Legacy systems and Agentic AI are often seen as opposing forces, but Luis Blando, CPTO, OutSystems, says together they can unlock institutional knowledge, accelerate innovation and power the next phase of Digital Transformation.
Today’s CEOs are being pulled apart. They are racing to deliver AI breakthroughs while struggling to keep fragile legacy systems alive amid shrinking budgets and a dwindling pool of talent.
On the surface, those priorities look irreconcilable. One faces forward, towards Agentic AI and autonomous systems. The other faces backwards, towards COBOL, mainframes and technical debt.
Conventional wisdom says these are conflicting goals. Yet they are actually the same problem.
Legacy systems are often perceived as blockers to AI adoption. They are, however, better understood as an untapped goldmine of institutional knowledge and process logic. They encode decades of business rules that no documentation repository can fully capture.
Agentic AI, deployed on the right kind of platform, can turn that dormant knowledge into a strategic asset.
The question then is not how to escape legacy. It is how to safeguard it. To ensure that the thousands of hours of development that went into the system continues to provide value, both now and well into the future.
Agentic AI is ready, yet the surrounding architecture is not
Agentic AI systems that can reason, plan and act with minimal human input are no longer theoretical.
Our latest OutSystems survey of 550 software executives shows that almost half of organisations globally are already integrating agentic AI into applications and workflows.
A further 28% are actively piloting solutions and only a tiny fraction have no plans at all. The technology is advancing quickly and the appetite to reap the benefits dangled by the new technology is clearly apparent.
And yet many initiatives stall. Our survey uncovered that only 40% of European organisations have integrated Agentic AI so far, compared with 50% in North America and 60% in Asia.
So why is the continent moving slowly compared with the rest of the world? Regulation, especially regarding the EU AI Act, is clearly an issue. Yet there is another factor at play too in legacy-heavy environments that are difficult to connect, observe and safely automate.
The problem is not that AI is immature. It is that the surrounding architecture was never designed for AI-driven autonomy.
Legacy is not dead weight, it just needs reactivating
For most enterprises, legacy systems are still the beating heart of their operations. They process payroll, manage supply chains, handle claims and reconcile financials. Over decades, they have absorbed the information of thousands of complex processes and policy changes. Each bit of code is there because something real happened in the business that needed to be codified and addressed.
That accumulated logic is precisely what we want Agentic AI to understand.
Imagine an agent that does not just read a COBOL-based inventory log but can reason based on it. This technology would detect that a pattern of orders will create a shortage two weeks from now, model the impact on key customers and automatically initiate a purchase order in a modern Enterprise Resource Planning system.
For example, in insurance claims handling, an agent can access policyholder data from a legacy mainframe, integrate it with modern inputs like image recognition and automatically approve a claim via a modern payment API. This can now happen without a major overhaul or even a complete reboot that potentially could compromise operational continuity, with AI taking care of the algorithmic overhaul that otherwise would be too complex.
In this model, the legacy system remains the authoritative system of record. At the same time, the AI agent becomes the catalyst for action, executing decisions and orchestrating modern workflows around that trusted data.
To reach this ideal, we do not need to rip and replace systems. Rather, we need a way to connect Agentic AI to the data and behaviours locked inside those legacy platforms, then orchestrate actions across old and new systems in a controlled way.
The ‘Agent Workbench’ needs to be a bridge, not a bulldozer
This is where the concept of an Agent Workbench becomes powerful. It provides a unified environment where teams can design, connect and govern AI agents that span both modern and legacy systems. It offers a controlled alternative to the sprawl of scattered automation scripts, brittle point-to-point integrations and ungoverned RPA bots.
Instead of building one-off integrations, organisations define reusable connectors, guardrails and workflows that enforce consistency.
The Workbench ensures safe orchestration by embedding the things that matter operationally – observability, sandboxed execution, fallback and retry rules, circuit breakers and rate-limiting. Agents can call legacy APIs, read logs, interact with modern SaaS and hand off to humans to reduce risks of uncontrolled fully autonomous operations. Governance and security are not added later; they are built into how every agent is designed and executed.
In our research, 95% of organisations said they plan to increase AI investments over the next 12 months and roughly two-thirds are already seeing improvements in software quality and developer productivity. But the same research shows that 64% of executives cite governance, security and compliance as their top concerns and 44% worry about AI and tool sprawl creating new technical debt.
An Agent Workbench directly addresses that tension. Let CIOs say ‘yes’ to more agentic use cases without the potential headaches caused by the proliferation of ungoverned tools. Crucially, it consolidates how agents are built and run and makes it feasible for a small central team to oversee a large, distributed automation footprint.
A unified AI development platform, when built on abstraction and governance layers long used in low-code platforms and now extended for agentic systems operating at enterprise scale, can simplify connections to legacy systems. It turns them into visual components that agents can interact with while abstracting complexity, enforcing governance and security by design and making workflows reusable and maintainable.
It enables developers and increasingly less technical employees to create multi-step agentic workflows without having to rewrite the underlying systems.
The goal is not to make legacy disappear overnight. It is to make it composable.
Agents as digital maintenance crews
There is another dimension where legacy and AI intersect: talent.
Around the world, organisations are struggling to hire and retain developers who understand older languages and platforms. At the same time, demand for change on those systems has never been higher.
The worry is that without the right technical expertise to manage systems, every modification feels risky.
Fortunately, agentic AI can help rebalance this equation. With the right platform, agents can act as digital maintenance crews. They can undertake a variety of roles including monitoring system health, scanning logs, flagging anomalies, suggesting patches and even generating code changes that human developers can review and approve.
They can automate repetitive tasks like regression checks, configuration comparisons or dependency updates. Essentially, they can undertake all the tasks that consume time but rarely create visible value. This is not about replacing human expertise. It is about amplifying it.
AI reduces dependency on scarce legacy specialists by offloading analysis and routine changes, but it does not eliminate the need for expert oversight.
For the first time, junior developers working in a modern low-code environment augmented by agentic tools can contribute meaningfully to the maintenance and evolution of legacy systems. They can do this without needing 10 years of COBOL experience. They can ask natural-language questions about how a process behaves, generate changes and rely on the platform to enforce guardrails and reviews.
In a world where 69% of software executives expect AI to create new specialised roles and 63% anticipate significant reskilling within development teams, as our joint research highlights, this augmentation model is critical. It turns talent scarcity into an opportunity to build more adaptable, AI-literate teams.
Two problems, one strategy
For years, CIOs have treated ‘modernise legacy’ and ‘adopt AI’ as separate initiatives, competing for budget, talent and attention. In reality, they are deeply intertwined.
Modernising the core is almost impossible at the speed the business demands without AI. And AI, if it ignores the core, will never deliver more than surface-level wins.
Success in the Agentic AI era is not about companies having the fewest legacy systems. It is about learning how to make those systems work smarter.
Legacy is not the end of the story. With the right platform and the right mindset, it is where the next chapter of AI-driven Digital Transformation begins. The AI winners will not be those who ignore legacy, but those who strategically incorporate it into their AI architecture, turning it into their AI’s greatest asset.
