What are the pressure points of designing the AI-ready internet network?

Julian Skeels, Chief Digital Officer at Expereo, recently spoke at TechEx Global and challenged some deeply held assumptions about networks and AI. We sat down with him afterwards to get his insights and key takeaways for CIOs looking to make AI a success this year.

What are the key messages from your talk at TechEx?

For years, enterprise networks were treated as background infrastructure. They were there to keep the packets moving and stay out of the way. But that era is over.

Because AI breaks that model.

AI systems don’t pause. Inference, monitoring, retrieval, and remediation run continuously. When the network degrades, outcomes degrade immediately. Today, the CIOs I speak with are struggling with networks that were never designed for this role. Not because they are slow, but because they are unpredictable.

These are the pressure points CIOs are now confronting. And they are forcing leaders to unlearn assumptions that no longer hold.

Why is an enterprise’s internet network no longer plumbing?

For most of our careers, applications failed whilst the networks were technically “up.” Availability was binary. Either the link was live or it wasn’t.

AI changes that.

AI workloads operate continuously and performance degradation does not trigger graceful failure. It produces incorrect outputs, delayed decisions, and exponential cascading impact across systems. When the network falters, the business falters with it.

Alongside AI, security and growth hotspots, digital sovereignty is moving from policy to architecture. Reliance on large external cloud ecosystems is becoming both a legal and strategic risk. Leaders will need to regain control by diversifying suppliers, adopting sovereign-compliant options where obligations demand it, and embed sovereignty into procurement and risk frameworks. The rise of digital nation states will accelerate this trend. Soon, sovereignty will separate companies that rent their data from those that truly own it. Digital provenance will soon be a requirement for global enterprises, meaning organizations must demonstrate control with evidence, not just claim it.

Why is speed the wrong measure of AI networks?

AI-ready networks shouldn’t be measured just by how fast they are. That’s because success is based on certainty.

Bandwidth still matters, but it is no longer the defining constraint. AI does not fail because data moves too slowly on average; it fails because data behaves unpredictably.

In my experience, the most important features of an AI-ready network are:

1. Determinism
2. Predictable routing
3. Consistent latency
4. Controlled behavior under load

AI systems depend on stable conditions to operate correctly, especially at scale.

Designing networks assuming peak throughput without controlling for variance creates fragile systems that look strong on paper but fail in production.

Why does variance, not outages, kill AI in production?

Averages hide pain.

But you do not experience an average network. You experience the worst five minutes.

AI workloads are highly sensitive to jitter, packet loss, and tail latency. Even brief brownouts can disrupt inference pipelines and real-time decisioning. Traditional SLAs mask this reality by smoothing performance into metrics that do not reflect your lived experience in production.

Design for averages, and AI fails silently.

Design for variance, and AI remains stable even when conditions change.

How can procurement-led network design create operational fragility?

Many enterprise networks are optimized for procurement simplicity, not operational truth.

Single network providers, uniform contracts, and clean SLAs may make networks easier to buy and govern, but performance does not behave uniformly across regions. The last mile dominates experience, because geography matters.

Simplicity at purchase can create fragility in production.

Why must “always on” be engineered, not promised?

AI workloads punish even brief degradation. AI systems require early detection of degradation and the ability to intervene before outcomes are affected.

And measuring availability after the fact is too late.

So availability must be engineered through path diversity, proactive monitoring, and real-time response, not assumed by contract.

Why do security and networking collapse into one control fabric?

AI removes quiet periods. Systems operate continuously, but so do attackers.

In AI systems, every connection is a real-time security decision.

Periodic controls and after-the-fact analysis are insufficient for real-time systems. Security must reside directly in the data path and be enforced continuously, with data moving under continuous inspection and behavioral monitoring.

This collapses traditional boundaries between network and security architecture. Performance, protection, and policy enforcement must operate as one system.

Sovereignty is a data flow problem, not a storage decision

Data sovereignty is not defined by where data sits. It is defined by how data moves.

Modern network architecture for AI workloads must make data flow observable and provable. That means enterprises need to be able to demonstrate where data travelled, under which policies, and across which paths.

If you cannot observe the path, you do not have sovereignty; you just have paperwork. This is especially critical in regulated regions such as EMEA, where expectations extend far beyond data residency and into operational proof.

What makes AI fail for enterprises?

You’ve put together strong models, capable teams, and well-funded initiatives. Prototypes work. Demos impress. Your AI shouldn’t fail on paper, but somehow it does. Why?

Nine times out of ten, AI fails when it meets the network.

Production exposes weaknesses in the network’s ability to move data reliably, securely, and predictably under real-world constraints.

That’s why the network often determines how far AI ambition can go.

What should technology leaders do to become AI-ready?

Most AI conversations start with models in a lab, and that’s fine, but readiness in production really begins with data movement.

I always ask one question: Can you move the right data to the right place, under the right rules, every single time?

A few practical steps:

1. Map AI data flows end-to-end; not just for training, but for inference and retrieval
2. Treat “always on” as a design constraint, not an SLA target
3. Design for variance, not averages. Users experience the worst five minutes of the network, not the average
4. Make sovereignty enforceable through policy and routing, not documentation
5. Collapse network and security into a single control fabric. Threats operate in real time
6. Instrument everything end-to-end so behavior is visible and provable
7. Build multi-network resilience instead of depending on a single last-mile provider

Why does Network-as-a-Service align with AI systems?

Network-as-a-Service treats connectivity as a continuously managed system, not static infrastructure.

Instead of owning complexity, enterprises consume global connectivity through a single operating layer that spans design, delivery, performance, change, and optimization across regions and providers.

This mirrors how AI systems operate. Continuous. Adaptive. Governed in real time. For CIOs, this shifts effort away from coordination and toward outcomes.

Why do large enterprises choose managed connectivity?

According to IDC, 45% of large enterprises now use managed network services. As networks scale across countries, clouds, and use cases, internal teams spend more time coordinating providers than advancing strategy.

Managed connectivity provides access to specialist expertise, standardized performance, and predictable outcomes without increasing operational overhead.

From network pressure to network advantage

CIOs do not have a connectivity problem. They have a predictability problem. As AI becomes operational, networks can no longer be passive infrastructure. They must act as an intelligent, governable layer that supports continuous change without friction.

This is the shift Expereo helps enterprises make. Turning global connectivity from a pressure point into a platform for progress.

If you are rethinking how your network supports AI, resilience, and global scale, Expereo can help you redesign for certainty.