AI agents are beginning to be incorporated into real operations, on an increasingly larger scale. This changes the business challenge and the next obstacle focuses on organizational readiness.
Every company must carefully plan how work is assigned, how decisions are governed, how systems are directed, and where human responsibility lies, as agents do more in the real world.
Vice President of Data and Artificial Intelligence at Kyndryl UK&I.
The gap between ambition and preparation is increasingly difficult to ignore. The Kyndryl Readiness Report found that 87% of business leaders expect AI to completely reshape career paths and role responsibilities. But right now, only 29% said staff can use AI effectively, while 62% said they were still in the “experimentation phase” with AI.
The success of AI agents will depend less on capability in isolation and more on whether the enterprise is adequately positioned to govern, orchestrate, and operationalize them.
That means alignment between business intent, decision rights, data access, sovereignty, workflow design, governance, and human oversight. Without that foundation, agents will struggle to deliver value at scale and may introduce new risks.
Operational complexity in practice.
That is why it is also worth being precise about what is meant by agent AI and why it creates operational complexity in practice. Agent systems are not simply generative AI with a more sophisticated interface. They are systems that can plan, take action, coordinate tasks, and coordinate multi-step workflows with limited human involvement.
This makes them more capable, but also materially more difficult to govern. Once technology can reason, invoke tools, coordinate systems, and act, the question is no longer simply “what it can do,” but whether we are equipped to execute it.
Smaller agent deployments typically rely on limited data sets and touch relatively few systems. At scale, that simplicity disappears. More systems, more integrations, more operational variation and greater governance demands add pressure. Many pilot projects lose momentum at this point because the scaling of AI forces organizations to address long-standing organizational and architectural debt.
If we want these systems to work in complex business environments, we also need to stop collapsing the discussion into an LLM discussion. LLM is important, but it is only one component of a much broader agent system architecture. Real enterprise deployments depend on orchestration, tooling, context, memory, workflow logic, policies, permissions, runtime controls, identity management, observability, and human escalation paths all working together.
If we focus the conversation too narrowly, we risk failing to design critical components that determine whether the system can operate safely and effectively in production.
For me, one of the most important practical points is that we must design with the most complex operational workflows in mind, especially in mission-critical environments, but sequence the deployment intelligently so that we can de-risk deployments as we scale.
Scale appropriately
As agents scale, they put pressure on companies in four ways.
The first is data overload, as sensitive and unstructured information becomes more available, reusable and exposed.
Second is integration tension, because each additional agent increases dependency on existing platforms, interfaces, and operational processes.
Third is operational stress, as an increasing number of autonomous components interact, complexity compounds and failure scenarios multiply.
Fourth is governance tension, where oversight models designed for static systems struggle to keep pace with dynamic, adaptive behavior.
Managing this well depends on some practical disciplines:
Set clear decision boundaries: establish which agents can decide, what needs to be escalated, and what remains firmly under human control.
Scale Design Orchestration: Multi-agent environments need coordinated workflows, shared context, and clear checkpoints to avoid deviations, duplication, and compliance failures.
Incorporate the intervention into the operating model from the beginning; Supervisory control should not be treated as an emergency measure. It should be incorporated through thresholds, alerts, approvals, and disconnection and rollback mechanisms.
Assign responsibility to designated roles and systems of record: If a decision cannot be traced, challenged and defended, it is not ready for production.
This is where control needs to get much closer to runtime. It is not enough for policies to remain in a political document, in a governance forum or somewhere in the background. The policy must be machine readable, testable and enforceable. Permissions should be tailored to the context and escalation should be set from the beginning.
Basic operating requirements
In addition to this, telemetry, orchestration, real-time monitoring and AIOps are now core operational requirements. As agent AI becomes part of the daily workflow, telemetry must go beyond uptime and response times. Organizations now require visibility into behavior, alignment to intent, workflow dependencies, exception trends, and policy enforcement by code.
Testing must also evolve. If a system is dynamic, context-sensitive, and capable of taking different paths, we can't test it as if it were a deterministic workflow with a bit of layered AI. We're not just testing responses, we're testing behavior.
This creates a difficult but important leadership balance. Too much autonomy without enough control creates unmanaged risks. Too much control without enough autonomy slows the realization of value. The goal is not unrestricted freedom or rigid lockdown, but limited autonomy: agents operating at high speed with clearly enforced policies and controls, with escalation paths and explainability built in.
Strategic partnerships are also becoming part of the underlying architecture. No organization alone can manage orchestration, integration, governance, platform interoperability, and operating model redesign at the speed now required. The most effective partnerships must be co-designed around shared responsibility for results, resilience, and speed to value.
Just as important, organizations need to take the time to align. That means bringing people together from the beginning through design cycles, governance forums, and cross-functional roadmaps. Internal and external stakeholders must be involved from the beginning: architects to validate the stack, engineers to scale it, and risk leaders to ensure compliance is addressed.
When those inputs are coordinated early, agency initiatives are less likely to stall, regardless of how advanced the technology appears.
Pulling as one to achieve the vision
The potential of AI agents will not be unlocked through experimentation alone. It will be done by organizations prepared to redesign the systems around them, including the governance, architecture, operating models, and leadership disciplines needed to turn autonomy into business value.
Pacing organizations understand that value at scale depends on being structurally prepared to run AI-native operations.
We've ranked the best data visualization tools.
This article was prepared as part of TechRadar Career Insightsour channel to feature the best and brightest minds in today's tech industry.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing, find out more here:
