Why the CoE Era Is Quietly Ending

When large pharmaceutical companies first stood up dedicated AI groups in 2021 and 2022, the rationale was straightforward. There were not enough machine learning engineers on the open market, business units did not know what “good” looked like in an AI proof of concept, and the compliance apparatus around GxP-relevant models was being built almost from scratch. Consolidating talent, tooling, and governance under a single Center of Excellence protected against duplication and gave the enterprise a defensible answer when regulators asked who owned model risk. McKinsey’s 2024 research on generative AI operating models found that more than fifty percent of enterprises had adopted a centrally-led model for generative AI even when their broader data and analytics function was already decentralized, precisely because early-stage AI felt too specialized to trust to line functions.1

Four years later, three things have changed at once. First, the technology has become dramatically easier to consume. Enterprise AI platforms, low-code AI builders, and embedded copilots inside Microsoft 365, Salesforce, Veeva, and Benchling now let non-specialists build capabilities that previously required a data science squad. Gartner has projected that by 2026 developers outside the formal IT organization will account for at least eighty percent of the user base for low-code development tools, and enterprise AI app builders are actively driving that citizen-developer curve upward.2

Second, the CoE has become a bottleneck. Practitioners writing across enterprise AI communities describe the same pattern: approval delays and knowledge bottlenecks emerge as AI experts inside the CoE cannot support every team; product teams and the CoE debate priorities instead of shipping; and the central team’s distance from frontline operations produces solutions that lack the domain depth clinical, regulatory, or commercial owners actually need.3 One widely cited enterprise AI analysis put it bluntly: a CoE without a defined scope becomes either a bottleneck because it tries to own everything, or an irrelevant advisory body because nobody is compelled to follow its guidance.4

Third, regulators are re-anchoring accountability at the point of use. At Pharma Meets AI in Barcelona in April 2026, industry commentary highlighted a decisive shift from standalone AI governance functions to embedded, operational governance that sits inside business processes and is owned by business function leaders.5 The FDA’s evolving position on AI in pharmaceutical quality systems and EMA guidance both push accountability toward the process owner rather than a distant AI office. When the quality unit is required to attest to an AI system’s fitness for purpose in a GxP-relevant workflow, the CoE cannot be the signing officer.

50%+ of enterprises adopted a centrally-led gen AI model in the early phase, per McKinsey research1
80% of low-code platform users projected to sit outside formal IT by 2026, per Gartner2
30% governance readiness among enterprises already deploying AI, per Deloitte’s 2026 State of AI in the Enterprise6

None of this means the CoE was a mistake. Central teams built the platform standards, the governance vocabulary, the MLOps templates, and the responsible AI rubrics that made federation possible in the first place. The mistake is treating the CoE as permanent infrastructure rather than a phase of a maturity curve. The rest of this article is about recognizing the phase transition, planning it deliberately, and preserving what should not be lost.

Six Signals It Is Time to Sunset Your AI CoE

A federation decision is rarely a single event. It is a cluster of small frictions that accumulate until leadership realizes the organization has quietly outgrown its original model. Across published enterprise AI operating-model analyses and our own observations working with life-sciences clients, six signals surface repeatedly.3 7

1. Business functions are requesting their own AI teams

Commercial analytics stands up a data science pod. Medical Affairs hires an insights engineer. Manufacturing quietly funds a plant-level modeling role. When multiple functions are separately building shadow capability rather than routing work to the CoE, the market has spoken. Treat this as a leading indicator, not an act of defiance. The functions are telling you that domain proximity beats central speed for the problems they now care about.

2. The CoE has become a queue rather than a capability

The most reliable operational signal is intake congestion. When new project requests are triaged into slots eight, twelve, or eighteen months out, and when demand consists mostly of variations on solved patterns rather than genuinely novel problems, the CoE is functioning as a throughput constraint. Assembly’s operating-model analysis warns that choosing the wrong AI operating model structure costs enterprises twelve to eighteen months of rework, and staying centralized past the natural inflection point is one of the most common versions of that mistake.3

3. The model portfolio has outgrown central bandwidth

Deloitte and other operating-model observers have noted that the hub-and-spoke structure becomes the right answer when the portfolio exceeds roughly fifteen to twenty active initiatives, at which point centralizing execution creates more coordination cost than it saves.6 Above that threshold, a purely central CoE cannot maintain quality, oversight, and pace simultaneously. Something breaks. Usually it is quality first, then pace, then talent retention.

4. Foundation models and platforms have commoditized the core

Techniques that once justified a dedicated PhD-level team are now available inside Veeva Vault workflows, Benchling apps, Snowflake Cortex, Salesforce Einstein, and Microsoft 365 Copilot. When the answer to “why do we need the CoE for this?” is increasingly “we don’t, we just need a business analyst who knows what good looks like,” federation is overdue.

5. Governance load is exceeding technical work

Deloitte’s 2026 State of AI in the Enterprise report shows that governance readiness sits at only thirty percent among companies already deploying AI, compared to forty-three percent for technical infrastructure and forty percent for data management.6 Where the CoE originally spent most of its time on model development, mature CoEs increasingly spend the bulk of their time on policy, third-party model review, evaluations, and audit response. That work is real, but it is governance work, not build work, and a governance office does not need to hoard build capacity to be effective.

6. The talent economics are inverting

Central AI teams that no longer produce visible business outcomes lose their most valuable engineers first. When your senior data scientists start taking rotations into business functions or leaving for smaller shops that offer more delivery agency, the CoE has become a career trap rather than a career accelerant. Federation, done thoughtfully, is often what retains that talent by moving them closer to the problem.

Watch for false positives. None of these signals in isolation justifies wind-down. A single grumpy business unit demanding its own team is not evidence of CoE failure. What matters is the pattern: three or more of the six signals present at once, sustained over two or more quarters. Federation off a single quarter’s frustration is how organizations lose institutional AI knowledge and end up rebuilding the CoE eighteen months later under a different name.

The CoE-to-Federation Maturity Model

Federation is not a binary event. It is a staged transition, and the same CoE can be sunset gracefully or catastrophically depending on which stage the enterprise is actually operating in. McKinsey’s operating-model work describes a progression in which companies choose to progressively increase domain teams’ ownership of generative AI development, moving from centralized to federated and finally to decentralized structures.1 Our five-stage view synthesizes that progression with what we see in pharma and biotech engagements.

STAGE 1

Exploratory

No CoE yet. Scattered pilots inside individual functions, no shared standards, no coherent governance. Typical of pre-2021 pharma AI. Move to Stage 2 by consolidating.

STAGE 2

Centralized CoE

Dedicated central AI team owns strategy, delivery, MLOps, and governance. Reports to CIO, CDO, or CEO. Correct answer when portfolio is under ten active initiatives and business units lack maturity.

STAGE 3

Hub-and-Spoke

Central hub retains governance, platform, and standards. Business functions run their own spokes with embedded engineers. The transitional stage where most pharma CoEs sit in 2026.

STAGE 4

Federated

Business functions own delivery end to end. A slim central team retains governance, platforms, and standards but no build capacity. Federation is complete but coordination is still explicit.

STAGE 5

Embedded

AI is a normal capability inside every function, indistinguishable from analytics or engineering. Governance rides on standard platforms. The “CoE” no longer exists as an identifiable unit.

GUIDANCE

Where most pharma sits today

Large pharma is generally at Stage 3 with residual Stage 2 muscle memory. Mid-size biotech is generally at Stage 2 aspiring to Stage 3. Sunset conversations belong at the Stage 3-to-4 transition.

Two cautions on the maturity model matter. First, life-sciences maturity research on federated infrastructures identifies six discrete domains that must mature in parallel: governance and policies, operations and performance, data management and security, operational infrastructure, clinical or research infrastructure, and outreach and communication.8 A CoE that has advanced technical capability but immature governance is not ready to federate; it is ready to be corrected. Second, published work on operating-model transitions warns that operating-model design is the beginning of a maturation process that unfolds over two to four years and that transitions between stages are where programs most commonly stall or regress.9 Sunset is a two-year project, not a two-quarter project.

How to self-diagnose your stage in one afternoon. Interview five business function leaders and ask three questions: Who do you go to when you want an AI capability built? Who signs off that the model is fit for purpose in your workflow? Who is on the hook if the model produces a bad outcome? If all three answers point at the CoE, you are at Stage 2. If build points at the CoE but sign-off and accountability point at the function, you are at Stage 3. If all three point at the function, you are already federated whether the org chart admits it or not.

What Stays Central: The Retained Function After Federation

The most consequential design decision in a CoE sunset is not what to move but what to keep. Federation is not the elimination of central capability. It is the deliberate concentration of central capability on the things that genuinely benefit from enterprise scale, and the deliberate transfer of everything else. Microsoft’s Cloud Adoption Framework guidance on AI Centers of Excellence and analyses of platform-team transitions consistently identify five retained functions.10 11

Retained Central Function What It Does After Federation Why It Cannot Federate
AI Governance Office Owns responsible AI policy, model risk framework, external regulatory engagement, AI inventory and register, executive AI committee secretariat. Consistency across functions and single point of accountability to the board and regulators.
Platform Engineering Runs enterprise AI platforms (MLOps, evaluation harnesses, agent frameworks, LLM gateways, prompt libraries, RAG infrastructure, feature stores). Economies of scale; avoids each function reinventing infrastructure that has already been solved.
Standards and Reference Architecture Publishes reference patterns, evaluation criteria, model cards, data contracts, and reusable components for functions to consume. Prevents divergence that becomes an audit and integration nightmare within eighteen months.
Third-Party and Vendor AI Review Screens vendor AI capabilities before contracting, maintains an approved-vendor list, negotiates enterprise licenses for foundation models and platforms. Purchasing leverage and consistent security, IP, and data-use review across the enterprise.
Talent and Community Runs guilds, chapters, and communities of practice; owns AI literacy curricula; manages rotational programs and career pathing across functions. Talent development benefits from cross-function scale; individual functions cannot sustain career depth alone.

What is conspicuously absent from that list: model development, use-case ideation, business analysis, deployment, and operational support. Those move to the functions. A well-designed retained-central-function typically has between eight and twenty-five people at a large pharma, compared to fifty to two hundred at the peak of a centralized CoE. The dramatic headcount reduction is the point. Central mass is expensive without being federated. Concentrated central specialists work.

The SD perspective. Most pharma organizations we work with underestimate how much of what the CoE currently does is actually reviewable, teachable, or platform-mediated work rather than genuinely central expertise. When you honestly categorize CoE activity into “must be central,” “should be a platform,” and “belongs in the function,” the “must be central” bucket is usually smaller than leadership expects, and the platform bucket is usually larger than the platform team expects. That is the intellectually honest starting point for a federation plan.

The Twelve-Month Federation Playbook

Sunsetting a CoE well is closer to a controlled decommissioning than a reorganization. Announced abruptly, it destroys morale, scatters talent, and leaves in-flight models orphaned. Executed over roughly twelve months with a phased structure, it can preserve delivery pace, transfer capability cleanly, and hand credible governance to the business functions. The following playbook is drawn from published operating-model transition patterns and adapted for the regulatory realities of pharma.9 11

1

Months 1-2: Diagnosis and Declaration

Confirm the maturity stage. Interview business function leaders. Audit the model portfolio. Map every current CoE activity to “central,” “platform,” or “function” buckets. Communicate the transition openly to CoE staff before rumors form. The single most damaging thing leadership can do is let the CoE learn about its sunset from a slide deck.

2

Months 2-4: Design the Target State

Define the retained central function (governance, platform, standards, vendor, community). Define which functions receive federated capacity, in what shape, and reporting where. Redesign the enterprise AI committee to reflect the new accountability. Draft the operating agreement between retained function and federated teams.

3

Months 4-6: Transfer Model Ownership

Every production model gets an assigned business owner and a functional custodian. In-flight builds are triaged: complete with CoE, transfer to function, or terminate. Documentation and evaluations are brought to a standard that a receiving function can maintain. GxP-relevant models receive extra scrutiny to preserve audit trail continuity.

4

Months 5-8: Talent Placements

Match CoE staff to receiving functions based on domain fit and career preference. Publish placements publicly with clear job descriptions. Retention conversations happen simultaneously for the retained central team. Do not force placements that neither the individual nor the receiving function wants; that path leads to attrition on both sides.

5

Months 6-9: Platform Handoff

The MLOps stack, LLM gateway, evaluation harness, and shared components move to platform engineering ownership. Service-level agreements are written between platform and function teams. Cost allocation and chargeback models replace unlimited free consumption.

6

Months 8-11: Governance Recut

Responsible AI policy is republished in the federated context. AI inventory is centralized in the retained function but populated by federated teams. Regulatory sign-off pathways are redrawn to route through function quality units with retained-office consultation. Audit response drills are rehearsed under the new model.

7

Months 10-12: Sunset and Reflect

The formal CoE ceases to exist as a named organizational unit. A retrospective is conducted with function leaders, retained central staff, and the executive AI committee. The community of practice takes on the cross-function knitting that the CoE used to do informally. The retained function publishes its year-one operating plan.

The GxP handoff is the highest-risk step. A CoE that owns validated models within a GxP-relevant workflow cannot simply transfer them to a function without deliberately re-establishing the validation chain, the intended-use documentation, and the change-control ownership. Regulatory inspectors will ask who owns the model, who validated it, and who is competent to change it. If the CoE is the answer today and the function is the answer tomorrow, that transition must be documented, evidenced, and defensible. Do not sunset the CoE faster than the quality management system can catch up.

Capability Transfer: Where the Talent Goes

Every conversation about CoE sunset eventually becomes a conversation about people. The instinct in most reorganizations is to protect roles and preserve reporting lines. That instinct works against federation. The clearer honest answer is: some people belong in functions, some belong in the retained central team, some belong on the platform, and some may not belong at the enterprise at all in its post-federation shape. Being direct about that from the beginning is kinder than pretending.

Where each CoE role tends to land

Original CoE Role Typical Federation Landing What Changes for the Individual
ML/AI Engineer Embedded in business function Deeper domain immersion, tighter delivery cadence, dotted line to platform team.
Data Scientist Embedded in business function Same as above; often paired with a business analyst.
MLOps / Platform Engineer Retained platform engineering team Explicit service orientation; SLAs; internal customers rather than internal projects.
Responsible AI Lead Retained AI governance office Role shifts from advisor-per-project to policy owner and audit lead.
AI Solution Architect Retained standards team or platform Fewer projects, more reference architecture and reusable pattern authorship.
AI Program Manager Business function or retained governance Portfolio coordination shifts from central intake to cross-function community facilitation.
Vendor / Partnerships Lead Retained third-party review team Concentration on vendor screening, enterprise licensing, and community of practice for buyers.
Head of CoE Chief AI Officer, retained function head, or transition role The most delicate placement; ideally becomes the head of the retained function or CAIO. The wrong answer is a lateral into an ambiguous role that leaves them without a mandate.

A well-run federation typically preserves seventy to eighty percent of CoE headcount somewhere in the enterprise. Ten to twenty percent naturally exit during the transition because the shape of the work no longer suits them. That is not a failure; that is the transition working. The failure mode is the reverse, where the wrong people stay and the right people leave.

The community of practice replaces the CoE as connective tissue

Once engineers are distributed into functions, cross-pollination has to be engineered deliberately. Every mature federated organization runs a community of practice or guild structure: monthly practitioner forums, shared code and prompt libraries, quarterly show-and-tell, and rotational placements. The retained central team hosts these gatherings but does not own the agenda. Practitioners set what they need to talk about, and the retained office removes friction. This is where the informal knowledge that used to travel through CoE hallway conversation gets re-created.

Communities of practice in a federated model are less about lectures and more about problem-solving with peers. A common cadence is a monthly forty-five-minute forum where two practitioners each present a real problem, an audience of fifteen to thirty debate the solution, and the outputs go into a shared knowledge base that platform engineering treats as backlog for the next quarter. Rotational placements of six to nine months let a data scientist in Commercial spend time inside Regulatory or Manufacturing, which is how domain knowledge becomes portable rather than trapped in function silos. Executive support for these rituals matters more than budget. A community of practice with a small stipend but visible leadership sponsorship consistently outperforms a well-funded community whose members quietly disengage because the CFO or CIO never shows up.

Compensation, career pathing, and the sunset conversation

One under-discussed dimension of federation is career pathing. In a centralized CoE, senior engineers had a clear promotion track inside a growing central organization. In a federated model, that track fragments across functions. Without deliberate design, ambitious engineers reach a ceiling in their function and leave. The retained office should own an enterprise-wide AI career framework, define principal and distinguished engineer tracks that span functions, and offer explicit paths from function into the retained office and back. Federation without a career framework quietly bleeds senior talent for the eighteen months after sunset, and by the time leadership notices, the replacement pipeline is thin.

Risk Management During Sunset

Federation done badly produces four failure modes that show up in the literature and in practice with disappointing regularity. Each one is manageable if leadership names it out loud in the planning phase.

Duplication and drift

Observers of federated operating models have consistently warned that organizations moving immediately to federated structures without staging report higher duplication and more compliance incidents than those that transition through a hub-and-spoke intermediate stage.6 Two functions build overlapping capabilities. Standards diverge. Evaluations become non-comparable. Governance drifts silently until an audit exposes it. The countermeasure is explicit shared standards, a shared platform, and a retained office with real teeth on reference architecture and reusable components.

Governance vacuum

If the CoE was the de facto governance function and federation is not accompanied by a genuine retained governance office, the enterprise ends up with less rigorous oversight than before rather than more distributed rigor. The industry conversation around AI governance 2.0 has moved firmly toward integrated assurance that embeds risk and compliance structures directly within business functions supported by continuous monitoring and clearly defined operational accountability.5 Federation without that scaffolding is not federation; it is abandonment.

Institutional knowledge loss

CoEs accumulate tacit knowledge: which vendors ghosted us in 2023, which architecture patterns did not survive contact with commercial data, which regulatory reviewer prefers a particular evidence format. If people leave during the sunset without deliberate knowledge transfer, that context walks out the door. Document what can be documented, rotate people through overlap periods deliberately, and hold retrospectives before the CoE is dissolved rather than after.

Executive re-centralization instinct

Twelve to eighteen months after federation, when the first governance incident happens or the first function overspends on foundation model calls, executive instinct is to rebuild the CoE. Resist. The correct answer is almost always a strengthened retained function, better platform guardrails, or a clearer policy, not a return to centralized delivery. Recentralization after federation is unusually expensive because talent has re-anchored in the functions and will not cheerfully move back.

The healthy end state. Eighteen months after sunset, the retained central function is smaller than the original CoE, spends most of its time on governance, platforms, standards, vendors, and community rather than delivery, and is invisible enough to the business that only the CIO and quality unit notice it exists. AI capability shows up on the org chart under Commercial, Medical, Regulatory, Quality, Manufacturing, and R&D. The community of practice runs monthly. Audits pass. Model velocity has increased. Nobody is nostalgic for the CoE.

Case-Study Patterns from Life Sciences

Individual company organizational designs shift constantly, and specific engagements are usually confidential, but three cross-industry patterns show up consistently in published life-sciences work and in operating-model observations from major consultancies. Each is instructive for pharma leaders sizing up their own transition.

Pattern 1: The R&D-Anchored Federation

Some large pharma companies have led federation from the R&D side because that is where AI investment concentrated first. Sanofi’s public discussion of its Modulus manufacturing platform, its internal AI council, and its embedded AI capability across R&D and manufacturing illustrates the pattern of running dedicated capability inside business domains while a central council governs policy and portfolio.12 The retained central mass is deliberately small, and business function ownership is explicit. IQVIA’s own trajectory with the IQVIA.ai platform, which by launch in March 2026 had over 150 internal AI agents adopted by 19 of the top 20 pharma companies, reflects a related pattern on the vendor side: shared platform infrastructure enabling function-level consumption without every consumer standing up its own AI engineering group.13

Pattern 2: The Federated Data Substrate

Life-sciences federation has been driven partly by data physics. A single human genome is roughly two hundred gigabytes, and multi-omic data at population scale is too massive to move and too sensitive to share, which forces distributed processing.9 Federated learning approaches, most visibly Eli Lilly’s Lilly TuneLab initiative using NVIDIA FLARE to let partners benefit from Lilly’s models without sharing raw data, illustrate how the technical architecture nudges the organizational structure toward federation. When data must stay where it is, so must capability, and central CoEs cannot pretend otherwise.14

Pattern 3: The Governance-First Federation

The most successful transitions we observe start by strengthening governance before touching delivery. The retained AI governance function is stood up as a distinct organizational unit with an explicit mandate, published policies, and an inventory of every deployed model. Only once that governance infrastructure is in place does the delivery-side federation begin. This is the inversion of the common instinct, which is to federate delivery first and figure out governance later. The BCG-observed correlation between hub-and-spoke maturity and higher realized AI value, combined with Deloitte’s finding that governance readiness lags technical infrastructure, suggests governance-first is the safer sequencing.6 15

Pattern 4: The Vendor-Consolidation Trigger

A quieter but powerful trigger for federation is a major vendor consolidation. When a pharma company standardizes on an enterprise AI platform, an LLM gateway, and a shared evaluation harness, the technical justification for a large central team evaporates. The platform does what the CoE used to do in code, and the specialist skill required to consume that platform inside a business function drops sharply. Practitioner analyses of enterprise AI platforms consistently note that platform maturity and CoE necessity are inversely related: once the platform team can offer self-service infrastructure to functions, the CoE ceases to be the delivery gate.16 Pharma leaders sizing up federation should look at platform investments as accelerants of the sunset conversation rather than parallel to it.

Pattern 5: The Agentic Reframing

McKinsey’s work on the agentic organization argues that the next paradigm of AI in the enterprise is not a centrally managed model portfolio but a distributed fabric of AI agents embedded in workflows, coordinated by shared standards but built and operated where the work lives.17 If this framing proves right, the CoE model becomes structurally obsolete rather than merely fatigued, because agents that live inside a clinical operations workflow or a manufacturing MES cannot be built or governed by a distant central team without unacceptable latency and context loss. Leaders whose AI roadmaps are shifting toward agentic architectures should treat that shift as an additional reason to accelerate federation planning, not delay it.

A note on borrowing from tech. Life-sciences leaders sometimes look at how technology companies federate AI and want to copy the model. The mechanics translate, but the governance context does not. Pharma federation must survive a regulatory inspection that a consumer tech company will never face. Every playbook borrowed from Big Tech needs a GxP overlay, a change-control conversation, and an audit-trail plan before it is safe to run.

What the case patterns share

Across all five patterns, three constants show up. First, governance intensifies during federation rather than relaxing. Second, the retained central function is smaller but more senior than the original CoE. Third, the transition is driven by a combination of technology maturity, portfolio scale, and business-function readiness rather than by any single trigger. Leaders looking for a clean external justification to sunset a CoE will not find one. The justification is always cumulative and always partly a judgment call. What the patterns provide is confidence that others have made the same call and lived to talk about it.

Conclusion

The AI Center of Excellence was the right answer to the problem pharma faced in 2021 and 2022. It is increasingly not the right answer to the problem pharma faces in 2026. Foundation models are commoditizing, tools are embedded, business functions are demanding delivery agency, regulators are anchoring accountability at the point of use, and the CoE queue is the single most cited reason interesting work does not ship. Federation is not the failure of the CoE model; it is its natural graduation. The organizations that plan the sunset deliberately, over twelve months, with a clear retained central function, preserve their governance rigor, retain their best talent, and hand credible capability to the business functions that own the P&L. The organizations that either cling to a centralized model past its usefulness or federate abruptly without a retained governance office both end up in the same place eighteen months later: rebuilding under a new name, with less trust and more scars.

Sakara Digital works with pharma and biotech organizations thinking about the operating model behind their AI capability. If you are looking at your current CoE and asking whether it is time to stage a federation, or whether you are already federated in practice and need to formalize it, we are happy to have that conversation. Independent perspective is often what the internal debate is missing.