Agentic AI is moving fast. As organizations struggle to scale from endless proofs-of-concept to sustainable scale, one question remains under-explored: what is the ‘right’ mental model, for agents in the workplace? We hear a variety of mental models wherein agents are often referred to as ‘personas’, ‘projects’ or ‘processes’; attempts to describe a shared operating logic for what an agent is and how it should change work. It feels a bit like the parable of the blind men and the elephant, wherein each person touches one part of the beast and mistakes it for the whole.
In practice, teams fall back on familiar metaphors, such as treating agents as digital employees (“Look! We gave the bots employee IDs!”), or as a slightly smarter form of automation. Those metaphors may feel intuitive, but they encourage the wrong design decisions, the wrong governance posture, and the wrong expectations between business and technology teams.
The ‘agent as employee’ mental model is the most appealing because it mirrors how enterprises already allocate work. It also helps with a narrow subset of requirements, such as identity, permissions, and access management. However, it breaks down quickly because it frames agents as tactical work units rather than reusable cognitive assets.It also encourages anthropomorphism, which can inflate perceived autonomy and accountability, and push organizations toward managing individual agents instead of managing a portfolio as a governed system. Most importantly, it reinforces the organizational status quo. It implies that enterprises can swap humans for agents while leaving structures, decision rights, and controls intact. That temptation is costly because agentic value usually comes from redesign, not substitution.
At Forrester, our approach to agentic mental models treats agents as having a dual identity. First, an agent is a skill: a discrete, bounded unit of cognitive work that is modular, reusable, and extensible across adjacent use cases. This skill lens gives leaders a clean way to answer the essential question, “what does this agent do?”It forces a concrete definition of what the skill does, what triggers it, what data and inputs it takes in, and what outcomes it produces. Skills are also designed for modularity and callability, so they start to look less like human roles that are ‘hard-coded’ into specific departments, and more like callable cognitive services that take their place within an enterprise’s portfolio of assets.
Such modularity lets teams build a coherent library of capabilities, avoid duplicative one off agents, and govern change with discipline as each skill evolves, is reused across functions, and is swapped out when better implementations emerge.
Second, an agent is a product that’s intentionally managed over time. The product identity is what gives the underlying skill a roadmap, a lifecycle, and a set of dependencies to be owned and extended, release after release. This product framing shifts the agentic skill from a one-off implementation to a durable enterprise asset whose capability surface expands across successive releases to handle broader, more complex use cases.
Those dependencies include the controls and services that make the skill viable in production, such as policy enforcement, telemetry, evaluation and quality gates, and governance mechanisms.
That said, defining the right mental models for AI agents only solves half the problem. Today, human competencies sit in HR taxonomies while agent skills live in technical inventories. As cognitive work becomes more of a collaboration between humans and AI, we fall short of a common language to describe such work in a shared language that spans people and machines.
We need a new approach that closes that gap by making skills the common unit of design and governance. In this approach, both human skills and nonhuman skills are inventoried as discrete capabilities with clear intent and boundaries, then assembled into higher order constructs. A skill is a discrete capability you can invoke on demand, whether it lives in a person or in an agent.
Examples include “negotiate terms,” “classify customer issues,” “draft a compliant response,” or “spot anomalies”.
Once you can name skills cleanly, you can build an inventory that spans both the workforce and the agent portfolio, and you can manage that inventory like any other strategic asset. From that skills inventory, the organization can further define composite concepts such as ‘roles’ and ‘workflows’.
Roles are a nonlinear combination of skills aligned to a distinct set of work outcomes, especially tasks and decisions. Workflows are sequences of steps (cognitive and deterministic, involving humans or non-humans) that move work from trigger to outcome.
For example, an ‘exceptions approver’ role may combine skills such as ‘interpret policy intent’, ‘assess risk exposure’, ‘weigh business impact’, ‘approve within limits’ and so on. The path an exception follows is an example of a workflow: request submitted, context gathered, risk scored, recommendation produced, decision made, approvals recorded, controls updated, and audit trail stored.
In a hybrid setup, a system composed of skills dynamically organized into roles and workflows can assemble context, draft the rationale, and propose a decision, while a human, who may be part of the workflow, could own the decision and accountability.
At Forrester, we call this conception of skills, roles, and workflows the Cognitive Operating Model. It gives leaders and architects a practical way to describe and redesign cognitive work.
This discipline is increasingly timely as firms such as Moderna have started to bring technology and human resources under a single leader, in part to prepare for a workplace where execution is shared across humans and AI. As agentic systems move from novelty to infrastructure, this kind of decomposition becomes essential to redesigning jobs, restructuring teams, and rebuilding core processes around a callable mix of human and nonhuman skills.
The author is Leslie Joseph, Principal Analyst, Forrester.
Disclaimer: The views expressed are solely of the author and ETCIO does not necessarily subscribe to it. ETCIO shall not be responsible for any damage caused to any person/organization directly or indirectly.
