Your AI Strategy Needs A Rebuild Before Agents Break It

Today, absolutely everyone seems to be talking about AI agents, digital workers that are always-on and capable of doing many jobs, from coding to customer service.

They clearly have the potential to be transformative. But I’m finding many businesses are coming up against a common problem: the foundations their technology stacks are built on simply aren’t up to the job. This is because they were built for an entirely different job.

Today, enterprise IT infrastructure is more often than not built around cloud, human-centered workflows and security models designed around people.

In simple terms, they’re intended for an era when technology was designed to support the work of people, not machines.

With autonomous, always-on AI taking on jobs that traditionally only we’ve done, existing architecture is looking increasingly creaky. In my experience, this is one of the main reasons that agentic AI projects fail to make the jump from pilot to operational use.

However, recent research, such as Deloitte’s Tech Trends 2026, observes that organizations with the most mature AI strategies are starting to reshape their infrastructure around the possibilities of agents.

This isn't always easy to do, and there are some tricky challenges that often crop up. Here, I'll go over some of them and highlight what the priorities should be for professionals and leaders looking to make sure their IT infrastructure is agent-ready.

What This Means

Imagine trying to play basketball on a cricket pitch. That kind of gives an idea of the sort of challenge an agent has operating on traditional IT infrastructure.

The cloud environments hosting enterprise AI systems were designed to run applications and databases, not conduct hugely compute-intensive tasks like generative AI workloads. Computing AI training and inference tasks at enterprise scale requires specialist GPU hardware designed specifically for large-scale AI processing.

This might not cause problems during pilots and trials because there’s less work to do. But when a customer service chatbot is scaled from working with a small niche of use-cases to running the entire department, more money has to be spent on expensive, specialized hardware.

On top of this, most enterprise workflows were created for humans, with the understanding we’d be hands-on in terms of decision-making. Agents are threatening to change this, as they operate autonomously, make plans to hit goals, and interact with third-party systems on our behalf.

Then there’s security. Traditionally, cybersecurity measures control the way people access IT systems. Now, they have to work with machines, understanding when authorized agents need to interact with systems, and defending against unauthorized agentic threats.

Last but not least, there’s the issue of governance. Processes in place to make sure humans stick to the rules and implement appropriate safeguards might not be effective for machines.

A simple example of how an agent might face all of those changes:

An AI agent charged with running a customer service job, like triaging initial customer contacts, needs a budget to pay for the compute tokens that will be crunched by your large-language model.

Then there will be workflow challenges, because inevitably there will be situations where human nuance, insight and 360-degree vision of the landscape are critical.

Security challenges crop up because the agents need to access protected information to do their job. So security infrastructure has to be able to authenticate and manage agentic users as well as human ones.

And governance is an issue in situations where organizations may have statutory responsibilities towards customers, safeguarding, or preventing illegal activity like money laundering.

In many cases, the problem isn’t that AI projects don’t work—they often do—but their effectiveness is limited by the environment supporting them. Making it hard to demonstrate ROI that justifies scaling.

This means if you want to make AI start showing real results, you need to make strategic choices that go far beyond the models and tools in your AI layer.

So What’s The Solution?

First, it’s important to recognize that agentic AI isn’t simply a feature that can be “bolted on” to existing tools or processes. Its arrival on the scene signals a fundamental shift in the way we work.

As we’ve seen, the challenge isn't building better AI models; it's about making sure the supporting IT infrastructure, workflows and security are built with machines in mind, as well as people.

This could mean putting specialized hardware environments together to handle intensive AI workloads, and an architecture tuned to the persistent buzz of agentic activity.

Instead of forcing agents to follow human workflows, businesses must redesign workflows around human-AI collaboration. It’s important to always first think of AI as a way to augment human work, rather than replace it. So start by looking for tasks where humans can use AI to be more productive, efficient and innovative.

Security and governance models must adapt to work in a world where machines as well as people are actors in enterprise systems. This means designing frameworks to monitor, authenticate and regulate agent activity, keeping it running safely within guardrails.

Looking Forward

For me, it’s really exciting to see AI agents breaking out of experimental and pilot stages and taking on real responsibilities.

But as they get more complex and ambitious, the importance of making sure they’re operating on a sound footing grows.

I predict that the organizations that see the best results from AI in the next two to five years will be the ones that fundamentally rethink infrastructure to accommodate agents.

Previous waves of digital transformation, like the arrival of computers, the internet and cloud, made it necessary to rebuild our technological foundations. And lots of organizations that didn’t realize this aren’t around today. Agentic AI has the potential to be as transformative as any of those, and just as likely to need a rethink from the ground up.