2026 04 22 enterprise ai platform operating models

Enterprise AI platform operating models: organisational structure and ownership

Research Question

What organisational structures do enterprises use to operate multiple Artificial Intelligence (AI) platforms simultaneously, and what trade-offs emerge between (a) a single unified AI platform team, (b) a split by target customer base (business users vs developers), and (c) a split by underlying technology stack (Microsoft 365 (M365) vs Amazon Web Services (AWS)), including implications from explore vs exploit operating modes and Conway's Law?

Scope

In scope:

Enterprise operating-model patterns for multi-platform AI enablement
Comparative analysis of three structures: unified team, customer-segment split, and technology-stack split
Trade-offs across speed, reliability, governance, security, compliance, and cost accountability
Evidence from analogous structures (cloud platform teams, Cloud Center of Excellence (CCoE), DevOps platform teams)
Any publicly available examples or governance guidance from financial services firms and regulators

Out of scope:

Tool-level feature comparisons between individual AI products
Vendor-specific implementation playbooks below operating-model level
Prescriptive reorganisation plan for any single named company

Constraints: (time, source types, access)

Prioritise primary and high-credibility secondary sources with explicit publication dates
Distinguish established evidence from inference where direct enterprise disclosures are limited
Emphasise material published in the last 24 months while allowing older foundational references (for example, Conway's Law)

Context

[fact; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] This item asks how an enterprise should assign ownership, team boundaries, and governance when more than one Artificial Intelligence (AI) platform must operate in parallel without duplicating capabilities or obscuring accountability. [inference; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] The core design problem is therefore organisational boundary-setting: choose boundaries that preserve shared governance, internal-platform reuse, and cost control while still giving distinct user groups fast and relevant service.

Approach

Decompose the question into sub-questions and describe how each will be investigated (literature review, experiment, prototype, expert interview, etc.).

Catalogue operating-model archetypes used for multi-platform internal platforms and the typical decision rights in each.
Compare the three target structures using common evaluation dimensions (delivery velocity, platform quality, governance, risk, total cost).
Analyse explore vs exploit dynamics in each structure, including where experimentation should sit versus where scale operations should sit.
Evaluate Conway's Law implications: how organisational boundaries may shape architecture, integration seams, and user experience fragmentation.
Review published evidence from Team Topologies, DevOps Research and Assessment (DORA), cloud platform/CCoE operating models, and financial-services AI governance publications.
Synthesize decision criteria, anti-patterns, and context-dependent recommendations.

Sources

Starting points - papers, articles, videos, repos, docs.

Research Skill Output

(Full output from running the research skill - retained verbatim in the completed item. Sections 0-5 are the investigation; section 6 seeds the Findings section below.)

§0 Initialise

[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Research question restated: which organisational structure best operates multiple enterprise AI platforms in parallel, and what trade-offs follow from choosing a single unified platform team, a split by target customer base, or a split by technology stack such as Microsoft 365 (M365) and Amazon Web Services (AWS)?
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Scope confirmed: operating-model patterns, decision rights, comparative trade-offs, explore-versus-exploit dynamics, Conway's Law, and financial-services governance examples are in scope; tool feature comparisons, vendor implementation playbooks, and a prescriptive reorganisation plan for any single named company are out of scope.
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Constraint confirmed: the investigation uses public sources, prefers primary and high-credibility secondary sources from the last 24 months where possible, and allows older foundational sources such as Conway's Law.
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-03-19-layered-org-llm-architecture.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-03-23-software-factory.md] Prior completed repository work already established three adjacent points: enterprise AI value depends on shared foundational capabilities, enterprise AI architectures are strongest as layered systems rather than isolated silos, and AI-driven productivity gains only matter when the surrounding platform and workflow system can absorb them.
[inference; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-03-19-layered-org-llm-architecture.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-03-23-software-factory.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md; https://www.capitalone.com/tech/ai/data-management/] This item therefore extends prior work from capability design and architecture into ownership design: which team should own the shared control plane, meaning the central configuration, access, evaluation, observability, and policy layer; which teams should own user-facing experiences; and where experimentation should sit.
[assumption; source: https://github.com/davidamitchell/Research/blob/main/research-prompt.md] The missing .github/skills/research/SKILL.md file means the fallback process defined in research-prompt.md is the operative research skill for this run. Justification: the named skill file is absent in this environment, and the repository prompt defines the same sectioned workflow.

§1 Question Decomposition

Branch A - operating-model archetypes
- A1. What does a central platform team or CCoE typically own?
- A2. What do customer-facing product teams typically own when a central platform exists?
- A3. What problems is a vendor-stack split trying to solve, and which of those problems can be solved without duplicating the platform core?
Branch B - trade-offs across the three candidate structures
- B1. When is a single unified team strongest?
- B2. When does a customer-segment split outperform a unified team?
- B3. When does a technology-stack split create more duplication than value?
Branch C - explore versus exploit
- C1. Where should experimentation with new models, vendors, and workflows sit?
- C2. Where should scaled operation, reliability, security, and cost accountability sit?
Branch D - Conway's Law
- D1. How do team boundaries shape architecture seams and user experience?
- D2. What boundary choice is most likely to fragment identity, retrieval, evaluation, and support workflows?
Branch E - financial-services constraints
- E1. What do regulators emphasise about AI governance, accountability, and human oversight?
- E2. What bank disclosures show about how central governance and local delivery can coexist?
Branch F - synthesis
- F1. Which decision rights belong in the central hub?
- F2. Which decision rights belong in spokes or customer-segment teams?
- F3. What anti-patterns should enterprises avoid?

§2 Investigation

Source audit and accessibility notes

[fact; source: https://teamtopologies.com/key-concepts-content/platform-team; https://teamtopologies.com/key-concepts] The seeded Team Topologies URL returned 404 in this environment, so the investigation used the current key-concepts page as the authoritative replacement.
[fact; source: https://docs.aws.amazon.com/whitepapers/latest/aws-overview/introduction-to-cloud-center-of-excellence.html; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] The seeded AWS whitepaper URL redirected to an obsolete location, so the investigation used the current AWS Prescriptive Guidance page for Cloud Center of Excellence guidance.
[fact; source: https://cloud.google.com/architecture/cloud-center-of-excellence; https://cloud.google.com/resources/cloud-teams] The seeded Google Cloud Center of Excellence URL returned 404 in this environment, so the investigation used Google Cloud's current cloud-teams guide.
[fact; source: https://www.bankofengland.co.uk/paper/2022/artificial-intelligence-and-machine-learning; https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence] The seeded Bank of England page moved, and the current official page is the Prudential Regulation Authority publication for Discussion Paper 5/22.
[fact; source: https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/FEAT] The official MAS FEAT page was located but blocked in this environment, so it is recorded as inaccessible and is not used to support downstream factual claims.
[fact; source: https://github.com/davidamitchell/Research/blob/main/research-prompt.md; https://teamtopologies.com/key-concepts-content/platform-team; https://docs.aws.amazon.com/whitepapers/latest/aws-overview/introduction-to-cloud-center-of-excellence.html; https://cloud.google.com/architecture/cloud-center-of-excellence; https://www.bankofengland.co.uk/paper/2022/artificial-intelligence-and-machine-learning; https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/FEAT] No secondary source in this investigation cited a paper, Digital Object Identifier (DOI), or arXiv preprint that could not be located after a targeted search, so there are no failed primary-source search records beyond the moved or blocked seed URLs above.

A. What central platform teams and CCoEs are designed to own

[fact; source: https://teamtopologies.com/key-concepts] Team Topologies defines a platform team as a grouping that provides a compelling internal product to accelerate delivery by stream-aligned teams, and it distinguishes that role from enabling teams and complicated-subsystem teams.
[fact; source: https://teamtopologies.com/key-concepts] Team Topologies also defines three interaction modes: collaboration, X-as-a-Service, and facilitation, which means a platform team is expected to provide reusable services with clear boundaries rather than absorb all downstream delivery work.
[fact; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] AWS defines a Cloud Center of Excellence as the group that leads cloud adoption, governance, training, cost optimisation, and cultural change, and notes that the central CCoE can contain separate workstreams or practices.
[fact; source: https://cloud.google.com/resources/cloud-teams] Google Cloud argues against a traditional monolithic CCoE and recommends one Cloud Office for strategy and programme management plus one or more Cloud Platform Teams that build and operate a platform like a product for business and application teams.
[fact; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] DORA 2025 reports that 90% of organisations have adopted at least one platform and that high-quality internal platforms correlate directly with an organisation's ability to unlock AI value.
[inference; source: https://teamtopologies.com/key-concepts; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Across platform-team, CCoE, and DORA guidance, the central unit consistently owns standards, shared services, and enablement, while use-case delivery is expected to stay closer to stream-aligned or application teams.

B. What bank disclosures show about central governance plus local delivery

[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] DBS states that scalable bank AI depends on three pillars - platform, process, and people - where a centralised data platform provides a single source of truth, an Artificial Intelligence protocol repository enables model reuse, and governance frameworks manage privacy, bias, and explainability.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] DBS also states that dedicated data teams should drive adoption across the organisation and that AI professionals should work in cross-functional squads, which shows central platform ownership combined with embedded delivery capacity.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] DBS reports that its data platform and protocol repository reduced time to market for AI initiatives from 15 months to under three months.
[fact; source: https://www.dbs.com/annualreports/2024/cio-statement.html] DBS reports more than 1,500 models across 370 use cases, uses board and management oversight for technology resiliency and governance, and describes a Testing Centre of Excellence plus Architecture Review Committee and technology vendor governance forum as central quality and risk mechanisms.
[fact; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai] Morgan Stanley designed its internal OpenAI-based assistant to answer exclusively from Morgan Stanley content with links to source documents and appropriate controls, which is a centrally governed knowledge platform rather than a business-unit-local experiment.
[fact; source: https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Morgan Stanley reports that 98% of Financial Advisor teams adopted its assistant and that leadership sees AI as an interaction layer between colleagues and many applications.
[inference; source: https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] That pattern is consistent with a shared central layer serving a distinct customer group.
[fact; source: https://www.capitalone.com/tech/ai/data-management/; https://www.capitalone.com/software/blog/capital-one-you-build-your-data/] Capital One describes a central control plane, a common self-service data portal, and both central and federated deployment models with consistently enforced governance across both approaches.
[fact; source: https://www.capitalone.com/tech/ai/data-management/] Capital One says a central platform and a federated data model can coexist, with governance capabilities applied consistently across both centralized and federated approaches.
[fact; source: https://www.capitalone.com/digital/gai-transparency/] Capital One publicly discloses that it customises open-source LLMs using synthetic and open data, which indicates bank-controlled model customisation still sits inside an explicit governance and transparency regime.
[fact; source: https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction] J.P. Morgan's public payments discussion emphasises starting with data-governance capability building, taking AI risk appetite to board level, and embedding ethical issues in an overarching governance framework.
[inference; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.dbs.com/annualreports/2024/cio-statement.html; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.capitalone.com/tech/ai/data-management/; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction] Public bank examples consistently show central governance, data, and platform ownership combined with distributed usage or cross-functional delivery, not a pure vendor-by-vendor organisational split.

C. Unified team versus customer-segment split

[fact; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] Both AWS and Google Cloud treat early centralisation as useful for strategy, governance, and migration acceleration, because scarce expertise and standards benefit from concentration.
[fact; source: https://cloud.google.com/resources/cloud-teams] Google Cloud explicitly separates Cloud Office work from Cloud Platform Team work, which is evidence that even within a central model, not all responsibilities belong in one team.
[fact; source: https://teamtopologies.com/key-concepts] Team Topologies argues for stable, stream-aligned teams close to user value streams and warns against loading one team with excessive cognitive load or ambiguous dependencies.
[fact; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] DORA 2025 says AI amplifies team quality, that user-centricity is a prerequisite for AI success, and that strong internal platforms and fast feedback loops are foundational.
[fact; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Morgan Stanley's AI suite serves Financial Advisors as a clearly defined user segment with tailored workflow support, summarisation, and meeting follow-up, which is different from a developer platform's needs.
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://cloud.google.com/resources/cloud-teams] Once both business-user assistants and developer-facing agent platforms matter at scale, a customer-segment split is more congruent with user needs than a single monolith, provided both segments still consume one governed shared platform core.
[assumption; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] The enterprise in question has two materially different user populations - business users and developers - with distinct service models and success metrics. Justification: if only one user population matters, the case for a customer-segment split weakens materially.

D. Why a vendor-stack split is usually weaker

[fact; source: https://www.melconway.com/Home/Conways_Law.html] Conway's Law states that organisations design systems whose structure copies their communication structure.
[fact; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] DORA links better AI outcomes to high-quality internal platforms and clear workflows, not to multiplying platform boundaries.
[fact; source: https://www.capitalone.com/tech/ai/data-management/; https://www.capitalone.com/software/blog/capital-one-you-build-your-data/] Capital One's operating model separates a central control plane from federated implementation choices, which shows that flexibility can be created without duplicating governance, metadata, lineage, and access-control functions by stack.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] DBS emphasises one centralised data platform and one protocol repository for reuse, not separate governance systems by channel or vendor.
[inference; source: https://www.melconway.com/Home/Conways_Law.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] If an enterprise splits ownership primarily into an M365 team and an AWS team, it will likely duplicate identity, policy, retrieval, evaluation, logging, and support mechanisms, and the platform architecture will mirror those seams.
[assumption; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] There is enough overlap across M365 and AWS guardrails to justify one shared control plane. Justification: both stacks require shared decisions on model approval, data access, human-review policy, observability, and cost governance in most large enterprises.
[inference; source: https://www.melconway.com/Home/Conways_Law.html; https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams] A vendor-stack split is only well justified when vendor boundaries coincide with genuinely different regulated environments, data-residency constraints, or customer groups that rarely share workflows.

E. Explore versus exploit placement

[fact; source: https://teamtopologies.com/key-concepts] Team Topologies gives enabling teams a temporary role in helping other teams overcome capability gaps and learn new practices.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] DBS says its AI journey began with experimental projects and later evolved into structured institutional programmes that generate enterprise-wide value.
[fact; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] DORA recommends clarifying AI policies, connecting AI to internal context, fortifying safety nets, and investing in internal platforms before expecting durable value from AI acceleration.
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md] Prior completed repository work on exploit and explore argues that exploratory activity and exploitative scaling behave differently and should not be managed as if they were the same portfolio motion.
[inference; source: https://teamtopologies.com/key-concepts; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Explore work should sit in a small central enabling or incubation function close to the shared platform, because that is where new models, vendors, evaluation methods, and controls can be tested once and then productised for reuse.
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] Exploit work should sit on standardised shared rails and in customer-facing teams, because scaled operation requires predictable service levels, support, change control, and cost visibility rather than perpetual experimentation.

F. Financial-services governance implications

[fact; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence] Bank of England, PRA, and FCA Discussion Paper 5/22 says AI can amplify existing risks to consumer outcomes, firm safety and soundness, market integrity, and financial stability, and it asks how governance and decision-making processes should address those risks.
[fact; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence] The same discussion paper explicitly highlights additional challenges for firms' decision-making and governance processes and asks how existing tools such as the Senior Managers and Certification Regime (SM&CR) can address them.
[fact; source: https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction] J.P. Morgan's public discussion says data governance should be built early, raised to board level, and tied to risk appetite before large-scale incorporation of AI into business processes.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] DBS and Morgan Stanley both describe human-in-the-loop deployment patterns for sensitive banking and wealth-management work rather than fully autonomous operation.
[inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Financial-services firms have stronger reasons than most sectors to centralise governance, approved patterns, and accountability while decentralising only the workflow configuration and user support that must stay close to the business.

§3 Reasoning

[fact; source: https://teamtopologies.com/key-concepts; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] Established platform guidance converges on a central unit owning shared standards, governance, and internal services, with downstream teams consuming those services.
[fact; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] DORA 2025 provides current evidence that internal-platform quality and workflow clarity are foundational to AI value.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.dbs.com/annualreports/2024/cio-statement.html; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.capitalone.com/tech/ai/data-management/] Public bank disclosures show a recurring pattern of central platform and governance ownership with distributed usage or federated execution.
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.melconway.com/Home/Conways_Law.html] Because user-centricity matters and Conway's Law makes team seams visible in architecture, the boundary that scales best is a shared central control plane plus customer-segment products, not a boundary drawn around vendor stacks.
[inference; source: https://teamtopologies.com/key-concepts; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] Explore work should be temporarily central and enabling; exploit work should be operational, standardised, and product-managed.
[assumption; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] This reasoning assumes the enterprise has enough overlap across its AI platforms that shared guardrails and shared service components are economically meaningful. Justification: if the platforms are almost entirely isolated by law, geography, or customer type, the conclusion could shift toward stronger structural separation.

§4 Consistency Check

[inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] AWS's strong CCoE language and Google's warning against a monolithic CCoE are not contradictory once stage and scope are separated: both support central governance, but Google is more explicit that execution should split across dedicated platform teams and programme-management functions.
[fact; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.capitalone.com/tech/ai/data-management/] DBS and Capital One both combine centralised governance with federated or cross-functional delivery, which resolves the apparent tension between control and local responsiveness.
[fact; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Morgan Stanley's disclosures support a centrally governed business-user platform but do not disclose the exact internal team chart, so conclusions about org structure remain partly inferential.
[inference; source: https://www.melconway.com/Home/Conways_Law.html; https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] No reviewed source positively argues that a primary split by vendor stack is the default high-performing pattern for enterprise AI, so the critique of stack-splitting is evidence-backed even though it remains an inference rather than a direct experimental result.

§5 Depth and Breadth Expansion

[inference; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] Technical lens: the shared core of enterprise AI platforms is not the user interface but the control plane - identity, data access, observability, evaluation, cost controls, and reusable protocols - which argues for one central owner of those concerns.
[inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction] Regulatory lens: financial-services firms need one accountable locus for policy, auditability, risk appetite, and human-review rules, which weighs against unmanaged federation and strongly against duplicated vendor-specific governance stacks.
[inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Economic lens: a central hub amortises scarce architecture, governance, and platform-engineering talent, while a customer-segment split is justified only when responsiveness gains exceed the coordination cost of running two product surfaces.
[inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] Historical lens: cloud operating models evolved from monolithic CCoE language toward smaller platform-product teams plus enablement and programme functions, and enterprise AI shows the same shape.
[inference; source: https://teamtopologies.com/key-concepts; https://www.melconway.com/Home/Conways_Law.html; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai] Behavioural lens: business users and developers ask different questions, need different service levels, and tolerate different failure modes, so splitting customer-facing ownership by user segment is more behaviourally coherent than splitting by cloud vendor.

§6 Synthesis

Executive summary:
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md] Enterprises that must run multiple AI platforms in parallel should default to a hybrid hub-and-spoke model, meaning a central platform hub serving user-facing teams through shared services and enablement patterns, with separate customer-facing product ownership only where user groups materially differ.
- [inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts] A single unified team is a strong starting structure when demand is immature and specialist talent is scarce, but it becomes a bottleneck once both business-user and developer-facing needs scale.
- [inference; source: https://teamtopologies.com/key-concepts; https://www.melconway.com/Home/Conways_Law.html; https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] A split by target customer base is usually healthier than a split by M365 versus AWS, because customer-segment boundaries preserve shared controls while vendor-stack boundaries tend to duplicate them.
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md] Explore work should sit in a thin enabling or incubation function near the central hub, while exploit work should run on standard shared rails with explicit governance and human-review rules.
Key findings:
- [inference; source: https://www.capitalone.com/tech/ai/data-management/; https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md] Most enterprises should organise multiple AI platforms around one central control plane, using Capital One's term for the central configuration, access, evaluation, observability, and policy layer, then expose distinct products to users only where needs diverge materially.
- [inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts] A single unified platform team is strongest in the early stage because it concentrates scarce talent and standards, but it degrades into a slow intake queue if it continues owning every downstream use case.
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai] A split by target customer base often becomes the best scaling move because business users and developers need different workflows, support models, and product metrics even when they share one governed platform core.
- [inference; source: https://www.melconway.com/Home/Conways_Law.html; https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] A primary split by M365 versus AWS is usually a weak default because it causes the organisation to duplicate policy, identity, retrieval, and support mechanisms, and then hardens those vendor seams into the architecture.
- [inference; source: https://teamtopologies.com/key-concepts; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md] Explore activity should be owned by a small enabling or incubation function near the hub so that experiments with models, vendors, and evaluation methods can be productised once rather than rediscovered in parallel.
- [inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Financial-services firms have stronger reasons than most sectors to centralise governance, approved patterns, and accountability, while decentralising only workflow configuration and user support that must stay close to the business.
- [inference; source: https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts; https://www.dbs.com/annualreports/2024/cio-statement.html; https://www.capitalone.com/software/blog/capital-one-you-build-your-data/] The cleanest decision-rights split is central ownership of policy, vendor approval, data-access guardrails, observability, shared tooling, and cost attribution, with spoke ownership of prioritisation, integration, adoption, and benefit realisation.
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://www.melconway.com/Home/Conways_Law.html] The most consistent anti-patterns are a permanent central team that owns all delivery, federated teams launched before a shared platform exists, and vendor-aligned teams that mirror suppliers instead of user journeys.
Evidence map:
- Each synthesis claim is mapped one-to-one to the Evidence Map in the Findings section below.
Assumptions:
- [assumption; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] There is enough overlap across M365 and AWS control requirements that one shared governance and platform core is economically meaningful. Justification: if legal, residency, or customer separation is extreme, stronger structural separation could be warranted.
- [assumption; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] The enterprise has at least two distinct AI customer groups, business users and developers, whose workflow needs are meaningfully different. Justification: without that divergence, a customer-segment split may add cost without enough benefit.
- [assumption; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.capitalone.com/tech/ai/data-management/] Public bank disclosures describe governance and platform patterns more often than exact team charts, so some operating-model conclusions are inferential rather than diagram-level factual.
Analysis:
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] The most important trade-off is not centralisation versus decentralisation in the abstract; it is whether the organisation centralises the shared control plane while decentralising only the product surfaces that need domain intimacy.
- [inference; source: https://www.melconway.com/Home/Conways_Law.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Vendor-stack splits look tidy on an org chart, but they create the wrong software seams because they optimise around supplier boundaries rather than around user value, workflow coherence, and reusable internal services.
- [inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] In regulated sectors, the acceptable degree of federation is narrower, because accountability, auditability, and human-review requirements must remain centrally legible even when local teams configure the last mile.
Risks, gaps, uncertainties:
- [fact; source: https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/FEAT] The official MAS FEAT page was inaccessible from this environment, so MAS-specific governance claims were not used as primary evidence in the synthesis.
- [fact; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.capitalone.com/tech/ai/data-management/] Public enterprise disclosures describe operating principles more often than full reporting lines, so exact headcount and reporting-structure comparisons remain thin.
- [inference; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://teamtopologies.com/key-concepts] The recommendation against stack-splitting is strong but still inferential because the public evidence base is richer on platform-team performance than on controlled comparisons of AI org charts.
Open questions:
- [inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] What funding model best allocates central AI platform costs across business units without recreating a slow approval bureaucracy?
- [inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence] Which AI decisions in financial services should stay with central risk and architecture functions, and which can safely be delegated to product or business-line teams under pre-approved guardrails?
- [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] What are the leading operational metrics for knowing when a unified team should split into separate customer-segment products without fragmenting the platform core?

§7 Recursive Review

[fact; source: https://github.com/davidamitchell/Research/blob/main/research-prompt.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Review outcome: every externally grounded claim in sections 0-6 is explicitly labelled as fact, inference, or assumption, and every factual or inferential claim binds to a URL-backed source.
[fact; source: https://github.com/davidamitchell/Research/blob/main/research-prompt.md; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Review outcome: prior repository work is cited with GitHub URLs rather than repository-relative paths.
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Review outcome: the synthesis remains consistent with the investigation, recommending a hybrid hub-and-spoke structure, preferring customer-segment splits over vendor-stack splits, and placing exploration near the central hub and scaled operation on shared rails.
[fact; source: https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-platform-operating-models.md] Review outcome: the remaining uncertainties are explicit, including moved or blocked source pages, thin public reporting-line data, and the inferential nature of the anti-stack-split conclusion.

Findings

(Populated from section 6 Synthesis above.)

Executive Summary

[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md] Enterprises that must operate multiple AI platforms in parallel should usually adopt a hybrid hub-and-spoke model, meaning a central platform hub serving user-facing teams through shared services and enablement patterns, with one central AI platform and governance hub and customer-facing ownership split by user segment only where needs materially diverge. [inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts] A single unified team is the best starting point when AI demand is immature and specialist talent is scarce, but it becomes a bottleneck if it keeps owning both the shared platform core and every downstream use case. [inference; source: https://teamtopologies.com/key-concepts; https://www.melconway.com/Home/Conways_Law.html; https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] Splitting by target customer base is usually healthier than splitting by M365 versus AWS because customer-segment boundaries preserve a shared control plane, while vendor-stack boundaries tend to duplicate controls and then harden those seams into the architecture. [inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md] Explore work should sit in a thin enabling or incubation function near the hub, while exploit work should run on standardised shared rails with explicit governance, observability, and human-review rules.

Key Findings

[inference; source: https://www.capitalone.com/tech/ai/data-management/; https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-04-22-enterprise-ai-capability-model.md] High confidence. Most enterprises should organise multiple AI platforms around one central control plane, using Capital One's term for the central configuration, access, evaluation, observability, and policy layer, then expose separate products to users only where needs diverge materially.
[inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts] High confidence. A single unified platform team is strongest in the early stage because it concentrates scarce talent and standards, but it degrades into a slow intake queue if it continues owning every downstream use case after demand scales.
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai] Medium confidence. A split by target customer base often becomes the best scaling move because business users and developers need different workflows, support models, and product metrics even when they share one governed platform core.
[inference; source: https://www.melconway.com/Home/Conways_Law.html; https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Medium confidence. A primary split by M365 versus AWS is usually a weak default because it causes the organisation to duplicate policy, identity, retrieval, and support mechanisms, and then hardens those vendor seams into the architecture.
[inference; source: https://teamtopologies.com/key-concepts; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://github.com/davidamitchell/Research/blob/main/Research/completed/2026-02-28-exploit-explore-ai-portfolio-framework.md] High confidence. Explore activity should be owned by a small enabling or incubation function near the hub so that experiments with models, vendors, and evaluation methods can be productised once rather than rediscovered in parallel.
[inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] High confidence. Financial-services firms have stronger reasons than most sectors to centralise governance, approved patterns, and accountability, while decentralising only workflow configuration and user support that must stay close to the business.
[inference; source: https://cloud.google.com/resources/cloud-teams; https://teamtopologies.com/key-concepts; https://www.dbs.com/annualreports/2024/cio-statement.html; https://www.capitalone.com/software/blog/capital-one-you-build-your-data/] High confidence. The cleanest decision-rights split is central ownership of policy, vendor approval, data-access guardrails, observability, shared tooling, and cost attribution, with spoke ownership of prioritisation, integration, adoption, and benefit realisation.
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://www.melconway.com/Home/Conways_Law.html] High confidence. The most consistent anti-patterns are a permanent central team that owns all delivery, federated teams launched before a shared platform exists, and vendor-aligned teams that mirror suppliers instead of user journeys.

Evidence Map

Claim	Source	Confidence	Notes
[inference] Most enterprises should centralise the AI control plane and split only the customer-facing products that truly differ.	Capital One data management; Team Topologies; DORA 2025; Google Cloud cloud teams; AWS CCoE; DBS responsible AI; Enterprise AI capability model	high	Convergence across platform-team, cloud, bank, and prior repository capability evidence.
[inference] A single unified team is a strong starting shape but becomes a bottleneck if it owns all downstream delivery indefinitely.	AWS CCoE; Google Cloud cloud teams; Team Topologies	high	Strong analogy from cloud and internal platform design.
[inference] A customer-segment split is often the healthiest scaling move once business users and developers have materially different workflows.	Team Topologies; DORA 2025; Morgan Stanley OpenAI milestone	medium	User-centricity plus one concrete business-user platform example, but limited direct comparative evidence.
[inference] A vendor-stack split usually creates the wrong seams because it duplicates the shared platform core and embeds vendor boundaries in the architecture.	Conway's Law; Capital One data management; DBS responsible AI; DORA 2025	medium	Strong logic and corroboration, but not a controlled experiment.
[inference] Explore work should sit near the hub in an enabling function, while exploit work should move onto standardised shared rails.	Team Topologies; DBS responsible AI; DORA 2025; Exploit-explore AI portfolio framework	high	Strong agreement between organisational, operational, and prior repository portfolio evidence.
[inference] Financial-services firms need more central governance and accountability than most sectors, even when delivery is locally configured.	Bank of England DP5/22; J.P. Morgan AI in payments; DBS responsible AI; Morgan Stanley Debrief	high	Regulator guidance and bank disclosures align.
[inference] The cleanest decision-rights split is central policy and guardrails with spoke ownership of prioritisation, integration, and value capture.	Google Cloud cloud teams; Team Topologies; DBS CIO statement; Capital One You Build, Your Data	high	Repeated hub-plus-local pattern.
[inference] Permanent central-delivery ownership, premature federation, and vendor-aligned teams are the leading organisational anti-patterns.	Team Topologies; Google Cloud cloud teams; Conway's Law	high	Anti-patterns follow directly from recurring source warnings.

Assumptions

[assumption; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] Assumption: There is enough overlap across M365 and AWS control requirements that one shared governance and platform core is economically meaningful. Justification: if legal, residency, or customer separation is extreme, stronger structural separation could be warranted.
[assumption; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] Assumption: The enterprise has at least two distinct AI customer groups, business users and developers, whose workflow needs are meaningfully different. Justification: without that divergence, a customer-segment split may add cost without enough benefit.
[assumption; source: https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.capitalone.com/tech/ai/data-management/] Assumption: Public bank disclosures describe governance and platform patterns more often than exact team charts, so some operating-model conclusions are inferential rather than diagram-level factual. Justification: the public evidence is richer on principles and controls than on reporting lines and headcount.

Analysis

[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/resources/cloud-teams; https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html] The evidence was weighted toward sources that describe operating responsibilities directly, not toward generic strategy commentary, which made Team Topologies, AWS, Google Cloud, and bank disclosures more important than consultancy narratives. [inference; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.capitalone.com/tech/ai/data-management/] The strongest pattern across sources is central ownership of the reusable platform core plus local ownership of the workflow edge, so competing interpretations that favour either total centralisation or unmanaged federation were rejected as weaker fits to the evidence. [inference; source: https://www.melconway.com/Home/Conways_Law.html; https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] The main trade-off is between governance efficiency and local responsiveness, and the best way to manage that trade-off is to centralise the control plane while letting customer-facing teams optimise the experience for their user segment. [inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence; https://www.jpmorgan.com/insights/payments/security-trust/ai-payments-efficiency-fraud-reduction; https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch] Financial-services evidence was given additional weight because the sector's governance requirements are stricter and therefore expose accountability needs that more lightly regulated sectors can ignore for longer.

Risks, Gaps, and Uncertainties

[fact; source: https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/FEAT] The official MAS FEAT page was inaccessible from this environment, so MAS-specific governance content was checked for accessibility but not used as primary evidence.
[fact; source: https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html; https://www.capitalone.com/tech/ai/data-management/] Public enterprise disclosures describe principles and controls more often than exact reporting lines, so precise org-chart recommendations remain partly inferential.
[inference; source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report; https://teamtopologies.com/key-concepts] The recommendation against stack-splitting is strong but still inferential because there is no public controlled comparison of vendor-aligned and customer-aligned AI platform organisations.
[inference; source: https://www.capitalone.com/tech/ai/data-management/; https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/responsible-ai-in-banking-gaining-a-competitive-edge.html] The conclusion could change in an enterprise where M365 and AWS are separated by law, geography, or customer base strongly enough that shared governance and shared services are no longer economical.

Open Questions

[inference; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-center-of-excellence/introduction.html; https://cloud.google.com/resources/cloud-teams] What funding and chargeback model best supports a central AI hub without recreating a slow approval bureaucracy?
[inference; source: https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence] Which AI decisions in financial services should always remain with central risk and architecture functions, and which can safely be delegated under pre-approved guardrails?
[inference; source: https://teamtopologies.com/key-concepts; https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report] What leading indicators show that a unified team has reached the point where a customer-segment split is justified?