The direct answer: no. A midsize company does not need an AI Center of Excellence before implementing its first AI marketing workflow. Start with a small, time-bound, cross-functional implementation cell tied to one measurable outcome. Centralize only what should be reused—policy, access, evaluation, approved data, observability, integration patterns, and incident response—while marketing and the business owner keep workflow decisions. Formalize a CoE when repeated coordination across multiple live workflows costs more than a shared function.
Starting with the org chart reverses the investment sequence. The company funds meetings, use-case catalogs, generic training, and a platform before proving an operating decision. At the other extreme, unconstrained local teams produce duplicate vendors, uncontrolled data, incomparable evaluations, and pilots nobody operates. Use a progressive operating model instead: project squad, implementation cell, federated hub, then a formal CoE.
Use four stages, not a binary choice
- Project squad — one workflow, owner, deadline, and bounded budget; suitable for proving value and operations.
- Implementation cell — fractional specialists support two or three workflows and create reusable standards.
- Federated hub — a small core maintains controls and common components while domain teams own outcomes and change.
- CoE or platform team — a broad portfolio, dedicated capacity, internal service, funding, and SLAs; justified when reuse, risk, and demand recur.
Decision rights matter more than the label. Microsoft describes an AI CoE as an internal team supporting adoption, integration, governance, and common practice. Those capabilities are useful at scale, but they can begin in a smaller cell. NIST calls for clear roles, communication, and executive accountability; it does not require a centralized department.
Write the Implementation Cell Contract
- Outcome — business metric, baseline, comparator, and decision date.
- Workflow — inputs, decisions, outputs, exceptions, volume, and affected people.
- Rights — who can prioritize, approve data, change a rule, release, pause, and stop.
- Shared foundation — identity and access, approved sources, evaluations, logs, cost, versions, fallback, and incidents.
- Domain owner — marketing owns the outcome, content, exceptions, and adoption; technology does not inherit the commercial target.
- Capacity — named hours from marketing, data, technology, security, legal, analytics, and finance.
- Evidence — quality, downstream outcome, unit cost, human intervention, risk, and operating load.
- Exit — dissolve, keep a simple rule, repeat the cell, federate, or fund a CoE.
The Contract prevents an AI team without an internal customer. For qualification, marketing or revenue operations owns policy and SLA; data resolves identity and quality; technology integrates; security and privacy approve controls; analytics designs comparison; finance validates contribution. The cell coordinates delivery but does not replace permanent owners.
An eight-week implementation scope
- Week 1 — select one problem, owner, baseline, and stop criterion.
- Week 2 — map the workflow, exceptions, systems, data, and human decisions.
- Week 3 — set decision rights, minimum controls, vendors, and cost ceiling.
- Week 4 — build the smallest flow in shadow mode with logs and an evaluation set.
- Week 5 — test quality, security, integration, latency, cost, and fallback.
- Week 6 — release to a bounded cohort with human review proportional to risk.
- Week 7 — measure outcomes, adoption, rework, incidents, and operating load.
- Week 8 — stop, revise, expand, or transfer; record what genuinely deserves reuse.
The UK government's People Factor guide treats scale as a partnership among leaders, managers, interdisciplinary delivery teams, employees, and users. The practical implication is that adoption and work redesign cannot be outsourced to a technical hub. Include the people who perform and receive the work—not only those who buy or configure technology.
The trigger for a formal CoE
- Three or more production workflows depend on the same evaluations, data, integrations, or controls.
- Two or more business units or countries repeat architecture, vendor, or risk decisions.
- A shared backlog now requires dedicated capacity and an SLA rather than project favors.
- Incidents, audits, model spend, and vendor changes require continuous operations.
- Common components demonstrably reduce delivery time or cost, rather than merely promising reuse.
- Business owners still fund and own outcomes; the hub does not become a parking lot for experiments.
Without those signals, an early CoE adds fixed cost and distance from the problem. When the signals exist and nothing is centralized, every workflow repeats integration, evaluation, and review. Compare the annual cost of the hub with duplicated labor, avoidable incidents, deployment time, and the value of workflows it unlocks. Licenses are not value, and pilot count is not an outcome.
Budget, cost, and funding
For the first workflow, fund a temporary cell from the business case plus a small shared foundation. Separate discovery, build, integration, evaluation, change, operations, and contingency. In a federated model, charge domains for variable use and fund mandatory standards centrally. A formal CoE needs a service catalog, capacity, SLAs, and a prioritization rule; without those, it becomes overhead competing for influence.
Major cost drivers are availability of the right people, data quality and access, number of systems and vendors, environments and identity, evaluation volume, regulatory needs, observability, human intervention, support, and change frequency. An enterprise tool may reduce duplicate buying, but it may also impose integration and capacity a first workflow does not need.
Measure the organization, not only the model
- Outcome — conversion, revenue, margin, time, or capacity attributable to the workflow.
- Delivery — time from decision to use, gate pass rate, and rework.
- Reuse — components actually adopted by another workflow and hours avoided.
- Operations — availability, latency, incidents, rollback, unit cost, and exception backlog.
- Adoption — correct use, abandonment, override, trust, and net time saved.
- Portfolio — workflows stopped, expanded, and maintained; capital trapped in pilots.
- Governance — documented decisions, expired exceptions, active owners, and approval time.
GAO organizes AI accountability around governance, data, performance, and monitoring. Use those four lenses to review the cell; do not turn them into four committees. A short decision record with evidence and an owner is more useful than a long policy with no workflow mechanism.
Connect the cell to the first outcome
Choose the first workflow with https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow, set autonomy at https://makinai.co/insights/en/copilot-workflow-ai-agent-marketing-what-build-first, prepare production at https://makinai.co/insights/en/ai-marketing-pilot-production-operating-scale, and find value leakage at https://makinai.co/insights/en/marketing-team-uses-ai-results-unchanged-value-leak. MAKINAI's service is at https://makinai.co/services/en/ai-strategy-transformation-consulting. Bring one workflow, its current owners, and two coordination costs; we can design the smallest cell and its scale trigger.
This design does not guarantee ROI or replace legal, employment, privacy, or sector requirements. A content workflow does not require the same structure as a sensitive customer decision. The sources support accountability, human-centered scale, governance, performance, and monitoring. The four stages, Contract, and triggers are MAKINAI editorial recommendations to validate in context.