Agents in the Enterprise: From Experiment to Operating Model

Dr. Wendy Ng, CISSP

AI Futurist | Thought Leader | Scientist

If you have worked with emerging technologies in an enterprise setting, you will recognise a familiar pattern. Capability arrives as a useful tool, then becomes infrastructure, and eventually forces a governance model that can operate at its tempo.

In 2023, I stated that artificial intelligence was approaching a point where it would begin to reshape multiple sectors, including space, and that its longer-term value would depend less on individual tools and more on how effectively it could be operationalised at scale. More than two years on, the direction of travel is clear. The more practical question is whether organisations can adopt AI without creating a parallel operating model that no-one fully owns.

AI is increasingly treated as foundational technology, sitting alongside cloud platforms, identity systems and operational tooling. Some organisations have progressed cautiously, shaped by legacy environments and regulatory obligations. Others have moved faster, driven by delivery pressure. Either way, AI is no longer peripheral.

The differentiator is shifting. Model capability matters, but sustained outcomes tend to come from integration discipline, clear control boundaries, and governance that is designed to work in production, not in slide decks.

Early enterprise adoption focused on bounded use cases, for example chat interfaces, anomaly detection and predictive analytics. Useful, but constrained.

Agentic systems change the shape of the problem. Agents can plan and execute sequences of actions across multiple systems with limited human involvement. In practical terms, this introduces three characteristics enterprises need to treat seriously:

·      Initiation: agents can trigger actions rather than simply respond

·      Context: agents can retain state across tasks and workflows

·      Coordination: agents can orchestrate activity across tools and platforms

 Once agents are placed into workflows that touch production environments, financial processing, compliance evidence or security controls, they become part of the control plane. The tempo changes. A human-driven process can tolerate ambiguity and delay. An agent operating continuously cannot.

The main barrier to effective agent adoption is rarely technical feasibility. It is organisational readiness, particularly governance models designed for slower cycles.

In traditional IT operations, centralised governance worked when change was infrequent. In modern environments, teams ship small, incremental changes with rapid feedback and higher tempo. Blog 62 makes the point in the DevSecOps context: governance needs to adapt to cadence, not fight it. The same principle applies to agents.

Local teams are closer to the workflow and the operational blast radius. Central functions still matter, but not as the sole gatekeeper. Their role is to set principles that reflect risk appetite, provide reference patterns, and maintain oversight without becoming a bottleneck.

If an agent can take action, it needs permission boundaries, traceability, and lifecycle management. Without this, scale becomes unmanaged scale.

This is not “human in the loop” as a slogan. It is where authority should remain human, what evidence is required for auditability, and what confidence thresholds are acceptable for automation.

Looking Towards 2026

In 2026, agents are likely to be less visible but more deeply embedded. They will increasingly be assumed capabilities within platforms, with attention shifting towards outcomes.

Organisations that progress fastest will not be those chasing maximal autonomy. They will be those that integrate agents into existing identity and control frameworks, define clear permission boundaries, and keep governance aligned with delivery tempo.

Autonomy will expand selectively where tasks are structured and evidence can be recorded. Fully autonomous operation in ambiguous domains will remain the exception, largely due to accountability requirements rather than tooling limitations.