The AI Center of Excellence Playbook for ChatGPT Enterprise

Every enterprise AI program eventually hits the same scaling problem.

In the first year, the AI team does everything: runs the ChatGPT Enterprise deployment, builds the showcase Custom GPTs, trains the early adopters, handles the governance questions. It works, because the scope is manageable and the enthusiasm is high.

Then the program grows. GPT and Project creation velocity accelerates. Business units start making their own decisions. Governance requests pile up. The central AI team can’t keep pace. The program either stagnates under centralized control or fragments under decentralized chaos.

The organizations that escape this trap build an AI Center of Excellence (CoE): a governance and enablement structure that scales AI capability without scaling the central team linearly. Here’s how to do it for a ChatGPT Enterprise environment.

What an AI CoE Actually Is (and Isn’t)

An AI Center of Excellence is not:

• A team that builds AI for everyone else
• A committee that approves AI requests
• A training program that runs once a year
• A compliance function that reviews AI for risk

An AI CoE is:

• A capability network that distributes AI expertise across the organization
• A governance framework that enables fast decisions at the right level
• A measurement system that tracks AI capability maturity organization-wide
• An enablement engine that continuously improves builder quality

The distinction matters because the wrong model: the “central team builds it all” model: creates a bottleneck. The right model creates a multiplier.

The Three-Layer CoE Structure

Effective AI CoEs for ChatGPT Enterprise environments typically operate in three layers:

Layer 1: The Core AI Team (3-8 people)

This is the central team that owns the program. Their job is not to build GPTs for business units: it’s to build the infrastructure that lets business units build well.

Core AI Team responsibilities:

• Own the ChatGPT Enterprise deployment and admin configuration
• Define and maintain the governance framework (risk tiers, approval workflows)
• Build and maintain the AI portfolio intelligence system (discovery, classification, metrics)
• Run the AI Champion program (see Layer 2)
• Manage vendor relationships (OpenAI, tooling providers)
• Report to leadership on portfolio maturity and ROI

Key roles:

• AI Transformation Leader: Program strategy, executive communication, CoE coordination
• AI Engineer: Pipeline development, integrations, tooling
• AI Governance Analyst: Risk assessment, compliance review, policy maintenance
• AI Learning & Development Specialist: Training programs, builder enablement, content

Layer 2: AI Champions (10-50 people, depending on org size)

AI Champions are high-performing GPT builders embedded in business units who serve as the CoE’s distributed presence. They are not full-time AI roles: they’re existing employees who’ve demonstrated strong AI capability and are formally recognized and supported by the CoE.

AI Champion responsibilities:

• Serve as the first point of contact for AI questions in their business unit
• Review and provide initial approval for new GPT builds in their unit
• Surface pain points and use case opportunities to the Core AI Team
• Participate in monthly AI Champion community of practice calls
• Mentor new builders in their unit

How Champions are identified: This is where AI Transformation Intelligence becomes critical. Champions shouldn’t be self-nominated or manager-nominated: they should be data-nominated. The Builder Score (a composite of GPT quality, adoption, governance hygiene, and volume) surfaces the employees who are already building the best AI, regardless of title or department.

Formalizing these naturally-occurring champions into a CoE structure gives them recognition, community, and influence: and gives the CoE distributed reach without headcount.

What Champions get:

• Formal recognition as an organizational AI expert
• Priority access to new features and capabilities
• Dedicated training budget for AI certifications
• Direct line to the Core AI Team for questions and escalations
• Profile page in an internal AI directory

Layer 3: All Builders (everyone with GPT creation access)

The third layer is the entire population of employees who can create Custom GPTs and Projects. This isn’t a defined team: it’s the universe of potential AI builders.

The CoE’s job with this layer is to:

• Make good building easier than bad building (quality templates, starter prompts)
• Provide lightweight self-service training that’s immediately applicable
• Create clear paths for getting help and governance approval
• Make the portfolio visible so builders can find what already exists before building something new

Most enterprises make the mistake of treating this layer as a compliance problem (how do we control what they build?) rather than an enablement opportunity (how do we help them build better?).

The Governance Framework: Four Risk Tiers

One of the most valuable things a CoE can establish is a tiered governance model that matches review intensity to actual risk. Applying the same review to a personal productivity GPT and a customer-facing decision-support AI is both inefficient and ineffective.

Tier 0: Auto-Approved (Low Risk)

Profile: Personal-use GPT, no external tools or integrations, no regulated data, single creator as sole user, minimal system prompt.

Process: Auto-approved on creation with standard notice of terms.

Example: “Help me write better meeting agendas”: a personal productivity assistant for one person.

Tier 1: AI Champion Review (Standard)

Profile: Team-use GPT (multiple users in one team), no external integrations, non-regulated domain, reasonable system prompt.

Process: AI Champion in the relevant business unit reviews within 3 business days. Checklist-based review focused on appropriate use, data handling, and documentation quality.

Example: A sales team’s proposal writing assistant, connected to no external systems, used by 10 reps.

Tier 2: CoE Review (Elevated)

Profile: Department-wide or org-wide GPT, OR external API integrations, OR regulated domain adjacency (HR, legal, finance, compliance).

Process: Core AI Team reviews within 5 business days. Includes security review of any external connections, compliance review for regulated domains, and documentation requirements.

Example: An HR GPT that answers employee policy questions, referencing an internal policy Knowledge base.

Tier 3: Security + Legal Review (High Risk)

Profile: Customer-facing GPT, OR direct access to regulated data (PII, PHI, financial records), OR decision-making capability in regulated domains, OR external partner integrations.

Process: Full security, legal, and compliance review. SLA is 15 business days. May require external audit for highly regulated industries.

Example: A customer service GPT that can access account information and make recommendations on financial products.

This tiered model is only sustainable if it’s automated at the classification stage. Manual triage of every GPT is not scalable. The CoE needs a system that classifies risk tier automatically based on GPT configuration and semantic analysis, and routes accordingly.

The Metrics Cadence

An AI CoE without a measurement cadence is just a governance committee. Metrics make the CoE accountable to outcomes: and make AI transformation legible to leadership.

Weekly (AI Team internal):

• New GPTs created and their initial risk classification
• High-risk GPTs pending review
• Builder Score changes (new champions emerging, existing champions declining)

Monthly (CoE Leadership):

• Portfolio Maturity Distribution (Production/Functional/Experimental split)
• Top 10 GPTs by usage and quality score
• Builder recognition: Champion of the Month
• Training completion rates and impact correlation (did training improve GPT quality?)
• Process coverage map: which core workflows have AI coverage at what quality level?

Quarterly (Executive Leadership):

• Portfolio Maturity trend (are we improving?)
• Duplication rate trend (are we consolidating?)
• Estimated value delivered (based on top GPT case studies)
• AI capability distribution (are we developing talent broadly?)
• Governance health (risk tier distribution, review SLA compliance)

Annually (Board/Executive):

• AI Transformation Index: composite score of portfolio quality, builder capability, governance health, and business value
• Investment vs. returns: license cost, headcount, training investment vs. documented value creation
• Capability roadmap: where we’re investing next and why

The Common Failure Modes

Failure Mode 1: The CoE becomes a bottleneck. If every GPT requires CoE approval, velocity dies. The tiered model above exists specifically to prevent this. Default to enabling, not controlling.

Failure Mode 2: Champions are chosen for seniority, not capability. If your AI Champions are department heads who volunteered rather than proven builders who were identified by data, you have a political committee, not a capability network.

Failure Mode 3: The CoE owns the roadmap but not the intelligence. A CoE that makes decisions without portfolio intelligence: without knowing what exists, who built it, how it’s performing: is governing in the dark. AI Transformation Intelligence is the nervous system of the CoE.

Failure Mode 4: Training is event-based, not continuous. A two-day AI bootcamp produces a two-day bump in builder quality. Continuous, targeted learning: matched to each builder’s actual skill gaps as measured in their GPT output: produces compounding improvement.

Failure Mode 5: No executive mandate. An AI CoE without executive sponsorship gets ignored when it conflicts with business unit priorities. The CoE needs a named executive sponsor who can enforce governance standards when needed.

Getting Started: The 60-Day CoE Launch

If you’re starting from scratch, here’s a practical first 60 days:

Days 1-14: Baseline Assessment Run a full portfolio audit. Classify every existing Custom GPT and Project. Identify your top 10 builders by Builder Score. Document your current governance (or lack thereof). This is your starting point.

Days 15-30: Framework Design Define your risk tiers and approval workflows. Draft the AI Champion criteria and program structure. Get executive sponsor sign-off on the governance model.

Days 31-45: Champion Identification and Launch Invite your data-identified top builders to the AI Champion program. Run a kickoff session. Give them their first assignment: review and classify the existing GPTs in their business unit.

Days 46-60: Public Launch Communicate the CoE to the organization. Publish the governance framework, the GPT and Project directory, and the path to approval. Open the Champion nomination process for the next cohort. Start your weekly metrics cadence.

By day 60, you have a functioning CoE structure, a governed portfolio baseline, an embedded champion network, and a measurement system. Everything after that is iteration.

---

AgentsOrg is built to support exactly this operating model: providing the portfolio intelligence, Builder Scores, risk classification, and metrics dashboard that make an AI CoE functional rather than theoretical.

git clone https://github.com/ori129/agentsorg.git
cd agentsorg && make up

---

Building an AI CoE and running into a challenge this playbook doesn’t cover? Post in GitHub Discussions: we’re compiling community learnings into future versions of this guide.