AI agents in live operations require new standards and management

AI agents are beginning to move into live operations, increasingly at scale. This changes the enterprise challenge with the next hurdle centering on organizational readiness.

Every business should carefully plan how work is assigned, how decisions are governed, how systems are directed, and where human accountability sits, as agents do more in the real world.

VP Data and AI at Kyndryl UK&I.

The gap between ambition and preparedness is becoming harder to ignore. Kyndryl’s Readiness Report found 87% of business leaders expect that AI will absolutely reshape career paths and role responsibilities. But right now, only 29% said that staff can use AI effectively, while 62% said they were still in the ‘experimentation phase’ with AI.

Success with AI agents will depend less on capability in isolation and more on whether the enterprise is properly 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 fresh risks.

Operational complexity in practice

This is also why it is worth being precise about what is meant by agentic AI and why it creates operational complexity in practice. Agentic systems are not simply generative AI with a more sophisticated front end. They are systems that can plan, take action, co-ordinate tasks and coordinate across multi-step workflows with limited human input.

This makes them more capable, but also materially harder to govern. Once technology can reason, invoke tools, coordinate across systems and act, the question is no longer simply ‘what can it do’ but whether we are equipped to run it.

Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!

Smaller agentic deployments usually rely on bounded data sets and touch relatively few systems. At scale, that simplicity disappears. More systems, more integrations, more operational variation and greater governance demands all add pressure. Many pilots lose momentum at this point because scaling AI forces organizations to confront longstanding architectural and organizational debt.

If we want these systems to work in complex enterprise environments, we also need to stop collapsing the discussion into an LLM discussion. The LLM matters, but it is only one component in a much broader agentic system architecture. Real enterprise deployments depend on orchestration, tools, context, memory, workflow logic, policy, permissions, runtime controls, identity management, observability and human escalation paths all working together.

If we center the conversation too narrowly we risk under designing critical components that determine whether the system can work safely and effectively in production.

For me, one of the most important practical points is that we should design with the most complex operational workflows in mind, especially in mission critical environments, but sequence implementation intelligently so we can de risk deployments as we scale.

Scaling appropriately

As agents scale, they place pressure on enterprises in four ways.

The first is data strain, as sensitive and unstructured information becomes more widely available, reusable and exposed.

Second is integration strain, because each additional agent increases dependence on existing platforms, interfaces and operational processes.

Third is operational strain, as growing numbers of autonomous components interact, complexity compounds and failure scenarios multiply.

Fourth is governance strain, where oversight models designed for static systems struggle to keep pace with dynamic and adaptive behavior.

Managing this well depends on a few practical disciplines:

Set clear decision boundaries: establish what agents are allowed to decide, what must be escalated and what remains firmly under human control.

Design orchestration for scale: multi agent environments need coordinated workflows, shared context and clear control points to prevent drift, duplication and compliance failures.

Build intervention into the operating model from the outset; supervisory control should not be treated as an emergency measure. It should be built in through thresholds, alerts, approvals and kill-switch and roll back mechanisms.

Assign accountability to named roles and systems of record: if a decision cannot be traced, challenged and defended it is not ready for production.

This is where control must move much closer to the runtime. It is not enough for policy to sit in a policy document, a governance forum or somewhere in the background. Policy needs to be machine readable, testable and enforceable. Permissions need to flex with context, and escalation needs to be built from the start.

Core operational requirements

Alongside this, telemetry, orchestration, real-time monitoring and AIOps are now core operational requirements. As agentic AI becomes part of day-to-day workflow, telemetry must go further than uptime and response times. Organizations now require visibility into behavior, alignment with intent, workflow dependencies, exception trends and adherence to policy by code.

Testing also needs to evolve. If a system is dynamic, sensitive to context and capable of taking different paths we cannot test it as if it were a deterministic workflow with a bit of AI layered in. We are not just testing answers, we are testing behavior.

This creates a difficult but important leadership balance. Too much autonomy without sufficient control creates unmanaged risk. Too much control without enough autonomy slows down value realization. The goal is neither unrestricted freedom nor rigid lock-down, but bounded autonomy - agents operating at speed with clearly enforced policies and controls, with escalation routes and explainability built in.

Strategic partnerships are also becoming part of the underlying architecture. No organisation can manage orchestration, integration, governance, platform interoperability and operating model redesign on its own at the speed now required. The most effective partnerships need to be co-engineered around shared accountability for outcomes, resilience and speed to value.

Just as importantly, organizations need to make time for alignment. That means bringing people together early through design cycles, governance forums and cross-functional road mapping. Internal and external stakeholders should be involved from the start: architects to validate the stack, engineers to scale it, and risk leaders to make sure compliance is addressed.

When that input is coordinated early, agentic initiatives are less likely to stall, regardless of how advanced the technology appears.

Pulling as one to reach the vision

The potential of AI agents will not be unlocked through experimentation alone. It will be realized by organizations prepared to redesign the systems around them - including governance, architecture, operating models and the leadership disciplines needed to turn autonomy into enterprise value.

The organizations setting the pace understand that value at scale depends on becoming structurally ready to run AI native operations.

We've rated the best data visualization tools

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

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: https://www.techradar.com/pro/perspectives-how-to-submit

VP Data and AI at Kyndryl UK&I.