Marcus Webb, a deputy CIO at a 4,000 person state human services agency, had budget for an AI Center of Excellence and a one-line directive from his secretary: stand it up by the third quarter. Six months later he had hired three data scientists, bought a vendor platform, and produced exactly one finished product, a chatbot pilot that two program offices refused to adopt. The CoE was technically alive and functionally orphaned. His mistake was not technical. He had built a team before he had built a mandate, a charter, or a single internal customer who felt they owned the outcome.

Marcus's story is common because "Center of Excellence" sounds like a hiring problem and is actually a governance problem. This lesson treats the CoE as something you charter, fund and measure, not something you staff and hope. We will build one model end to end, using the federal experience as the reference point because it is the most documented government example available, and because the failure modes it has already produced are the ones your agency is most likely to repeat.

What a CoE actually is

An AI Center of Excellence is a small central team that makes AI adoption faster and safer for the rest of the agency. It does this by concentrating scarce skills, setting shared standards, providing shared infrastructure, and shepherding the highest-value projects through governance. It is not a feature factory that builds every model in the building. The moment a CoE becomes the only place AI work happens, it becomes a bottleneck and dies of its own backlog, usually while its sponsor is explaining the queue to a legislative committee.

The distinction that matters is between capability the agency owns and capability the CoE owns. A CoE that succeeds leaves behind program offices that can specify, commission, evaluate and operate AI work with the center's support. A CoE that fails leaves behind a dependency. That difference shows up in the charter long before it shows up in the delivery record, which is why the charter is where this lesson starts rather than where it ends.

Why concentrate capability at all

There are five systematic reasons to concentrate AI capability rather than distribute it. First, AI talent is scarce and expensive, and a central pool can be deployed across use cases that no single office could justify staffing alone. Second, AI infrastructure such as training environments, machine learning operations platforms and governed data stores has high fixed cost and low marginal cost, so sharing amortizes it. Third, governance artifacts including risk management policy, evaluation standards, monitoring, model documentation and impact assessments are reused across use cases and should be produced once.

Fourth, uncoordinated pilots carry real risk: duplicated vendors, inconsistent security postures, divergent privacy reviews, conflicting public messages, and policy arbitrage between components that shop for the most permissive reviewer. Fifth, oversight expectations from Congress and from central budget authorities are increasingly articulated at the agency level, so a component-only approach produces expensive coordination at every touchpoint. The federal accountability framework published by the Government Accountability Office in June 2021 highlighted these benefits and noted that agencies with central AI capability mature their governance faster.

Concentration has its own risks, and naming them in the charter is more honest than discovering them in year two. A center becomes a bottleneck when demand exceeds its capacity. It drifts away from operational business problems when it stops sitting with the people who have them. And it develops a tendency to pursue technology for its own sake, because that is what the staff enjoy and what conference audiences reward. The mitigations are embedded staff who deploy into components, real product management discipline, and success criteria tied to mission outcomes rather than to deployments.

Where the model comes from

The federal Center of Excellence pattern predates AI. A Centers of Excellence program was launched at the General Services Administration in 2017 under the Office of American Innovation, across six focus areas: cloud adoption, infrastructure optimization, customer experience, contact centers, data and analytics, and artificial intelligence. It began with the Department of Agriculture as the pilot partner and expanded to housing, personnel, labor, energy, environmental, emergency management and defense organizations. The operating shape was federal employees leading, contractor teams augmenting, and the receiving agency funding the work through interagency agreements.

That model borrowed in turn from the digital service teams stood up in 2014, which themselves borrowed from the United Kingdom's Government Digital Service, established in 2011. The AI-specific flavor emerged around 2019 under the federal data strategy and accelerated when central guidance required agencies to designate chief AI officers. Agencies now reach for the CoE pattern to concentrate expertise, share infrastructure, avoid duplicate pilots and harmonize risk management. Named programs are reorganized and renamed frequently, so confirm the current status of any body you plan to model yourself on.

The federal variants are worth studying because they are different answers to the same question. A national AI institute at the Department of Veterans Affairs, stood up in 2019, consolidates research and application across the health system and publishes peer-reviewed work on clinical problems including suicide prevention, cardiology and radiology. The Chief Digital and Artificial Intelligence Office at the Department of Defense was created in February 2022 by merging the Joint AI Center, the defense digital service and the chief data officer function, and runs intelligence, data platform, generative AI and shared services activities. Energy concentrates AI for science around its supercomputing laboratories. Health and Human Services uses a council to coordinate its operating divisions.

Other governments have reached for the same shape. The United Kingdom placed an AI incubator inside the Cabinet Office alongside its central digital and data function, and Singapore locates its AI center inside its government technology agency. The consistent pattern across all of these, domestic and foreign, is that hub and spoke with strong central infrastructure and strong component domain teams outperforms both the purely central and the purely federated arrangement. That is a pattern worth borrowing, not a guarantee: the same structure fails whenever the central body has infrastructure but no authority, or authority but no delivery capacity.

Three operating models

Three models exist, and choosing the wrong one is the most expensive early mistake an agency makes. The choice should follow agency size, mission diversity and current maturity, and it is normal for the answer to change: the federal pattern is that agencies expect to evolve the model over three to five years rather than pick one permanently. Write the current choice and its rationale into the charter so that the next reorganization argues with a documented decision rather than with a rumor about what the center was supposed to be.

• Central. The CoE executes AI work directly on behalf of components. All AI staff report to the CoE lead and infrastructure is owned centrally. Advantages: full consolidation, maximum governance coherence, simplest accounting. Disadvantages: distance from domain context, queuing delays, and loss of component ownership.
• Federated. Each component runs its own AI team with light coordination from a small central office. Advantages: domain proximity and component accountability. Disadvantages: duplicated effort, inconsistent governance, and difficulty attracting strong talent to any single component.
• Hub and spoke. The center provides shared infrastructure, governance frameworks and talent rotations while components run domain teams that plug in. Advantages: balances consolidation with domain depth and supports local ownership. Disadvantages: requires sophisticated operating model design and becomes confusing when authorities are unclear.

Most mature federal AI programs converge on hub and spoke, and the pattern is visible wherever a central AI office coexists with component AI teams that have their own domain staff. The reason is not ideological. It is that the central body can afford the platform and the governance machinery while the components hold the mission knowledge, and neither one can deliver a working system without the other. A coordinating body with no build capacity at all is cheap and politically safe, and usually too weak to change anything.

The charter comes first

Before a single hire, write a charter the agency head signs. The charter is the difference between a team and a mandate. It answers five questions in plain language: what problems is this CoE allowed to work on, what decisions can it make without going up the chain, who funds it and for how long, who does it serve and how do they request help, and how will we know in twelve months whether it worked. Every one of those questions becomes a fight later if it is not answered now.

The fuller version specifies mission and theory of change; scope, including whether national security systems are in or out; authorities delegated by the deputy secretary or the chief AI officer, covering decision rights over intake, governance approvals and capability investment; the service catalog across advisory, delivery, governance, infrastructure and workforce development; the funding model; the staffing model; and the measurement approach. Authorities have to be concrete enough to avoid governance by permission, the arrangement in which the center can see everything and change nothing.

Funding in government is usually a combination: a direct appropriation where the legislature creates a line item, a working capital fund assessment on the components that consume CoE services, and project-specific interagency agreements for cross-agency work. Federal modernization funds have backed several AI center initiatives. Whichever combination you use, name the year one and year two figures and the review date in the charter itself, because the funding surprise in year two is what kills centers that survived their first year on enthusiasm. A first-year budget for a small hub-and-spoke center of the size described below, including a shared tooling environment and modest contractor support, runs roughly $1.8M to $2.6M depending on whether engineers are staff or contractors.

The charter also fixes the relationships that would otherwise be negotiated per project. Where the center does not cover national security systems, say so and name who does. Define the working relationship with the chief information officer, the chief information security officer, the chief data officer and the senior agency official for privacy. Civil rights partnership with the agency's civil rights office or general counsel is mandatory for rights-impacting AI, and the charter is where that partnership is established rather than requested.

Staffing the minimum viable CoE

You do not need twenty people. A credible starting center is five to seven roles, and at least two of them are not technical build roles. The list below counts six, which sits inside that band; the point of the band is that you can start at the bottom of it and add the seventh role once the intake pipeline tells you which kind of work is actually arriving. Do not staff to a ratio you read somewhere. Staff to the pipeline you can already see, and revisit the number at the charter review date.

• CoE lead. A respected insider who can navigate procurement, budget and politics. Technical literacy matters more than technical depth.
• Two AI or machine learning engineers. The people who can build and evaluate models, or rigorously oversee a vendor doing so.
• A data engineer. Most government AI projects fail on data plumbing, not algorithms. This is the most underhired role in the building.
• A product or program manager. Owns the intake pipeline, prioritization, and the relationship with program offices.
• A governance and risk specialist. Connects every project to your risk management process and compliance obligations so that risk review is built in rather than bolted on.

As the center grows, the federal pattern fills out four groups rather than one. Technical leadership adds a lead data scientist and a machine learning operations lead. Applied staff adds statisticians, user experience researchers and human factors specialists alongside the engineers. Governance staff adds a model risk officer, privacy and civil rights liaisons, an evaluation lead and a red team lead. Operations adds program management, communications and someone who owns the cost of the compute. Which of these you need first is a function of your portfolio, not of an external standard.

Hiring in government is slow. Plan for details from program offices, term appointments and contractor augmentation in the first year while permanent hires clear, and use the direct hire and mobility authorities your human capital office can support. Contract staff supplement federal capacity but cannot substitute for federal decision-makers, and a center whose judgment lives inside a contract loses that judgment when the contract ends. Culture matters as much as headcount: traditional IT organization design rewards predictability, while AI work requires tolerance for failed experiments, publication of negative results and genuine peer review.

The specific instruments are worth knowing by name before you need them. Direct hire authority for data science roles applies under either the data science or the information technology occupational series depending on how the position is written. Intergovernmental Personnel Act assignments bring academic researchers in on temporary detail. Expert and consultant appointments cover short, specialized engagements. Excepted-service hiring flexibilities apply where the position and the authority match. Salary caps remain the binding constraint at the top of the range, which is why senior technical leadership discussions usually end up involving executive-service positions, special pay rates for information technology occupations, and retention incentives rather than base pay adjustments alone.

Retention risk is real, because mid-career data scientists have outside options and know it. Visible impact, technical challenge and a strong learning culture are the levers most consistently cited, and they are worth building deliberately. They are not a guarantee: a practitioner who cannot get a development environment approved will leave a center with an excellent learning culture, and no amount of interesting work compensates for a promotion path that ends at the first supervisory grade.

The intake pipeline

The single behavior that separates Marcus's orphaned center from a working one is a real intake pipeline. Program offices propose problems, the center scores them, picks a few, and says no to the rest in writing. Saying no is the product. A center that takes every request becomes a help desk, and a center that decides by favor or by whoever asked loudest destroys its own credibility in a single budget cycle. Publish the criteria, publish the decisions, and let the record be inspected.

The standard pipeline has five stages. Discover is a light-touch submission capturing the problem statement, the owner, the expected beneficiaries, the timing and the rough scale. Evaluate applies the screening criteria. Plan develops the project charter, performs the mapping stage of your risk framework, completes privacy and civil rights review, and drafts acceptance criteria. Deliver executes with iterative demonstrations and measurement against those criteria. Operate transitions the system into sustained operation with monitoring, incident response and revalidation triggers.

Score each candidate on mission value, feasibility with the data you actually have, risk category under your governance guidance including whether the use is rights-impacting or safety-impacting, data readiness, replicability across other offices, and whether the center has the capacity to take it now. Replicability matters more than people expect: a project that solves one office's problem is worth less than one that produces a reusable pattern ten offices can adopt. Mature centers also publish kill criteria, so that a use case which fails evaluation or loses mission relevance is closed cleanly rather than starved quietly.

Aligning with the wider governance system

A center is one node in an agency's AI governance system, not the whole of it. Central guidance for federal agencies requires each covered agency to designate a chief AI officer, convene an agency AI governance board, publish an annual inventory of AI use cases, and implement minimum practices for rights-impacting and safety-impacting uses. The risk management framework published by the National Institute of Standards and Technology defines govern, map, measure and manage functions; it is voluntary and non-binding, and its value is as a common vocabulary that survives changes in policy above it.

The discipline here is subtraction. Map each charter activity to the corresponding framework function and policy practice, then eliminate anything the agency already does elsewhere. If an existing investment review board already examines IT spending, the center should embed AI-specific criteria into that review rather than create a parallel one. Privacy review runs through the senior agency official for privacy, security through the chief information security officer, and civil rights through the civil rights office or general counsel, with the center supplying the AI-specific analysis those officials need to do their jobs rather than substituting its own judgment for theirs.

Shared infrastructure and the platform question

A credible center provides infrastructure that components could not justify individually. At minimum that means an authorized cloud environment appropriate to the sensitivity of the data, training and serving platforms, a governed data store and feature store with lineage, a model registry with versioning, experiment tracking, evaluation harnesses including bias and fairness tooling, monitoring and drift detection, and integration with incident management. Federal authorization processes have been modernizing to keep pace with AI services, and gaps remain, which is why agencies commonly operate cutting-edge services under provisional authorizations while the full package catches up.

Two disciplines keep the platform from becoming the problem. The first is explicit avoidance of vendor lock-in: data formats should be portable, models should be exportable, and the center should not become contractually dependent on a single provider for the capability its charter promises. The second is honest total cost of ownership, which includes compute, storage, the human time to maintain the platform, and the governance tooling that is almost always underestimated because it is bought last. Classified environments add cross-domain requirements that lengthen every one of these timelines.

How you prove it worked

Measurement is where most centers fail, usually by reporting what they did rather than what changed. Measure three layers and report them to your sponsor on a fixed cadence. Output covers projects shipped to production, reusable assets published and staff trained. Adoption covers program offices actively using the center's products and patterns, where a pilot nobody adopts counts as zero. Outcome covers hours returned to mission work, backlog reduced, errors caught and cost avoided, always translated into the unit the sponsor cares about: cases cleared, citizens served, days saved.

The fuller scorecard federal centers use has four dimensions. Delivery tracks production deployments, cycle time from intake to deployment, and the share of deployments with documented outcome measurement. Talent tracks headcount against plan, retention, time to fill key roles and cross-training. Governance tracks the share of use cases with a completed impact assessment, with a fairness audit, and with red team findings remediated, plus governance board cadence and inventory completeness. Impact tracks mission outcomes attributable to deployments, public trust measures where they apply, and closure of oversight findings.

The metrics to refuse are as important as the ones to keep. Slide decks produced, conference appearances made and pilots launched that never reached production are vanity measures, and a reviewer who knows the field will read them as activity theater. Publishing performance data, with appropriate protection of sensitive information, is both a transparency obligation and a retention tool, because strong practitioners prefer visible work. Publishing it also makes the next budget conversation shorter, since the argument has already been made in public over four quarters.

What Marcus did next

For Marcus, the fix was retroactive but it worked. He wrote the charter he had skipped, got his secretary to sign it, killed the orphaned chatbot, and ran an intake round that surfaced an eligibility document classification problem one office desperately wanted solved. That office assigned a detailee. Eight weeks later the tool was clearing a documented 30 hour per week review backlog, the detailee had become the office's embedded AI champion, and the center finally had an internal customer who described the work as theirs rather than as IT's.

Use the outline below as the skeleton for the one-page document your sponsor signs. It is deliberately short. A charter nobody reads is the same as no charter, and the version that gets signed is the version a busy executive can finish in the time between two meetings.

• Mission statement. One sentence naming what this center exists to make possible.
• Operating model. Central, federated or hub and spoke, with the reason and the review date.
• Scope and boundaries. Problem classes in scope, and what is explicitly out, including national security systems.
• Sponsor and authority. Named senior owner, and the decisions the center can make without escalation.
• Funding. Year one and year two budget, the funding source, and the review date.
• Customers and intake. Who it serves, how requests are submitted, how they are scored, and how a no is communicated.
• Governance links. How projects flow through risk framework functions, privacy, security and civil rights review.
• Success metrics. Delivery, talent, governance and impact measures, and who receives them.
• Sunset and review. The date the agency decides to continue, change or close it.

Anti-Patterns

• The consulting center. Pure advisory with no delivery authority. It produces well-argued reports that components read and ignore, and its staff gradually become the people who write the strategy nobody executes. The fix is delivery capacity attached to real decision rights in the charter.
• The outsourced center. The entire function run by a contractor with no knowledge transfer to federal staff. It performs adequately until the contract ends, at which point the capability, the documentation and the institutional memory leave together. Contract staff can extend the center; they cannot be it.
• The pilot factory. Endless pilots, no production deployments, and metrics that count launches. This is the pattern Marcus was one quarter away from institutionalizing, and it is the single most common failure in government AI. Pre-commit cutover criteria before the pilot starts.
• The center of prestige. High-profile partnerships with well-known universities and model vendors, an impressive speaker circuit, and no domain depth. The tell is that nobody in a program office can name a problem the center has solved for them.
• The walled garden. No embedded component staff, no rotations, no transparency about what was accepted or rejected. Components respond by building their own AI quietly, which is exactly the duplication the center was funded to prevent.
• The technology-first center. Tools adopted before mission needs are understood, leaving components holding platforms they cannot use and licenses they did not ask for. Buying the platform is the easiest decision available and it feels like progress, which is why it happens first.
• The shadow IT center. Operating outside the authority of the CIO, the security officer and the privacy officer because those reviews are slow. It works until the first security or privacy finding, after which the center spends a year rebuilding trust it could have kept for the cost of three meetings.
• The bias-blind center. No civil rights partner, no fairness discipline and no route into legal review. The Dutch benefits scandal is the cautionary tale here, and the pattern it teaches is that harm to the public was compounded by an internal structure in which nobody's job was to look for it.
• The revolving door center. High leadership turnover and no institutional memory, so each new lead restarts the strategy and the intake criteria. Career ownership and written decisions are the only defenses.
• The measurement-free center. No scorecard, no outcome tracking and no defensible narrative when the budget examiner arrives. A center that cannot show its value does not get cut because it failed; it gets cut because nobody could tell.
• Charter last. Marcus's original error, and worth listing separately because it produces all of the above. Hiring before chartering means the team's mandate is inferred from whoever asks them for things first.

Practice Prompts

1. Write the one-page charter. Using the outline in this lesson, draft the charter for a center in your agency. Which of the nine items can you complete today from documents that already exist, and which one would require a conversation with someone who has not yet agreed to have it?
2. Choose the operating model out loud. Argue the case for central, federated and hub and spoke in your agency's specific circumstances. Which one does your agency's current funding mechanism actually support, and what would have to change to support a different one?
3. Score three real candidates. Take three AI ideas circulating in your organization and score each on mission value, feasibility, risk category, data readiness, replicability and capacity. Which one wins, and could you defend the ranking to the office whose idea came third?
4. Draft the no. Write the written rejection you would send for the lowest-scoring candidate. If you cannot write it in a way that keeps the requesting office willing to come back next quarter, your intake process is not ready to launch.
5. Audit for duplication. List the review bodies a new AI project must already pass through in your agency. Which of them could absorb AI-specific criteria, and which genuinely needs a new mechanism? Every new body you propose has to survive that question.
6. Build the scorecard. Choose two measures for each of delivery, talent, governance and impact that you could report this quarter without new tooling. Name the recipient. Then identify which vanity metric your organization currently reports that you would stop reporting.

Reflection

Think about the last centralized capability your agency created, whether for AI, data, security or customer experience. Did it end up owning the work or enabling it, and which was the charter's intention? Consider who in your agency could currently say no to a program office's AI request and make it stick. If nobody can, the mandate has not been granted yet. Ask yourself what your center would be able to show a budget examiner after four quarters, using only measures you already collect. And if the honest answer is that the strongest evidence would be a list of pilots, what would have to happen in the next two quarters for that list to become a list of production systems with named owners?

Glossary

• Center of Excellence. A small central team that accelerates and de-risks AI adoption for the rest of an organization by concentrating scarce skills, shared infrastructure and governance capability.
• Charter. The signed document specifying a center's mission, scope, authorities, services, funding, staffing and measurement. It converts a team into a mandate.
• Hub and spoke. An operating model in which a central body provides infrastructure, standards and rotations while component teams retain domain delivery and ownership.
• Governance by permission. The failure state in which a center has visibility into AI work across the agency but no delegated authority to require, block or change anything.
• Intake pipeline. The staged process by which candidate use cases are submitted, screened, planned, delivered and transitioned to operations, with published criteria and published decisions.
• Kill criteria. Conditions defined in advance under which a use case is formally closed rather than left to run without funding, attention or a decision.
• Working capital fund. A funding mechanism in which components pay assessments or fees for shared services, letting a center recover costs from the offices that consume its work.
• Rights-impacting and safety-impacting use. Risk categories in federal AI guidance that trigger minimum practices such as impact assessment, testing and human oversight before and during deployment.
• Balanced scorecard. A measurement approach covering delivery, talent, governance and impact together, so that strength in one dimension cannot disguise failure in another.
• Activity theater. The pattern of reporting pilots launched, decks produced and events attended as evidence of maturity, when none of them indicate a capability the agency can repeat.

Related Lessons

The workforce pillar of this lesson is expanded in AI Talent Development and Retention, which covers hiring authorities, career ladders and retention mechanics in depth, with Workforce Planning for AI sizing the roles beforehand. Establishing an AI Governance Board and Your Agency's AI Governance Structure describe the bodies your center has to align with rather than duplicate, and AI Use Case Inventory and Documentation (OMB M-24-10) covers the inventory obligation the intake pipeline feeds. Data Infrastructure for Enterprise AI and Shared Services and Infrastructure Models go deeper on the platform decisions, while Vendor Lock-In Prevention addresses the dependency risk directly. Moving from Pilot to Production is the antidote to the pilot factory, Measuring Organizational AI Maturity situates the scorecard in a wider assessment, and Cross-Agency AI Coordination covers what to do when the capability you need already exists somewhere else in government.

Closing

A Center of Excellence does not succeed by building the most AI. It succeeds by making the rest of the agency able to build AI well, and the structures that produce that outcome are unglamorous: a signed charter, delegated authority, a published intake process, a scorecard nobody can argue with, and a review date at which the agency decides honestly whether to continue. None of those require a platform purchase, and all of them are harder to obtain than one.

Marcus recovered because he did in month seven what he should have done in month zero. That order is expensive but survivable, and it is the more common story. If you are standing one up now, spend the first quarter on the charter, the intake criteria and the sponsor relationship, and hire against the pipeline those produce. If you already have one that feels orphaned, the diagnostic question is not how many staff you have. It is whether any program office in your agency would describe one of your systems as theirs.

Key Takeaways

• Charter before hires. A signed charter naming a sponsor with real ownership, delegated decision rights and a funding line is what turns a team into a mandate. Staffing first produces an orphan with a platform.
• Concentrate for five specific reasons. Scarce talent, shared infrastructure cost, reusable governance artifacts, the risk of uncoordinated pilots, and oversight expectations set at agency level. If none of those apply to your situation, a center is not the answer.
• Hub and spoke is where mature programs converge. Central models bottleneck, federated models fragment governance, and advisory-only models lack teeth. Expect to evolve the choice as the program matures, and write the current choice down.
• Hire the data engineer and the governance specialist. Government AI fails on data plumbing and compliance more often than on algorithms, yet these are the most underhired roles. Staff to your visible pipeline, not to a ratio.
• Saying no is the product. A scored intake pipeline with published criteria, written rejections and pre-committed kill criteria is what keeps a center from collapsing into a help desk or a pilot factory.
• Replicability beats one-off wins. A reusable pattern many offices can adopt is worth more than a solution to a single office's problem, and it is the only way a small center scales its effect.
• Subtract governance rather than adding it. Embed AI criteria into the review bodies that already exist, and route privacy, security and civil rights through the officials who already own them.
• Measure four dimensions, refuse the vanity ones. Delivery, talent, governance and impact together; pilots launched, decks produced and events attended are activity theater and a reviewer will read them that way.
• Contractors extend a center and cannot be one. Knowledge held inside a contract leaves with the contract, and federal decisions about federal AI need federal decision-makers.

Frequently Asked Questions

How small can a center be and still be credible? Smaller than most agencies assume, provided the mandate is real. A handful of roles covering leadership, build capability, data engineering, product management and governance can run a genuine intake pipeline and deliver a first production system. What cannot be scaled down is the charter. A twenty-person team without delegated authority accomplishes less than a five-person team whose sponsor has signed a document saying what it is allowed to decide.

Should the chief AI officer lead the center? Both patterns exist and both work. The officer may lead the center directly, or may be a peer who exercises oversight of it. What matters is that the arrangement is written down, because the ambiguous version produces a center that believes it has authority it does not have and an officer who believes the center is handling obligations it is not. Decide which pattern you are using before the first governance board meeting, not during the first disagreement.

Our components already have their own AI teams. Have we missed the window? No, and you are arguably in the better starting position. The hub and spoke model assumes exactly this: domain teams that hold mission knowledge, plus a center that supplies platform, standards and governance capacity none of them can afford separately. The work is to define what the center owns exclusively, what the components own, and what is shared, then to make the center genuinely useful to teams that did not ask for it.

What if we cannot get a dedicated appropriation? Most centers do not start with one. The common alternatives are a working capital fund arrangement in which consuming components pay for services, project-specific agreements for cross-agency work, and modernization funding for the initial stand-up. Each carries a different incentive: fee-for-service keeps a center honest about usefulness but discourages the shared governance work nobody wants to pay for, so name that risk in the charter and fund the shared work explicitly.

How do we stop the center becoming a bottleneck? Refuse the role of sole builder. The moment every model in the agency has to pass through the center's hands, the queue becomes the strategy and everything else waits. Embedded staff who deploy into components, reusable patterns rather than bespoke builds, published criteria that let offices self-select before submitting, and honest capacity limits in the intake decision are the mechanisms that keep the queue from becoming the product.

What is the earliest sign that a center is failing? Not the delivery count. It is that no program office describes any of the work as theirs. An orphaned center can look healthy for two or three quarters on the strength of hires made, tools purchased and pilots launched, and the failure only becomes visible at adoption. Ask each quarter which offices have assigned staff, committed money or taken ownership of a deployed system, and treat a quarter with no answer as the signal it is.