Enterprise AI Team Architecture

The AI Team architecture separates workforce design, execution, and customer governance so each can evolve independently.

Reference flow

Manage → Release → Execute → Work → Connect

The enterprise architecture separates five concerns:

1. Manage: Compound Leverage maintains source-controlled, versioned AI Team definitions.
2. Release: Customer-specific configurations move through staging, validation, approval, and production release.
3. Execute: Approved AI environments run the workforce through platform-specific packages or adapters.
4. Work: AI Teams organize Digital Employees, job architecture, skills, workflows, templates, prompts, scripts, scheduled routines, and knowledge references.
5. Connect: Approved tools, enterprise systems, Compound Leverage capabilities, and customer-controlled context provide the information and actions required to perform work.

At the workforce level, the logical flow is:

Human direction → AI Team → Digital Employee → Skill → Orchestration → Execution → Customer AI Control Plane → Authorized systems and tools

The flow is logical rather than a requirement that every request pass through a separate software service. A capable runtime can perform several steps in one execution while preserving the boundaries needed for management and governance.

Logical layers

Human management

Managers and subject matter experts establish goals, provide organizational judgment, resolve ambiguity, review exceptions, and approve gated actions. See Human Management of AI Teams.

AI Team

The AI Team defines the business function and groups the Digital Employees required to perform it.

Digital Employees and skills

Digital Employees own jobs. Skills contain specialized methods, decision criteria, quality checks, and expected outputs. Keeping these separate prevents the persona or job definition from becoming an oversized prompt.

Orchestration

Orchestration coordinates workflow state, sequence, handoffs, validation, blocked work, and human approval gates.

Execution layer

The execution layer makes jobs and skills executable through plugins, agents, scheduled tasks, coding environments, or model-native mechanisms.

Customer AI Control Plane

The Customer AI Control Plane governs organizational context, identities, permissions, policies, data access, tool access, routing instructions, and audit boundaries for both humans and Digital Employees.

Enterprise systems and tools

Approved data stores, applications, APIs, browsers, search systems, connectors, MCP servers, and other tools provide the capabilities needed to perform work.

Source and release management

Compound Leverage manages the reusable AI Team source while customer-specific releases can be configured and validated before production. This separates centrally maintained workforce capabilities from customer deployment decisions.

Execution environments

The same AI Team product can be packaged for approved execution environments such as ChatGPT or Codex, Claude, Microsoft Copilot, and local or private model environments. The execution environment is runtime, not the definition of the workforce.

Capability and job architecture

Digital Employees are organized into named AI Teams rather than treated as standalone agents. Their work can include roles and responsibilities, skills, context and instructions, SOPs and workflows, templates and prompts, scripts, scheduled routines, assignments, and knowledge references.

Customer and Compound Leverage boundaries

Compound Leverage can provide shared or proprietary capabilities while the Customer AI Control Plane governs customer context, identities, permissions, policies, enterprise data, approved tools, and systems.

Customer proprietary information can remain under customer control while the workforce retrieves only the context and capabilities authorized for a job.

Enterprise connections may include CRM systems, SharePoint, internal databases, APIs, native connectors, MCP servers, and other approved integrations.

Operating principle: Compound Leverage manages the workforce. Customers control their data and systems.

Model independence

A Digital Employee does not have to equal a separate model instance. One capable model may execute several logically distinct jobs while those jobs remain separate for ownership, evaluation, governance, auditability, and human management.

This is what allows the workforce definition to survive changes in model architecture.

See the architecture in practice

The AI Capture Team and AI Proposal Team show how these layers can be applied to complete business functions.