Cobalt Giraffe Research
Enterprise Agentic AI Capability Map
The capabilities required to design, operate and govern enterprise agentic AI at scale.
Working capability model / research in progress
Most organisations begin their AI journey with models, copilots and individual use cases.
But enterprise agentic AI requires a wider capability system.
Once AI systems can interpret objectives, retrieve knowledge, coordinate work, invoke tools, delegate tasks and take consequential actions, architecture has to address more than the model itself.
The capability map below sets out the core enterprise capabilities required to design, operate and govern agentic AI at scale.
Working / evolving capability model.Enlarge SVG (new tab) ↗PNG export (new tab) ↗
Six capability domains
The model groups the required capabilities into six domains:
Experience & Interaction
How people and business processes initiate, guide and approve AI-driven work.
Agent Execution & Orchestration
How agents plan, coordinate, delegate and execute tasks.
Knowledge, Data & Context
How agents retrieve and use enterprise knowledge, data and context.
Integration & Action
How agents connect to enterprise systems, tools and external services.
Governance, Risk & Control
How the enterprise manages identity, authority, autonomy, policy, risk and compliance.
Observability, Assurance & Lifecycle
How agent behaviour is monitored, evaluated, assured, versioned and governed over time.
Identity, Authority and Autonomy
Three control dimensions cut across the capability landscape.
- Identity
- Who or what is acting?
- Authority
- What is the agent permitted to do?
- Autonomy
- How independently may the agent decide and act?
These dimensions cannot be treated independently from risk and evidence.
The stronger the potential consequence of an agent's actions, the more important it becomes to constrain autonomy, enforce policy and retain sufficient evidence to reconstruct behaviour.
Capability before isolated use cases
A portfolio of AI use cases does not, by itself, create an enterprise AI capability.
Sustainable agentic AI depends on reusable capabilities that can be applied across products, teams and business domains.
That means building shared foundations for:
- Orchestration
- Knowledge access
- Integration
- Identity
- Policy
- Observability
- Assurance
The architectural goal is not to build one successful agent. It is to establish the capability system that makes many agents governable.
Governance must become executable
Governance cannot remain only in policy documents and approval boards.
As agentic systems become more autonomous, governance has to become part of the runtime architecture.
That includes:
- Agent registration
- Technical identity
- Entitlements
- Autonomy limits
- Policy decision
- Policy enforcement
- Audit
- Evidence
This is where organisational policy becomes operational control.
The emerging principle
The current working hypothesis is:
Enterprise agentic AI requires a coordinated capability system spanning interaction, orchestration, knowledge, integration, governance and assurance, with control strength increasing according to consequence and autonomy.
This capability map is an evolving Cobalt Giraffe research model rather than a claimed universal reference architecture.