Program Architecture

The eight-track model tells you what an AI transformation has to get right. It does not tell you who owns each track, who decides between them when they compete for budget, or how the whole program is staffed and funded. That is the job of the program architecture — the operating model that runs all eight tracks as one coordinated effort rather than eight feuding initiatives.

Getting this layer right is not optional overhead. IBM's study of Chief AI Officers found that organizations running a centralized or hub-and-spoke AI operating model can realize up to 36% higher ROI on AI initiatives than those that leave AI fully decentralized (IBM, 2025). The structure you choose to coordinate the work is itself a determinant of whether the work pays off. This page covers the three operating-model patterns and when each fits, the program-level governance that sits above them, the roles that staff the program, and how to allocate budget across the eight tracks as you mature — all anchored to the binding-constraint principle.

A program architecture is not an org chart — it is a decision-rights map. Its real output is clarity on who decides what, on what cadence, against what evidence. The boxes and reporting lines are downstream of that.

Three Operating Models: CoE, Federated, and Hub-and-Spoke

There are three canonical ways to organize an AI program, and the right one depends almost entirely on your maturity and the binding constraint you are trying to relieve.

The Center of Excellence (CoE) — fully centralized. A single dedicated team owns AI strategy, standards, platform, governance, and the scarce specialist talent, and the rest of the organization draws on it. The CoE concentrates expertise where it is scarce and enforces consistency — which is exactly what an early-stage organization needs and exactly what relieves the most common early constraints (no platform, no controls, no specialist talent). The data backs the concentration: only 31% of CFOs report having even a data-science center of excellence (IBM, 2024), so for most organizations a CoE is still a step up in coordination, not a legacy structure to dismantle.

The federated/embedded model — fully distributed. AI capability lives inside each business unit, which owns its own use cases, talent, and roadmap. This maximizes domain relevance and speed-to-business-context, but at the cost of duplicated effort, inconsistent governance, and platform sprawl — the too-parallel "boil the ocean" failure mode made structural. It only works when the foundation (governance, platform, data) is already mature enough that distributed teams can be trusted to build on it safely.

The hybrid hub-and-spoke — centralized core, distributed delivery. A central hub owns the things that must be consistent — governance, risk, platform, data standards, and the specialist core — while spokes embedded in each business unit own the things that must be local — use-case selection, workflow redesign, and adoption. This is where mature organizations converge, and the survey evidence shows it directly. McKinsey's State of AI found that organizations centralize the elements where consistency is non-negotiable and distribute the rest: risk and compliance is centralized by 57% of organizations and data governance by 46%, while technology talent (49%) and adoption of AI solutions (44%) are most often run on a hybrid basis (McKinsey, 2025). The pattern is not "centralize or decentralize" — it is centralize the guardrails, federate the delivery.

Model	Best when	Owns centrally	Owns locally	Risk
CoE (centralized)	Early maturity (Nascent–Developing); foundation is the constraint	Everything	Little	Bottleneck; distance from the business
Federated (distributed)	Rare; only with a mature, trusted foundation	Little	Everything	Sprawl; inconsistent governance; duplication
Hub-and-spoke (hybrid)	Scaling maturity; foundation is solid, value is the constraint	Guardrails: governance, risk, platform, data standards, specialist core	Delivery: use cases, workflow redesign, adoption	Hub/spoke boundary disputes

The progression maps onto the maturity ladder: most organizations should start as a CoE to relieve foundation constraints (AI Governance & Risk, Data Readiness, Technology Architecture & Platform), then graduate to hub-and-spoke as Execution (Workflow Optimization & Automation, AI Adoption & Culture, Talent & Capability Building) scales out into the business and the CoE risks becoming a bottleneck. Pure federation is rarely a deliberate destination — it is usually what bottom-up tool sprawl looks like before anyone imposes a hub.

Program Governance: Steering, Decision Rights, and Cadence

Above whichever operating model you choose sits the program's governance — the standing structure that keeps eight interdependent tracks pointed at the same outcomes. This is distinct from the AI governance track, which governs AI systems (model risk, acceptable use, compliance); program governance governs the transformation (priorities, funding, sequencing).

The closest evidence on senior ownership comes from AI governance, not program governance. McKinsey found CEO oversight of AI governance was one of the elements most correlated with self-reported bottom-line impact from gen AI, strongest at larger companies (McKinsey, 2025). Yet only 28% report that their CEO actually oversees it (McKinsey, 2025). Senior AI roles are spreading fast. One CEO survey found organizations reporting a Chief AI Officer rose from 26% in 2025 to 76% in 2026 (IBM, 2026). A Gartner poll found 55% of organizations now report having an AI board (Gartner, 2024). The McKinsey finding is about AI governance, but we read it as a sign that accountability near the top matters for the program too.

Three components make the governance real:

• A steering committee — chaired by the CDAO/CAIO, with the eight track owners and the business-unit leaders whose P&Ls the program touches. It owns the cross-track trade-offs no single track owner can make: which track gets the next dollar, which dependency is currently the binding constraint, and when to slow one track because its upstream is not ready.
• Explicit decision rights. Write down who decides, who is consulted, and who is merely informed for each class of decision — use-case funding, platform standards, risk exceptions, hiring the specialist core. Ambiguity here is what turns the hub/spoke boundary into a turf war. The general rule mirrors the operating model: the hub decides guardrails (standards, risk appetite, platform), the spokes decide delivery (which use cases, in what order, for their business).
• A cadence tied to evidence, not the calendar. A monthly steering review of the integrated assessment and a quarterly re-allocation of budget against measured value — fed by the measurement track. The cadence exists to let evidence pull the next investment, which is exactly how the framework says to pace staggered intensity across the tracks.

Decision rights beat reporting lines. A program with a clear RACI and a fuzzy org chart will outperform one with a beautiful org chart and ambiguous decision rights every time. Spend the design effort on who decides, not on who reports to whom.

Roles: From the CDAO to the Working Teams

A staffed program has four tiers, each accountable to the one above.

CDAO / Chief AI Officer — the accountable executive. Owns the program's outcomes, chairs the steering committee, sets risk appetite, and makes the build/buy/partner bets. The role is consolidating fast: the established CDO/CDAO function is now present in roughly 83% of large organizations, up from 12% in 2012 (MIT SMR / Wavestone, 2024), and 70% of CDAOs now have primary responsibility for building the AI strategy and operating model (Gartner, 2025). Whether titled CAIO or CDAO, what matters is that one named executive owns AI end-to-end and reports high enough to reallocate capital. For the operator's manual for this seat, see Chief AI Officer in AI Accountability.

AI Program Lead — the operator. Runs the program day-to-day: drives the cadence, maintains the integrated assessment, manages cross-track dependencies, and escalates the binding constraint to the steering committee. Where the CDAO sets direction and owns outcomes, the Program Lead owns execution of the operating model — the orchestration that keeps eight tracks from colliding. In a hub-and-spoke model this person typically runs the hub. AI Accountability covers this same seat as the AI Transformation Lead.

Track Owners — eight accountable leads. One owner per track, each accountable for that track's key output and its maturity score on the readiness assessment. Track owners are not necessarily full-time — early on, one person may own several adjacent tracks (e.g., Data and Platform). What is non-negotiable is that every track has a named owner; an unowned track is a guaranteed binding constraint waiting to happen.

Two of the eight already have a dedicated operator's manual in AI Accountability: AI Governance & Risk splits across the AI Governance Lead, AI Risk Officer, and AI Security Lead; AI Adoption & Culture is the AI Enablement Lead. AI Strategy & Leadership is the CDAO/CAIO seat above. The remaining five are typically owned by an existing role whose remit AI has expanded, not a new AI-specific seat: Data Readiness usually stays with the same CDO/CDAO who owns AI strategy — Deloitte's own CDO survey found data governance is the CDO's top stated priority (51%) (Deloitte, 2025); Technology Architecture & Platform is an engineering execution function (MLOps/platform engineering) that builds what the Program Lead and Security Lead decide, not a governance seat of its own; Workflow Optimization & Automation is typically owned by business-unit or operations leaders, self-evident enough that it rarely gets its own title; Talent & Capability Building usually falls to the CLO/HR function, though HR itself says it's only barely in the running for this work today (28%, just behind cross-functional task forces at 29%) (SHRM, 2026); and Measurement & Value Realization usually routes through the CFO rather than a dedicated title — the rare organizations that hand CFOs this accountability (just 2%) see 76% of them achieve substantial AI value, far above any other owner (Fortune, 2026).

Working teams — the delivery layer. Beneath the track owners sit the people who do the work, drawn from the capability stack: the specialist core (ML engineers, data scientists, MLOps, platform owners) that the hub concentrates, the builders embedded in spokes who compose AI into workflows, the workflow and change leads who run redesign and adoption, and the network of internal champions that carries fluency into each function. In a CoE these sit centrally; in hub-and-spoke the specialist core stays in the hub while builders, workflow, and adoption roles distribute into the spokes.

Budgeting Across the Eight Tracks by Maturity Stage

AI budgets are growing fast — 84% of leaders intend to increase AI investment in the coming year (Deloitte, 2026) — but growth is not a plan. The binding-constraint principle is also a budgeting principle: fund the track that is actually limiting outcomes, not the most visible or fashionable layer. That means the allocation should shift as the binding constraint moves down the dependency chain.

A recurring and dangerous error is to spend almost entirely on technology while starving the people-and-process layer where value actually compounds. The corrective is BCG's 10-20-70 rule of thumb: 70% of strategic focus on people and processes, 20% on technology, and 10% on algorithms (BCG, 2025). It is guidance on focus, not a literal line-item split, but a useful counterweight to platform-heavy budgets. The independent budget benchmarks show how far most organizations still are from that balance: in the Wharton survey of enterprise GenAI budgets, technology tools and systems take the largest share and roughly 31% goes to internal R&D, while employee training ranks far lower (Wharton / GBK Collective, 2025). Real budgets skew toward technology and away from the people-and-process layer — which is exactly the gap the 10-20-70 framing warns against, and exactly what should shift as you scale.

Early stage (Nascent–Developing). The binding constraint is almost always in the Foundation and Direction layers. Concentrate spend in a CoE on AI Governance & Risk, Data Readiness, and Technology Architecture & Platform — the guardrails and substrate everything else runs on — plus a light, fast AI Strategy & Leadership exercise to set intent. Underweight broad AI Adoption & Culture and large-scale Workflow Optimization & Automation spend here: rolling tools out before the foundation is ready produces logins without usage. Talent & Capability Building spend at this stage targets the specialist core, hired in to stand up the foundation.

Scaling stage (Emerging–Scaling). The foundation is solid; the binding constraint moves into Execution. Shift spend toward Workflow Optimization & Automation, AI Adoption & Culture, and broad Talent & Capability Building — the people-and-process layer — and toward Measurement & Value Realization so evidence can pace the next allocation. This is also where the operating model graduates to hub-and-spoke and budget follows: the hub holds a shrinking platform/governance line while the spokes carry a growing share of delivery and adoption spend. The danger here is the reverse of early stage — continuing to pour money into platform when the constraint has already moved to whether anyone is changing how they work.

Layer / tracks	Early (Nascent–Developing)	Scaling (Emerging–Scaling)
Direction (AI Strategy & Leadership, AI Governance & Risk)	Moderate — set intent + guardrails	Lower — established, maintain
Foundation (Data Readiness, Technology Architecture & Platform)	Heavy — relieve the constraint	Moderate — maintain + extend
Execution (Workflow Optimization & Automation, AI Adoption & Culture, Talent & Capability Building)	Low–moderate — specialist core only	Heavy — the value layer at scale
Feedback (Measurement & Value Realization)	Light but present from day one	Moderate — paces re-allocation

The mechanism that keeps this honest is the quarterly evidence-based re-allocation in the governance cadence: the measurement track shows which investments moved value, and the steering committee moves the next dollar to the current binding constraint. For the stage-by-stage detail of what to fund when, see the sequencing playbooks; for standing the program up, the 90-day launch; and for keeping stakeholders aligned through the shifts, the communication templates.

Key Takeaways

• The operating model is a performance lever, not overhead. A centralized or hub-and-spoke model can deliver up to 36% higher AI ROI than full decentralization (IBM, 2025). Design it deliberately.
• Match the model to your maturity. Start as a CoE to relieve foundation constraints, then graduate to hub-and-spoke as Execution scales — centralize the guardrails (risk 57%, data governance 46%), federate the delivery (talent and adoption run hybrid) (McKinsey, 2025).
• Senior, named ownership matters. In the closest evidence, CEO oversight of AI governance is among the practices most correlated with self-reported bottom-line impact, yet only 28% have it (McKinsey, 2025). The CDAO/CAIO role is consolidating fast (IBM, 2026; Gartner, 2025). Put one accountable executive over the whole program.
• Decision rights beat org charts. Govern the transformation (priorities, funding, sequencing) with a steering committee, an explicit RACI, and an evidence-paced cadence — distinct from governing AI systems.
• Staff four tiers: CDAO/CAIO (accountable), Program Lead (operator), eight named Track Owners, and working teams. Every track needs a named owner; an unowned track is a binding constraint waiting to happen.
• Budget the binding constraint, and let it move. Fund Foundation heavily when early, shift to the people-and-process Execution layer as you scale — people and process should get ~70% of the effort (BCG, 2025), yet real budgets still skew to technology with training ranking far lower (Wharton / GBK Collective, 2025). Re-allocate quarterly on measured evidence.

Sources

• IBM — The CAIO Study (IBM Institute for Business Value with Dubai Future Foundation), 2025. Organizations where CAIOs drive a centralized or hub-and-spoke AI operating model can realize up to 36% higher ROI on AI initiatives than those that leave AI fully decentralized. View source · verified 2026-06-20 · primary
• IBM — 2024 CFO Study (IBM Institute for Business Value), 2024. Only 31% of CFOs surveyed have launched a center of excellence for data science. View source · verified 2026-06-20 · primary
• McKinsey — The State of AI: How Organizations Are Rewiring to Capture Value, 2025. Risk and compliance is centralized by 57% of organizations and data governance by 46%, while technology talent and adoption of AI solutions are most often run on a hybrid basis. View source · verified 2026-06-20 · primary
• McKinsey — The State of AI: How Organizations Are Rewiring to Capture Value, 2025. Our survey analyses show that a CEO's oversight of AI governance ... is one element most correlated with higher self-reported bottom-line impact from an organization's gen AI use. ... That's particularly true at larger companies, where CEO oversight is the element with the most impact on EBIT attributable to gen AI. Twenty-eight percent of respondents whose organizations use AI report that their CEO is responsible for overseeing AI governance ... and 17 percent say AI governance is overseen by their board of directors. View source · verified 2026-09-14 · primary
• IBM — 2026 CEO Study (IBM Institute for Business Value), 2026. 76% of organizations report having a Chief AI Officer in 2026, up from 26% a year earlier. View source · verified 2026-06-20 · primary
• Gartner — Gartner Poll Finds 55% of Organizations Have an AI Board, 2024. A Gartner poll of more than 1,800 executive leaders found 55% of organizations have an AI board. View source · verified 2026-06-20 · primary
• MIT Sloan Management Review / Wavestone — Data and AI Leadership Executive Survey 2024, 2024. 84.3% of organizations have appointed a chief data officer or chief data and analytics officer, up from just 12% in 2012. View source · verified 2026-06-20 · primary
• Gartner — Gartner Survey Finds 70% of CDAOs Are Responsible for AI Strategy and Operating Model, 2025. 70% of chief data and analytics officers have the primary responsibility for building the AI strategy and operating model for the organization. View source · verified 2026-06-20 · primary
• Deloitte — Chief Data Officer Survey 2025, 2025. Data governance was the top priority for the year ahead at 51%. View source · verified 2026-06-20 · primary
• SHRM — The State of AI in HR 2026 Report, 2026. Upskilling and reskilling the workforce was the only area of AI implementation where HR functions were close to taking the lead (28%). View source · verified 2026-07-03 · primary
• Fortune (citing Return on AI Institute research) — Why CFOs—not chief AI officers—are the secret to getting real value from AI, 2026. only 2% of companies assign CFOs this accountability role, yet when they do, 76% achieve substantial value. View source · verified 2026-07-03 · primary
• Deloitte — State of AI in the Enterprise 2026, 2026. 84% of organizations are increasing their AI investments. View source · verified 2026-06-20 · primary
• BCG — The Widening AI Value Gap, 2025. Adherence to our 10-20-70 rule for technology transformations will help speed the journey: 70% of a business's strategic focus should be on the people and processes, 20% on the tech, and 10% on algorithms. View source · verified 2026-09-13 · primary
• Wharton / GBK Collective — 2025 AI Adoption Report (Gen AI Fast-Tracks Into the Enterprise), 2025. Technology tools and systems take the largest share of enterprise GenAI budgets and roughly 31% goes to internal R&D, while employee training ranks far lower. View source · verified 2026-06-20 · primary