The Operational AI Stack: What It Actually Takes to Run Agentic AI
- Terry Chana
- 5 days ago
- 6 min read
The conversation about agentic AI has moved faster than almost any technology topic of recent years.
What was - not that long ago - a research concept is now appearing on vendor roadmaps, in transformation strategies, and in board-level discussions. Just about every organisation I speak to is exploring how autonomous agents could improve service delivery, reduce operational overhead, and help teams manage growing demand.
The ambition is understandable. The challenge, however, is that many organisations ask what agents can do rather than what they need in order to operate successfully.
For many organisations, "agentic AI" has come to mean an autonomous capability applied as broadly and as quickly as possible. People are looking for a system that can make decisions, coordinate activity across multiple platforms, and complete work with minimal human intervention.
While that picture is accurate, what it misses is everything that needs to be in place for it to be viable.
This is the final article in the Operational AI Stack series. The first explored the Assist layer, how it supports teams with knowledge and insight, and where it stops. The second covered the Automate layer, how structured workflows create repeatable execution and why it's the prerequisite many organisations miss out. This article focuses on the Agentic layer, what it enables and where it creates value, but firstly, what it needs in place to deliver the outcomes you’re looking for.
What Agentic AI Is
The definition in the Operational AI Stack is clear: Autonomous AI agents execute and optimise outcomes across systems.
Each word of that definition matters.
Autonomous means the agent acts without needing a person to initiate or supervise each step
Execute means actions are taken, not just suggested; the agent moves work forward.
Across systems means this activity can span multiple platforms simultaneously: ITSM, ESM, CRM, ERP, analytics, and communications.
Optimise is where it gets interesting. The agent doesn't just execute a fixed sequence. It can evaluate context, adapt to changing conditions, and pursue the best path towards a desired outcome within defined boundaries.
This is what differentiates agentic AI from both assistants and workflow automation.
Assistants help people perform work.
Workflows help organisations structure work.
Agents help organisations achieve outcomes.
What the Previous Two Layers Were Building Toward
While so many organisations are rushing to pursue agentic AI, this Agentic layer only really makes sense as part of the whole Operational AI Stack.
The Assist layer made knowledge accessible and improved how people interacted with systems at the individual layer. But it didn't change how work moved between people or systems.
The Automate layer focused on improving how work moves through organisations by structuring processes. It gave operations a skeleton. Requests were routed. Approvals were tracked. Handoffs happened in systems, not in inboxes. And outcomes were logged and visible.
This structure is the foundation that the Agentic layer depends on.
An agent needs defined processes to act within.
It needs clear policies and consistent logic to apply.
It needs connected systems and structured handoffs to execute.
It needs observable, measurable outcomes so that success can be evaluated.
I’ve seen many organisations struggle to move beyond pilot projects because they don’t have these foundations in place. While the technology may be ready, the operational environment isn’t.
So, instead of achieving autonomy, these organisations end up with a more expensive version of the same coordination problem they started with, albeit they’re now running at speed and at scale.
What the Agentic Layer Does
The key shift from automation to agentic AI is moving from task execution to outcome execution.
Traditional automation asks:
"What action should happen when this event occurs?"
An autonomous agent asks:
"What outcome are we trying to achieve, and what actions are required to achieve it?"
Consider a service operation managing a high volume of requests across multiple systems.
At the Assist layer, employees handle these requests, supported by better information, recommendations, and access to knowledge. The work still relies on people.
At the Automate layer, the requests are classified, routed, escalated, and tracked through structured workflows. The process becomes more efficient and consistent, with people only involved to manage exceptions.
At the Agentic layer, the agent is working toward the outcome itself, determining what needs to happen and when. It gathers information, determines next steps, initiates actions across systems, monitors progress, adapts when conditions change, and escalates only when a decision requires human judgment.
The role of the people involved has changed entirely. They’ve not been removed, just repositioned. All the decisions that require context, consequence, relationship, or policy judgement remain with people. But at the Agentic layer, everything else just runs.
Why Most Organisations Aren’t Ready for Agentic AI
Most of the conversations around agentic AI focus on capability.
Agents can reason, make decisions, and take action.
But this leads some to assume that once the technology is available, organisations can immediately use it and gain value from it.
In practice, though, the challenge is now rarely the technology.
Most of the organisations I talk to discover that autonomous agents depend on capabilities and infrastructure they haven’t yet fully developed: connected systems, structured workflows, reliable data, clear governance, and measurable outcomes.
This is why many organisations can successfully demonstrate agents in pilot environments but struggle to scale them into production operations.
The question is not whether agents can work. The question is what limits their ability to deliver value in the real world.
The Limits of Agentic AI
It would be wrong to end this series by suggesting the Agentic layer is the endpoint.
As demand for digital services rises and resource constraints remain constant, we know Agentic AI can process at volumes that staffed services can't sustain and act on patterns across datasets faster than any manual review process. But the Agentic layer is constrained by the same thing that constrains every layer of this stack: the quality of what sits beneath it.
These autonomous agents depend on connected systems, reliable data, structured processes, effective governance, and clearly defined outcomes. These are the same operational foundations that our human teams have always depended on.
An agent can only be as effective as the environment in which it operates.
This is one of the most important realities of agentic AI.
An agent that operates on poor data produces confident, wrong answers. An agent running on an undefined process amplifies inconsistency. And an agent without clear outcome definitions optimises for the wrong things.
The technology is no longer the limiting factor. It hasn't been for some time. The constraint is operational. How work is structured. How data is governed. How outcomes are defined. How human oversight is designed.
These aren’t technology problems but operating model decisions. And they're the decisions that determine whether agentic AI delivers on its promises.
Where Human Oversight Sits
One of the most important decisions is where governance and oversight should sit.
We’ve seen many failures in these models reported and shared. Whether that’s agents processing transactions outside authorised limits, completing incorrect account actions unable to reverse them, or ignoring explicit instructions to follow an alternative path. None of these are edge cases; they are the predictable result of deploying agents without clearly defined boundaries.
But while oversight is key, governance at the Agentic layer is not a monitoring problem but a design issue.
The question isn't how to watch what agents are doing at every step. The question is what decisions and actions agents are permitted to take, and under what conditions a human steps in.
In most organisations, that boundary is narrower than initially assumed.
Policy-bound, data-driven, repeatable actions are candidates for agent execution.
Decisions involving ambiguity, consequences, regulatory interpretation, or relationship management should typically remain with people.
The objective is not unlimited autonomy; the objective is trusted autonomy.
Observability is still critical. If you can't see what an agent did, when it acted, and what outcome it produced, you can't govern it. An observable agent is one you can audit, calibrate, and trust. An unobservable one is a liability regardless of how well it performs.
Where This Leaves Most Organisations
I see organisations working through these layers, all at different stages.
While most are at Assist, a meaningful number are starting to explore the Automate layer seriously, investing in workflow development and automation. And now, some are beginning to run the Agentic layer beyond pilots, working on production-scale deployments.
The challenge, however, is that this gap between pilot and production is where the real work happens. And it's where most organisations discover that the barriers to success aren’t technical.
What is often missing is the operational readiness required to support them. Structured workflows. Connected Systems. Defined outcomes. Clear governance. And clarity about which decisions stay with people and which don't.
These foundations ultimately determine whether agentic AI is valuable or just another technology initiative that never progresses beyond experimentation.
Defining how digital services and AI-enabled workflows operate, perform, and scale.
