather than each BU rebuilding. Shared services include: the eval pipeline (lesson 062 - Anthropic Inspect + OpenAI Evals + promptfoo + Garak + PyRIT orchestration); the promptfoo eval registry where reference suites are versioned (lesson 059); the ATLAS technique library catalogued with reference test scenarios (lesson 064); the CycloneDX ML-BoM repository (lesson 029) where every production AI system's ML-BoM is registered; the AIMS evidence vault as a shared service; the Annex IV technical documentation template library (lesson 050); the model card and system card template library (lessons 026-027); the FRIA template library (lessons 045-047). Shared services are the economic-efficiency driver of the hybrid pattern, why hybrid at scale outperforms embedded on cost. Acme's shared-services investment 2026 Q1-Q3 cumulative: $2.4M, returning estimated $5.8M of avoided BU-level rebuilding cost.

Element 6 - Rotation Program (Six-Month Embeds in Both Directions)

The rotation program prevents the central spine from becoming insular and the embedded champions from drifting into BU-local conventions. The 2026 reference design is six-month embeds, in both directions. Central staff rotate into BUs, sitting alongside an AI Champion, learning the BU's deployment reality, then returning with that grounding. AI Champions rotate into central, sitting alongside the FRIA / IMV / Red Team, learning the central methodology, then returning to their BU as more-capable Champions. Acme runs four rotations annually, two central-to-BU, two BU-to-central. The rotation is a real role swap, not a shadowing program. Rotation is also a talent-pipeline mechanism: BU Champions who excel in central rotation are the recruiting pool for central-spine open roles.

Element 7 - Central Charge-Back Model (BU Pays for Services Consumed)

The charge-back model creates economic discipline around central consumption. Without it, central is seen as a "free tax", BUs over-consume because the marginal cost is zero. With charge-back, BUs pay for FRIA executions, IMV cycles, Red Team engagements, AIMS-vault storage, and shared-services usage in proportion to consumption. The charge-back is internal, money does not actually move; the cost is allocated against BU P&L for purposes of the annual operating budget. The mechanism's value is behavioral: BUs prioritize their AI risk activities deliberately; central receives signal on which BUs are heavy users; the CFO has a defensible breakdown of central cost by BU. Acme's charge-back goes live in Q4 2026 - Q1-Q3 was deliberately non-chargeback to build baseline; Q4 onwards each BU receives monthly statements against published unit rates (FRIA execution: $12,000; IMV cycle: $35,000; Red Team day: $4,500; AIMS storage: $0.18/GB/month).

Migration Paths - From the Extremes to Hybrid

Few enterprises start in Year 1 with a clean hybrid design. Most arrive at hybrid through migration from one of two starting positions. The migration paths are themselves operationally important, the wrong migration sequence can make the destination unreachable. The 2026 reference patterns are:

Migration A - From All-Centralized (Y0) to Hybrid (Y1-Y2) to Optimized Hybrid (Y3+)

Enterprises that stood up the CAIRO function in Y0 with a centralized design, typical in financial-sector firms, in firms that adopted ISO 42001 ahead of the EU AI Act deadlines, in firms whose initial AI risk hire came from an SR 11-7 / OCC background, generally find that centralized works until the AI system portfolio crosses approximately 25-30 production systems, at which point the bottleneck dynamics surface. The migration path:

• Y0, centralized. Build the central spine. Establish policy, standards, FRIA methodology, AIMS evidence vault, the AI Governance Committee. Run the first 10-25 AI deployments through the central function. Build credibility with the board AI subcommittee and with the BU leaders.
• Y1, identify embedded candidates and pilot. Recruit 2-3 AI Champions in the BUs running the most AI. Build the community of practice as a forum (initially small). Define decision rights as Tier 1 / 2 / 3. Begin shared-services rollout (eval pipeline first).
• Y2, scale embedded. Recruit AI Champions across all BUs. Run the first rotation cohort (one central-to-BU, one BU-to-central). Mature the community of practice into monthly cadence. Roll out the full shared-services stack. Begin charge-back baseline (not yet billing).
• Y3+, optimized hybrid. Charge-back live. Rotation routinized. Community of practice mature. Decision rights matrix annually reviewed. Central spine adjusted to the hybrid steady state. Internal AI Audit (3L) mature and reporting through the Audit Committee.

Migration B, From All-Embedded-Ad-Hoc (Y0), Centralize-then-Hybridize

Enterprises that arrive at the CAIRO appointment after AI has already proliferated across BUs in an ad-hoc way, typical in technology firms and consumer-product firms where AI feature velocity ran ahead of governance, face a different migration. The mistake is to leave the embedded pattern in place and try to "add a central function." That fails. The BU-level practices are entrenched, incommensurate, and resistant to centralized standards being layered on top. The correct migration is centralize-then-hybridize:

• Y0, discover and inventory. Run the 30-day shadow-AI sprint (lesson 021). Build the enterprise AI use inventory (lesson 020). Surface the incommensurate BU practices through the inventory exercise.
• Y1, centralize. Build the central spine deliberately and assertively. Establish single enterprise policy / standards / FRIA methodology that supersede the BU-local practices. Run the AI Governance Committee. Establish AIMS evidence vault. Mandatory enterprise compliance for all AI deployments from this point. The Y1 conversation is hard, BUs lose autonomy they had grown accustomed to. The board AI subcommittee must back the CAIRO publicly through this transition.
• Y2-Y3, hybridize. Once enterprise standards are bedded in and BUs understand the new baseline, recruit AI Champions inside the BUs. Re-introduce BU-level operational presence under the central standards. Build shared services. Roll out community of practice, decision rights, rotation, charge-back. The hybrid steady state is reached around Y3, two years after the central reset.

The centralize-then-hybridize path is harder politically than the centralized-to-hybrid path. It is also the only path that works from the all-embedded-ad-hoc starting point. Attempting to skip the centralize step, to go from ad-hoc embedded directly to hybrid, typically fails because the embedded layer continues to operate by its prior practices and the central spine never gains operational authority over what the BUs do. The 2026 case study literature includes multiple anonymized examples of this skip failing; the documented failure mode is consistent.

Cross-Walks - 3LoD, RACI, and Regulatory Framework Alignment

Three Lines of Defense (IIA Three Lines Model 2020 Update)

The hybrid pattern must preserve the Three Lines of Defense structure (lesson 073) or it fails an ISO 42001 Stage 2 audit on A.3 leadership + A.4 organizational structure. Mapping is explicit:

• 1L, business units as model owners. AI Champions sit in 1L by design. The BU leader owns the model risk for systems deployed in the BU. AI Champions execute the operational AI risk work inside 1L.
• 2L, central AI Risk Office. CAIRO, AI Risk Office leadership, FRIA team, IMV, Red Team, Policy + Standards, AIGC secretariat. Independent oversight of 1L. Reports to the board AI subcommittee.
• 3L - Internal AI Audit. Reports to the board Audit Committee (not to the CAIRO). Independent assurance over both 1L and 2L. Never embedded in BUs. Never reports through the CAIRO solid-line.

The embedded-champion-in-BU design is consistent with 1L placement because the AI Champion's role is operational execution under enterprise standards, not independent oversight. The dotted-line into the central AI Risk Office is for standards alignment, community-of-practice participation, and escalation, not for performance management or compensation. Performance management and compensation flow through the BU leader (1L manager). The structural distinction matters in any regulator engagement on independence questions.

RACI - The 12-Role x 8-Stage AI Lifecycle Matrix

The hybrid pattern fits the 12-role x 8-stage RACI (lesson 075) naturally if the central and embedded roles are named explicitly in the matrix. The 12 roles include AI Champion (embedded, 1L), Model Owner (1L, BU leader's direct report), AI Risk Office Reviewer (2L, central spine), IMV Lead (2L, central spine), Red Team Lead (2L, central spine), Internal AI Audit (3L, central spine reporting to Audit Committee), AIGC Chair (2L oversight), CAIRO (2L head), Legal Counsel (2L support), Data Protection Officer (2L support), CISO (2L support, cybersecurity), Business Sponsor (1L, BU leader). The 8 lifecycle stages run from intake through retire. The matrix's R / A / C / I assignments are tier-1 / 2 / 3 informed: tier-1 decisions show A in 2L (CAIRO or AI Risk Office Reviewer); tier-2 decisions show A in 2L with C to BU; tier-3 decisions show A in 1L with C to 2L and I to Internal AI Audit. The RACI matrix is the operationalization of the decision-rights tiering at the lifecycle-stage level.

Regulatory Framework Cross-Walks

The hybrid pattern cross-walks cleanly to each of the major regulatory frameworks and standards:

• EU AI Act Article 17 (QMS). The QMS must be documented in writing through policies, procedures, and instructions. The hybrid pattern's centralized policy / standards (element 1) satisfies the QMS documentation obligation; the embedded execution (element 2) plus shared services (element 5) satisfies the operative-QMS obligation; the decision rights (element 4) document the responsibility framework Article 17(3) requires.
• EU AI Act Article 26 (deployer obligations). Each BU as deployer (1L) is the operative owner of Article 26 obligations: instructions-for-use compliance, human oversight, input data quality control, monitoring. The embedded AI Champion is the operational interface that ensures the BU's deployer obligations are met under central standards.
• ISO/IEC 42001:2023 Annex A.3 (leadership) + A.4 (organizational structure) + Clause 7 (support). A.3 is satisfied by the CAIRO + board AI subcommittee structure. A.4 is satisfied by the explicit central spine + embedded champion design with decision rights documented. Clause 7 support is satisfied by shared services + community of practice + rotation.
• NIST AI RMF Govern 1.1 (organizational structures, policies, and processes) + Govern 5.1 (responsibilities and lines of authority). Govern 1.1 is satisfied by the seven-element design. Govern 5.1 is satisfied by the decision-rights matrix + 3LoD mapping + named-role accountability.
• SR 11-7 (governance pillar). Central spine 2L + 3L mapping aligns with SR 11-7's expectation that model risk management is institutionally independent of model development. Embedded AI Champions in 1L are model-development-side, not model-risk-management-side; the distinction is preserved.
• IIA Three Lines Model 2020 update. Direct mapping as described above. The 2020 update's emphasis on "lines of accountability and reporting" rather than rigid functional separation accommodates the embedded-champion-in-1L pattern explicitly.

Common 2026 Operating-Model Failures + the Acme Worked Example

Seven Common 2026 Operating-Model Failures

1. Embedded without central standards. Each BU runs its own AI risk practice. No enterprise aggregation. No common FRIA template. No shared eval pipeline. ISO 42001 Stage 2 surfaces this as a Clause 7 + A.4 finding. The market surveillance authority surfaces it the first time an Article 73 incident has to be aggregated across BUs. Remediation: centralize-then-hybridize migration path (above).
2. Centralized without embedded champions. Single central function trying to scale linearly with AI portfolio growth. Bottleneck. Then shadow-AI evasion. The "department of no" reputation. Remediation: introduce embedded champions in the highest-AI BUs first; iterate.
3. Hybrid without rotation. Central spine becomes insular. Embedded champions drift into BU-local conventions. Standards stop landing in the field; field reality stops landing in the standards. Remediation: institute the rotation program (element 6), six-month embeds in both directions, two annually each direction.
4. Ambiguous decision rights matrix. Every decision is renegotiated. Paralysis. The CAIRO and the BU leaders argue over every FRIA scope, every IMV tasking, every Red Team engagement. Remediation: write the tier-1 / 2 / 3 matrix (element 4), have it ratified by AIGC, embed it in the RACI matrix (lesson 075).
5. Internal AI Audit embedded in 2L. Most consequential failure. Internal AI Audit reporting through the CAIRO is a 3L-independence breach under the IIA Three Lines Model 2020 update. The Audit Committee cannot rely on AI Audit findings filtered through the CAIRO. Remediation: re-line Internal AI Audit to report to the board Audit Committee directly, with dotted-line coordination only to the CAIRO. The fix is a board resolution.
6. No shared-tooling investment. Each BU rebuilds eval pipelines, FRIA templates, model cards, ATLAS technique libraries. Massive cost duplication. Inconsistent quality. Remediation: shared-services investment (element 5), central operates the canonical tooling stack; BUs consume.
7. No charge-back. Central spine seen as free tax. BUs over-consume. The CFO challenges central-spine cost annually. The CAIRO has no defensible breakdown of central-spine cost by BU. Remediation: institute charge-back model (element 7): internal allocation against BU P&L, published unit rates, monthly statements.

The Acme 2026 Worked Example

Maya's design memo to the Acme board AI subcommittee, dated October 22, 2026, recommended the hybrid pattern with all seven elements implemented. The board approved on November 4, 2026. The structure as of Q4 2026:

• Central spine: 18 FTE. CAIRO + AI Risk Office (3) + FRIA / risk-assessment (3) + IMV (3) + Red Team (3) + Internal AI Audit (2, dotted-line to CAIRO; solid-line to board Audit Committee) + AIGC secretariat / AIMS custodian (2) + Policy + Standards (1). Annual run-rate $4.8M.
• Embedded: 12 AI Champions across 7 BUs at 25% time. Effective 3.0 FTE distributed. Annual run-rate $1.1M (in BU budgets, not central).
• Shared services: Eval pipeline (Anthropic Inspect + OpenAI Evals + promptfoo + Garak + PyRIT); ATLAS technique library; CycloneDX ML-BoM repository; AIMS evidence vault; template libraries (Annex IV, FRIA, model card, system card). Q1-Q3 2026 investment $2.4M; estimated avoided cost $5.8M.
• Decision rights: Tier-1 / 2 / 3 matrix ratified by AIGC October 2026; embedded in 12-role x 8-stage RACI; annual review on the AIGC calendar.
• Community of practice: Monthly first Thursday 14:00-16:00 CET. Recorded for AIMS retention. Mandatory attendance for AI Champions, FRIA team, IMV, Red Team, AI Risk Office leads.
• Rotation program: First cohort launched September 2026, two central-to-BU (FRIA specialist to Manufacturing; Red Team junior to Customer Success), two BU-to-central (HR AI Champion to FRIA team; Sales AI Champion to IMV). Six-month duration. Next cohort March 2027.
• Charge-back: Q4 2026 go-live. Published unit rates per service category. Monthly statements to BU leadership. CFO joint sign-off with CAIRO on the rate card. Q1 2027 first full-quarter consumption visibility review.
• CMMI maturity target: Level 3 (Defined) for the AI risk function by end of Y2 (2027). Level 4 (Quantitatively Managed) by end of Y3 (2028). The hybrid design is the structural prerequisite.

The Acme structure is not the only valid 2026 hybrid. It is a reference instantiation. Smaller enterprises may run a 6-9 FTE central spine with 4-6 embedded champions. Larger global enterprises may run a 30-50 FTE central spine with regional sub-spines and 40+ embedded champions across geographies and BUs. The seven elements scale; the specific staffing levels and rate cards do not. The board AI subcommittee resolution that ratifies the design, and the annual review that refreshes it as the AI portfolio grows, is what makes the design durable.

Penalty Exposure + Regulatory Cross-Walk

Operating-model design carries indirect rather than direct penalty exposure under the EU AI Act. There is no "Article X penalty for wrong operating model." The penalty exposure surfaces through the systems and processes the operating model supports: an incoherent operating model means the QMS does not satisfy Article 17, the FRIA process does not satisfy Article 27, the post-market monitoring does not satisfy Article 72, the incident reporting does not satisfy Article 73, the technical documentation does not satisfy Annex IV. Each of those Article-level failures carries its own Article 99 exposure, most at the €15M / 3% of global turnover Article 99(3) level. The aggregate exposure from an incoherent operating model is therefore the sum of the Article-level exposures the model fails to support, compounded by ISO 42001 Stage 2 findings that surface the structural problem to certification bodies and to regulators that read the certification record.

The 2026 ISO 42001 Stage 2 audit literature consistently flags operating-model coherence as a leading Clause 7 (support) + A.3 (leadership) + A.4 (organizational structure) finding. Auditors look for written decision rights, named roles, documented reporting lines, evidence of community-of-practice operation, evidence of independent audit access, and evidence of independent IMV. The hybrid pattern with the seven elements documented satisfies the audit; the embedded-without-standards or centralized-without-embedded variants do not. The 2027 regulator-engagement experience (early Member State market surveillance authority engagements with first-wave deployers) reinforces the pattern, the MSA's first questions are organizational. Who decides? Who signs? Who escalates? Who audits? The hybrid pattern with documented decision rights answers each cleanly; ad-hoc structures do not.

Key Takeaways

• The structural question every CAIRO must answer in Year 1 is where AI risk capability sits relative to the business: centralized, embedded, or hybrid. The 2026 mainstream answer is hybrid (federated) for enterprises with five or more BUs running AI in production.
• The five-dimension scorecard for choosing the archetype: independence preservation (centralized wins), speed-to-business (embedded wins), cost efficiency (hybrid wins at scale; centralized wins at small scale), talent leverage (centralized wins for scarce talent; hybrid wins for blended talent), cultural fit (depends on enterprise heritage). Scoring is a structured conversation with the board AI subcommittee, not a deterministic algorithm.
• The hybrid pattern is seven elements that must each be present: centralized functions (CAIRO + AI Risk Office + Red Team + IMV + Internal AI Audit with 3L exception); embedded AI Champions in BUs at 10-25% time; monthly community of practice; tier-1 / 2 / 3 decision rights matrix; shared services (eval pipelines, registries, repositories, AIMS vault, template libraries); rotation program with six-month embeds in both directions; central charge-back model creating economic discipline.
• Migration paths: from all-centralized (Y0) to hybrid (Y1-Y2) to optimized hybrid (Y3+), the standard financial-sector / ISO-42001-first path. From all-embedded-ad-hoc (Y0) to centralize-then-hybridize (Y1 = centralize; Y2-Y3 = hybridize), the typical Big Tech / consumer-product path. Attempting to skip the centralize step from ad-hoc embedded is a documented failure.
• 3LoD preservation is non-negotiable. 1L = business units as model owners with AI Champions sitting in 1L (not 2L embedded into 1L). 2L = central AI Risk Office reporting to board AI subcommittee. 3L = Internal AI Audit reporting to board Audit Committee, never to the CAIRO solid-line, never embedded in BUs. The 3L independence breach is the most consequential operating-model failure under the IIA Three Lines Model 2020 update.
• The seven common 2026 failures: embedded without central standards (chaos + no aggregation); centralized without embedded champions (slow + disconnected); hybrid without rotation (insular); ambiguous decision rights matrix (paralysis); Internal AI Audit embedded in 2L (3L breach); no shared-tooling investment (each BU rebuilds); no charge-back (central seen as free tax).
• The hybrid pattern cross-walks cleanly to EU AI Act Articles 17 (QMS) + 26 (deployer obligations); ISO/IEC 42001:2023 Annex A.3 + A.4 + Clause 7; NIST AI RMF Govern 1.1 + 5.1; Federal Reserve SR 11-7 governance pillar; IIA Three Lines Model 2020 update. Operating-model design is the structural prerequisite for each of those frameworks landing in operative practice.
• Penalty exposure from operating-model design is indirect, surfacing through aggregate Article 99 exposure where the QMS / FRIA / PMM / incident reporting / technical documentation each fail because the operating model does not support them. ISO 42001 Stage 2 findings on Clause 7 + A.3 + A.4 are the leading regulator-facing indicator of incoherent operating-model design.