AI Adoption & Operations

Legacy Systems and AI: Integration into Established IT Landscapes

Marin Niehues26 January 202611 min read

Robot arm in a production cell

AI can be integrated into established IT landscapes without replacing the legacy systems: an interface layer leaves them unchanged and connects new applications through defined services. This route delivers first value in weeks rather than years, because the interface layer grows step by step and nobody has to replace the core systems first. The real work lies in choosing the data sources and settling responsibilities.

Established systems determine delivery

A legacy system is an application that has grown over years, runs business-critical processes and can only be changed at high cost or risk. An ERP system handles order processing, a mainframe books transactions, a custom-built application steers scheduling. Each of these systems does its job reliably. Because any intervention costs operational risk and scarce specialist knowledge, nobody likes to touch them.

For AI initiatives this has a clear consequence: an application creates value where the data and decisions of daily business reside, and that means the legacy systems. From a delivery perspective one question therefore decides success: can the data and functions of the legacy system be used in a controlled way? If yes, a prototype shows within a few weeks in the real workflow what an application achieves. If no, the initiative first needs access, via data export or user interface automation. Without it, the AI application remains an island next to operations.

Three properties determine the integration effort:

  • Accessibility: are there documented interfaces, or does data exist only in screens and export files?
  • Change risk: how far may an intervention affect live operations, and who is accountable for outages?
  • Knowledge: do people in the company still understand how the system works, and who maintains it?

An inventory answers these questions before a model is selected.

Four integration patterns at a glance

Four patterns have become established for connecting AI to legacy systems. Access to the system and the permissible change risk decide which one fits.

The first pattern is the interface layer, also called API layer or middleware. An additional service in front of the legacy system provides functions and data in a uniform format. AI applications talk only to this layer, and the legacy system stays untouched. This pattern suits any system that offers a minimum of technical connectivity, such as database access or file interfaces. For AI agents, the Model Context Protocol is gaining importance here, an open standard through which language models address tools and data sources in a uniform way. It does not replace the interface layer. It defines how an agent uses it.

The second pattern is data replication. Relevant data is copied at regular intervals into a separate data platform and prepared there for AI applications. The legacy system's operation remains unburdened, and data freshness follows the copy cycle. For analyses, summaries and search across documents this is sufficient, as long as the application does not need real-time data.

The third pattern is user interface automation, known as robotic process automation or RPA. Software operates the legacy system as a person would, through its screens. The pattern bridges systems without interfaces and needs maintenance whenever a screen changes. Combined with AI, RPA handles execution while the model decides, for instance which case belongs in which screen.

The fourth pattern is extension within the system itself. The vendor or an internal team builds AI functions directly into the application. This offers the deepest integration and ties you to the system's release cycle.

What is the strangler fig pattern?

The strangler fig pattern replaces a legacy system function by function with a new system while both run in parallel. Martin Fowler named it in 2004 after the strangler fig, which grows around a host tree until it replaces it. An upstream layer routes requests per function either to the legacy system or to the new service. With each function replaced, part of the traffic moves across, until the legacy system can be switched off.

For AI initiatives the pattern is valuable for one reason: the interface layer that connects an AI application today is the same layer through which the later replacement runs. Building it cleanly is an investment in modernization at the same time. A big-bang replacement with a single cut-over date, by contrast, is the riskiest approach for a business-critical system.

Can AI modernize legacy code such as COBOL?

Yes, generative AI noticeably accelerates the modernization of legacy code such as COBOL, provided the result is tested against the behavior of the legacy system. Language models read large code bases, describe the business rules they contain in plain language and produce documentation that has often been missing for years. They propose translations into modern languages and generate test cases.

The limit lies in certainty. A translated program that looks plausible can calculate differently from the original in edge cases, for instance in rounding or date logic. Every initiative therefore needs three safeguards. First, old and new system run in parallel for a period with the same inputs. Second, automated tests compare the results. Third, experts who know the legacy system review the extracted business rules. The greatest lever lies at the start: analyzing and documenting an unknown code base shortens the planning of a replacement considerably. The article AI in software development describes how to assess AI productivity in development reliably.

Data access before model choice

In many initiatives the first discussion revolves around the model. For delivery a different order works better: data access first, then the application, the model last. A model can be swapped. Missing data access stops the entire initiative.

A simple test applies to every data source. First, it must be clear which system is the authoritative source for a piece of information. Second, a named person in the business unit must decide on the meaning of the fields. Third, quality and freshness must suffice for the purpose of the application. The article Data quality: why it decides AI success sets out how to assess data quality.

Established landscapes often contain several systems with overlapping data, each with its own state. An AI application that works on all of them produces contradictions that a person used to resolve silently. Setting the leading source is therefore one of the first decisions and rests with divisional leadership.

What does it cost to integrate AI into an ERP system?

The cost of integrating AI into an ERP system is driven mainly by three variables: the existing access, the number of connected functions and the effort for testing and operation. An ERP system with documented programming interfaces causes far less effort than a heavily customized system whose data is only reachable through screens. External consulting in the German market is typically billed at around 1,000 to 2,500 euros per consultant day, depending on seniority.

Time follows the same pattern. With a documented interface, a prototype in the real workflow is ready after four to eight weeks. Without one, two to three months are added for building access.

The sequence: understand, build, roll out

A staged plan limits risk, because each stage delivers a robust result before the next begins.

In the first stage the existing landscape is understood. Business analysis clarifies which decision or task the AI application improves and which systems are involved. It also quantifies the value of the improvement. This yields a business case that justifies the priority against other applications.

In the second stage a prototype proves the value in the real workflow. Real means: with actual data from the legacy system, used by staff of the business unit.

In the third stage the application moves into operation. It receives an accountable owner and monitoring, along with a route for incidents and user training. The article Delivering AI solutions describes how this transition succeeds. The article From pilot to scale sets out the whole path from use case to regular operation.

Niehues Consulting as integration partner

Niehues Consulting supports AI initiatives in established landscapes from inventory to handover into operation. Marin Niehues works as a management consultant from a delivery perspective: he builds organizations that deliver and brings AI into operation, in projects from VC-funded scale-ups to DAX 40 corporations and in 20+ countries.

The approach follows the Niehues Delivery Model with its phases understand, build and roll out. Two of its review questions matter most for integration into legacy landscapes. Does the prototype prove the value in the real workflow with real data from the legacy system? And which roles, decision paths and interfaces does operation require? The interface layer is therefore built together with the internal or external technical teams who will later run it. Business units are involved and enabled early, so that the application is actually used after the prototype. In the free 30-minute initial consultation you clarify your starting position and integration route, and you can book it via the contact page.

Frequently asked questions

What is a legacy system?

A legacy system is an application that has grown over years, runs business-critical processes and can only be changed at high cost or risk. Typical examples are mainframe applications in COBOL and heavily customized ERP systems. What makes it legacy is less its age than its condition: missing documentation, few experts and missing interfaces.

Do you have to replace a legacy system before using AI?

No. An interface layer or data replication makes the legacy system's data and functions available to AI applications without changing the core. Replacement makes sense when the system is at the end of its life for other reasons. The interface layer then serves as the transition under the strangler fig pattern.

How long does it take to connect an AI application to a legacy system?

With a documented interface it takes four to eight weeks to a prototype in the real workflow. Without one, two to three months are added for building access. An inventory at the start gives a far more accurate estimate than a guess on request.

What is RPA, and when does it make sense for AI?

RPA stands for robotic process automation, software that operates applications as a person would, through their screens. It makes sense for AI when a legacy system offers no interface and building one would be too expensive or too risky. The AI decides, the RPA executes, and every change to a screen requires maintenance of the automation.

What is the Model Context Protocol, and does it help with legacy systems?

The Model Context Protocol is an open standard through which language models and AI agents address tools and data sources in a uniform way. It helps with legacy systems once an interface layer exists, because an agent can then use its functions without one-off adaptation. It does not create access to the legacy system itself.

How do you integrate AI with an ERP system?

You integrate AI with an ERP system through an interface layer or data replication, with a fixed leading source and a clear meaning for each field. The AI application uses only the data that the improved process needs, such as customer master data, terms and order history for an order check. A full extract of all data is neither necessary nor useful.

What role does IT architecture play in integration?

It determines whether individual connections emerge or a reusable platform. The platform requires more coordination at the start and saves time with every further application. The article IT architecture for AI: platform instead of point solutions offers an assessment.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation