For the past two years, almost every conversation about AI has focused on the rapidly evolving technical side – models, copilots and increasingly, agents.

I can see why there is a lot of excitement, as Agentic AI promises something genuinely different. Rather than simply generating content or answering questions, AI agents can reason, make decisions, coordinate activities and act across systems and processes.

But there is a question I find myself asking organisations with growing frequency:

What exactly will those agents connect to?

Because while the conversation has moved on to agentic AI, many organisations are still operating on technology estates designed decades ago. If organisations want to realise the full value of agentic AI, they first need to remove the legacy obstacles standing in the way.

Legacy isn’t just a technology problem

When people hear the phrase “legacy technology”, they often think of ageing applications, unsupported software or outdated infrastructure. The reality is usually much more complicated.

Many legacy systems continue to perform critical functions extremely well. They process payments, manage citizen records, support case management workflows and run essential operational processes every day. Replacing them outright is rarely straightforward or even desirable.

What makes something “legacy” is often not its age; it’s its inability to adapt. Over time, systems become tightly coupled to business processes, organisational structures and regulatory requirements. Documentation becomes incomplete, knowledge becomes concentrated in a small number of individuals, integrations become fragile, and change becomes increasingly risky.

We’ve seen this repeatedly across both the public and private sectors. The underlying challenge is rarely just technical debt; it’s operational debt, process debt and organisational debt accumulating over many years.

Why agents struggle in legacy environments

An AI agent can only make decisions based on the information and systems available to it. That sounds obvious, but it has important implications.

If customer information exists across five separate platforms, an agent has no reliable view of the customer. If those five platforms each hold a different answer, an agent has no way to know which one to trust. If key decisions depend on spreadsheets, email approvals and undocumented workarounds, an agent cannot consistently execute processes. If data quality is poor, the outputs produced by agents will be poor as well.

As I noted in a recent discussion about AI adoption, one of the most common challenges organisations face is that the information needed to make decisions is spread across multiple disconnected systems. Before introducing intelligent automation, that problem must be addressed. Put simply, agents need context; and context is often exactly what legacy estates struggle to provide.

The four legacy obstacles organisations should tackle first

I believe organisations should focus on identifying the specific elements of their estate that prevent AI systems from operating effectively.

1. Fragmented information

Most organisations have no shortage of data. The problem is accessibility. Information is distributed across multiple systems, departments and ownership boundaries. Different teams maintain different versions of the truth. Valuable data exists but is difficult to discover, integrate or trust.

For agentic AI, this creates significant limitations, because agents are only as effective as the information available to them. Before deploying AI at scale, organisations should understand where key information resides and how it can be accessed safely and consistently.

2. Systems with no integration pathway

Many core platforms were built long before APIs, event-driven architectures or modern integration patterns became commonplace.

Agentic systems frequently need to interact with multiple applications to complete a task. They may need to retrieve information, update records, trigger workflows or coordinate activity across departments.

Sometimes integrations may exist but are poorly documented, inconsistent and unreliable for an autonomous agent to depend on. In other cases, there are no integration pathways at all, just a screen built for users. As organisations try to integrate these systems, they often discover that their biggest barrier to AI adoption isn’t the AI at all, it’s the architecture surrounding it.

3. Brittle business processes

Legacy technology often creates legacy behaviours. Processes evolve around system constraints rather than user needs. Teams compensate through manual workarounds, and critical decisions rely on human knowledge rather than documented rules.

These processes may function adequately today, but they are difficult to automate. Before organisations ask, “How do we introduce AI agents?”, they should first ask, “Do we truly understand how this process works?” Agentic AI performs best when operating within clear, observable and well-understood workflows.

4. Operational risk and governance concerns

Many organisations rightly approach AI cautiously. When systems underpin critical public services, financial transactions or regulated activities, introducing autonomous decision-making raises legitimate concerns around accountability, transparency and control.

The answer is not to avoid innovation. It is to create appropriate guardrails. Best practices like least-privilege access and audit logging aren’t new, but they matter more than ever when it’s an autonomous agent, not a person, doing the accessing.

That means understanding where autonomy is appropriate, where human oversight remains essential, and how governance frameworks need to evolve alongside technology. My colleagues Colin Eberhardt and Andy Scotland expand on the topic in this blog post.

Modernisation before transformation

One of the biggest misconceptions in technology is that transformation begins with a new platform. In reality, transformation begins with understanding your systems, your data, your users, and the constraints that make change difficult.

This is why I often encourage organisations to think in terms of capability building. Sustainable change comes from improving the foundations that allow new technologies to succeed, not from deploying the latest tool and hoping it creates value.

The opportunity ahead

The good news is that organisations do not need a perfect technology estate before they can begin adopting agentic AI; few have one. The goal is not wholesale replacement, but rather to identify the obstacles creating the greatest friction and address them pragmatically.

That might mean improving access to data, introducing APIs around critical platforms, simplifying complex workflows, improving governance, or modernising specific systems that prevent wider progress. These are not glamorous activities and they rarely generate headlines, but they are precisely the things that determine whether agentic AI becomes a transformative capability, or just another promising pilot that never reaches production.

The conversation around AI has shifted from experimentation to execution. As organisations move towards agentic adoption, the question is no longer whether agents can create value, it’s whether the environment around them will allow them to.

And that’s why removing legacy obstacles remains one of the most important foundations for successful AI adoption.