When AI projects struggle to scale, the usual suspects are data quality, cost and skills. The network rarely makes the list, yet it often decides whether AI performs as well in production as it did in planning. AI moves data in a different way from the applications your network was built for. When it scales across countries, clouds and providers, predictability becomes the thing that decides whether your investment pays back.
The data backs this up. Enterprise Horizons 2026, IDC’s research sponsored by Expereo, revealed that 46% of organizations whose AI beat expectations credited that strong network performance. Among those whose AI fell short, 26% pointed to an inadequate network. Yet only 5% of companies say their networks are fully ready for what’s coming.
So while boardrooms debate which models and tools to back, the outcome often depends on something far less visible. Here’s why, and what infrastructure leaders can do about it.
AI is exposing problems you’ve lived with for years
Getting data reliably from A to B has always been hard. For a long time, most businesses could work around the gaps. A slow link here, an unreliable provider there… People adapted, and the business kept moving.
“AI is now exposing the problems that we ignored for a long time, or hoped would go away”, said Julian Skeels, Chief Digital Officer at Expereo, speaking at the Gartner IT Infrastructure, Operations & Cloud Strategies Conference in London.
It’s the same story behind many of today’s big infrastructure conversations. Sovereignty, cloud repatriation and edge computing all come back to one question: can you move data reliably to where it needs to be? AI raises the stakes on that question. A workaround that was fine for a person becomes a bottleneck for a system that never stops.
Why AI traffic is different
Think about how a person uses a business application. They make a transaction or two a minute in a CRM. Large amounts of data come down the pipe. If something fails, they hit refresh, try again and move on.
AI doesn’t work like that. “Behind AI is millions and millions of micro-decisions that need to happen all the time”, Julian explained. Each one travels complex routes across your network to reach the data inference needs, and it has to happen globally.
The type of dependency changes too. Traditional networks mostly connect people to applications. AI increasingly connects applications to data, models to business systems, and automated processes to each other. One AI-driven workflow can rely on several systems spread across different clouds, countries and locations. If any link in that chain wobbles, the whole process feels it.
Moving workloads closer to the edge doesn’t remove the problem either. Processing data near where it’s created sounds like a fix. But the edge still needs to talk to the cloud and to other services, so the network path still decides the outcome.
Reliability and predictability are becoming the new currency of business.
Julian Skeels
Chief Digital Officer, Expereo
The lab isn’t the real world
Plenty of businesses can build an impressive AI prototype. The lab has the best data, the best engineers and the best users. Everything is set up to succeed, because everyone wants this exciting new thing to work.
Then the application goes live in new regions and, as Julian said, it all starts to change. “They try and deploy it in one continent, then another, and suddenly the assumptions built into that application become very, very apparent.”
Those assumptions are often about the network. The pilot ran on a fast, stable connection close to the data. In production, the same application might depend on different carriers, access technologies and routes in every country. Performance that looked guaranteed in the lab becomes unpredictable at scale.
A pilot proves the idea works. Production proves whether your network can carry it, at every location, every day. Planning for that gap early is far cheaper than discovering it after launch.
Faster pipes or predictable pipes
For years, global networking has been about speed: bigger pipes, faster links. Julian compared it to adding another lane to a motorway. It helps, up to a point – but the time is often lost at the junctions and intersections along the way.
For AI, that means looking at the full path your data takes. Where is it flowing? Under what policies? How consistent is its performance from one hour to the next?
Consistent global processes let us deliver service at scale. Close relationships with our customers and account teams tell us when a situation calls for a local or regional approach, and our people have the freedom to take it.
In many cases, faster is not always the solution. Predictable is what we need for the future.
Julian Skeels
Chief Digital Officer, Expereo
Latency, jitter and packet loss decide whether an AI-driven process runs smoothly or breaks. A link with huge capacity but unpredictable delays can do more damage than a smaller one that performs the same way every time. Consistent performance across the whole route matters more than headline bandwidth on any single link.
The hidden cost in the AI business case
AI apps get the attention and the budget because they’re easy to see. They have a slick front end and a compelling demo. Julian calls this the infrastructure paradox: “We really value the things that we can see.”
Every one of those apps assumes the infrastructure behind it will be reliable. When it isn’t, the cost shows up fast. At Gartner’s conference, Julian shared the example of a global manufacturing customer where one minute of lost network connectivity costs around $10,000, because the computer vision applications on the factory floor stop working.
And the damage goes beyond downtime. Unreliable connectivity slows adoption, because people stop trusting tools that fail them. It interrupts the workflows AI was meant to speed up and prevents the business reaching the return it was promised.
That cost rarely appears in the original AI business case. It should. Infrastructure investment deserves the same financial scrutiny as the AI it supports, and the same clear line to business value.
What infrastructure leaders can do now
AI is changing how organizations work across teams, not just within them.
AI is collapsing boundaries within organizations.
Julian Skeels
Chief Digital Officer, Expereo
Networking and security have already merged, because one can’t be managed without the other. AI is pulling infrastructure into the same conversation as strategy and finance. For infrastructure teams, that’s a chance to shape AI plans instead of inheriting them.
“What AI has challenged all of us to do is to get out of our seat, understand what is happening in the business, and help shape those decisions”, Julian said.
A few practical places to start:
- Talk in business outcomes. Network teams measure latency, packet loss and availability. The business feels delayed transactions, stalled processes and lost revenue. Translate one into the other.
- Map the full path. Look beyond individual links to how data flows end to end, across every cloud, country and provider.
- Plan for production, not the pilot. Test AI applications under real conditions in the regions where they’ll run, before the rollout plan is fixed.
- Design resilience in early. Continuity is far easier to build before a disruption than to buy during one.
- Invest where it counts. Not every AI use case needs a network upgrade. Focus on the processes where network performance limits value, and measure what’s at risk.
Building a network ready for AI at scale
Expereo designs and manages global networks for enterprises in more than 190 countries, as one accountable partner. That means one team responsible for performance across every provider and region, instead of a patchwork of contracts.
Enhanced Internet uses AI-powered routing to send your traffic along the best-performing path, not just the shortest one. It constantly monitors network performance and adjusts in real time, so your AI applications get consistent results wherever they run. And you get visibility of performance across your whole estate, so you can spot problems before your users or your AI feel them.