Control & Governance
There is at least one consensus about agentic technology advance: it’s too fast, and it’s pace is accelerating. A two-pronged challenge ensues: operational control and organizational governance.
At the operational level, agents’ behavior is disruptive by design, their purpose being to look for unexplored territories, address unexpected situations, and imagine unfathomable solutions. That makes traditional control policies irrelevant.
At the organizational level, agents’ capabilities cut across human cognitive ones, undermining the principles of task definition and upsetting the whole of business process design. It follows that traditional governance approaches must be upgraded to integrate organization and knowledge engineering.
Thought Control
Compared to deterministic systems set in bounded contexts and driven by unambiguous instructions, agentic ones are meant to consider unexplored contexts and unpredictable courses of action. Controlling such behaviors is by itself challenging, all the more so when even instructions and commitments are themselves open to interpretation. It follows that control policies addressing behaviors will lag in trying to disentangle agents’ driving threads according to their status: assigned, asserted, inferred, conjured up, or hallucinated. Lest they become irrelevant, guardrails must be set upfront, prior to execution. And given the inherent indeterminacy of scope and operations, such guardrails will have to address agents’ thoughts, namely what they know and what they intend to do with it.
Knowledge Governance
Compared to control, which addresses agents’ behaviors, governance considers their use: who they are, what they do, how they do it.
From an organizational perspective, agents are of two kinds, human or digital. As such, agents are characterized by two kinds of capabilities: cognitive (observation, reasoning, and judgment) and communication (natural or symbolic language).
Regarding activities, governance should focus on changes in environments, in driving intents and objectives, and in supporting systems. Regarding agents’ modus operandi, the focus should be on the cognitive capabilities involved: observation (all agents), reasoning (all agents), judgment (human agents).
Epistemic Knowledge
It thus appears that both control and governance depend on the epistemic dimensions of knowledge: in support of awareness and agency for the former, and of assignments and accountability for the latter.
Regarding the control of agents’ behaviors, epistemic distinctions are used to define awareness (what agents know) and agency (what agents can do).
Regarding governance, epistemic distinctions are used to assign responsibilities for facts (observations and actions), concepts (intents and rules), and categories (systems) to human and digital agents.
Such selective assignments will greatly enhance transparency (resources), traceability (reasoning), and accountability (judgment) of agents’ behavior.
Maturity Levels & Roadmap
Maturity Levels
Given that understanding of the specificity of agentic AI control and governance issues, it’s possible to lay out a maturity roadmap.
Starting from scratch, the first level is characterized by semantic interoperability: all agents, human or digital, can communicate through a common semantic space. With ontological prisms, this is achieved through thesauri.
The second level ensures application interoperability: all agents, human or digital, can rely on a common address space to consistently access resources managed at enterprise level. With ontological prisms, this is achieved by combining thesauri and taxonomies.
The third level addresses organizational integration: all terms in controlled vocabularies reference concepts with assigned ownership and stewardship. With ontological prisms, this is achieved by extending the second level with organizational (concepts) domains.
The fourth level addresses systems integration: all schemas used by applications reference categories with assigned ownership and stewardship. With ontological prisms, this is achieved by extending the second level with operational (categories) domains.
The fifth level addresses enterprise architecture integration: all ownership and stewardship are aligned across organizational and operational domains. With ontological prisms, this is achieved through ontologies.
Maturity Roadmap
The roadmap to maturity is divided into three parts: the first includes the first and second levels; the second, the third and fourth levels; and the third, the fifth level.
The first and second levels are meant to be achieved sequentially, semantics before applications. The third and fourth levels can be achieved in parallel, according to domain-based granularity.