As focus shifts from GenAI to agentic, many organisations are still pinning down their AI strategy. But defining the approach is only the beginning: successfully moving AI from pilot to company-wide deployment is a major challenge, with only 6% managing to make it work at scale.
Pilot AI projects tend to go smoothly because they’re limited in scope. These initiatives are often carried out in a single, controlled environment that doesn’t drive significant complexity. They’re chosen because they’re straightforward to deploy: few dependencies and integrations, one or two teams, and limited legacy tech to navigate.
But this doesn’t reflect the reality of wider deployment. As one VP at a transportation company said to us recently, “When we roll out the solution into multiple sites or different countries, then the real picture comes out. Every location has a different set-up, client systems are different – and this is the time when complexity becomes clear.”
This is where AI momentum can stall: when a promising use case collides with the reality of the enterprise. AI is not creating the complexity. It is exposing fragmented networks, systems and operating models that have built up over years.
There are many factors at play here, like poor data quality, inadequate governance and ineffective change management, but one of the major issues is complex, disjointed and legacy infrastructure. More than four out of five (82%) organisations in our recent survey say that their networks, security, Unified Communications and cloud systems are not fully integrated. And two-thirds have failed to exploit AI opportunities because of limitations in their infrastructure.
AI demands guaranteed performance
Supporting agentic AI isn’t just about bandwidth. Although this plays a role, to achieve the “breathable infrastructure” needed for AI to thrive, we need to ensure that the workloads have access to elastic, low latency, deterministic performance and integrated and secure connectivity.
One of our manufacturing customers, for example, has a static MPLS network with siloed factories, air-gapped OT and security based on the outdated perimeter model.
To get more value from the company’s data and to safely drive AI into its workflows and supplier ecosystem, the manufacturer needs to converge IT and OT without increasing risk, remove silos between systems and locations, guarantee network performance and responsiveness, secure data flows from plant to cloud and across geographical borders, and strengthen security while allowing users, AI agents and suppliers to access the information and systems they need.
The company’s current legacy infrastructure is complex and poorly integrated and simply can’t support these objectives. Its perimeter-based security model wasn’t designed for today’s operational models and working practices. Poor visibility compromises performance and creates blind spots and vulnerabilities.
How has enterprise infrastructure become so complex?
Organisations have evolved rapidly over recent years by adopting cloud services, new security frameworks, AI tools and digital ways of working. Hybrid working and BYOD have become standard in many organisations, bringing new connectivity and security challenges with them.
This often involves workarounds and temporary fixes which add to the fragmentation and complexity. There’s rarely time for teams to stop, review and optimise infrastructure, so configuration drift, tool sprawl and technical debt make the situation even worse.
The average enterprise estate now relies on 16 different vendors, platforms or tools, so comprehensive oversight is almost impossible, and accountability isn’t always clear.
Until infrastructure complexity is addressed, organisations will continue to prove that AI works in controlled pilots while struggling to make it work across the business. Simplifying the foundations is what turns AI potential into repeatable, enterprise-wide value.
Explore how infrastructure complexity could be affecting your organisation and take the Complexity Check: https://www.colt.net/go/beat-the-ai-stall
About the Author:
Giovanni Pirolli is Colt’s Principal Technology Advisor and a trusted adviser to global enterprises navigating digital transformation. He specialises in helping organisations simplify complex technology estates, strengthen security, accelerate AI readiness and turn strategic objectives into measurable business outcomes.