2025 brought a revolution in the approach to artificial intelligence in corporate environments. But what about organisations that cannot or will not move their data to the cloud? For many Polish companies and public institutions, the answer was: let’s build our own AI infrastructure that runs wherever our SAP systems run. This is the story of how it was achieved.
When the first signs of agentic capability in the SAP ecosystem appeared towards the end of 2024, many IT specialists asked themselves: does this only apply to those who have moved to the cloud? The answer, as it turned out, was far more nuanced - and far more promising for organisations working in on-premise environments.
The starting point: SAP on-premise and the need for intelligence
Let’s start with the facts. According to various estimates, the vast majority of SAP installations worldwide still run on-premise. This is not a matter of technological lag or lack of vision. It is often a deliberate decision driven by industry regulation, security requirements, the specifics of business processes, or simply the economics of a given organisation.
In Poland, the situation is particularly interesting. Public institutions often have to keep data on their own infrastructure because of data protection legislation and regulatory requirements. The financial sector and healthcare likewise operate in a tightly regulated environment, where the question of where data is physically located is of fundamental importance.
At the same time, pressure to deploy AI-based solutions is growing month by month. Competitors are using chatbots for customer service, automating decision processes, and generating reports and analyses in a fraction of the time it used to take. Organisations working exclusively in on-premise environments began asking a fundamental question: can we have both?
An architecture of possibility: the Model Context Protocol changes the rules
The breakthrough of 2025 turned out to be the Model Context Protocol, or MCP. It is an open standard that can be compared to a universal USB port for the world of artificial intelligence. MCP defines a common language that large language models can use to communicate with external systems and tools.
Before MCP gained traction, every AI integration required building a dedicated solution. Want to connect a language model to SAP? Build your own interface. Want to add another data source? Write more dedicated code. MCP changed that dynamic by offering a standardised way for AI agents to discover and use external tools.
What is particularly significant for on-premise environments is that MCP works both in the cloud and on local infrastructure. MCP servers can be deployed as containerised applications in any environment, opening the door for organisations that, for various reasons, cannot use public cloud services.
MCP servers for SAP: from concept to reality
2025 brought an explosion of MCP solutions dedicated to SAP environments. The developer community, supported by official SAP initiatives, created a whole spectrum of servers covering different use cases.
One solution deserves particular attention: it allows an MCP server to run directly in the ABAP layer of SAP ECC and S/4HANA systems. Importantly, it supports ABAP versions as far back as 7.01, meaning integration is possible with genuinely mature installations. The architecture is designed to avoid middleware, giving organisations full control over tool definitions, the data exposed, and the mechanisms for creation and updates.
That sounds technical, but the practical implications are enormous. An AI agent can now communicate with an SAP system using natural language. It can query tables and CDS views, call BAPI function modules, and even trigger transactions - all without installing an SAP GUI client and without leaving the organisation’s network.
The computing power challenge: enter NVIDIA
Integration is one thing, but running local language models requires appropriate computing power. This is where NVIDIA comes in with its Enterprise AI Factory vision - a full-stack, validated architecture that allows organisations to build their own AI factories on local infrastructure.
The concept of the AI factory, which Jensen Huang has consistently promoted for several years, takes on new meaning in the context of corporate environments. The data centre stops being a place for storing and processing data. It becomes a factory producing intelligence. Just as factories producing cars or electronics take in raw materials, process them, and deliver a finished product, an AI factory takes in data as the raw material and delivers insights, decisions and automation as the product.
In 2025 NVIDIA brought RTX PRO servers to market, built on Blackwell architecture cards, allowing organisations to convert existing data centres into AI-ready infrastructure without a complete rebuild. This is not a minor change. Disney, SAP, Hitachi and Hyundai Motor Group are just some of the companies that have already adopted these solutions.
For mid-sized enterprises and public institutions, the key factor is the ability to run open language models on relatively accessible hardware. The Fraunhofer Institute recommends models based on the Mistral 7B architecture, which run on hardware comparable in capability to an advanced gaming PC, available from around ten thousand euros. The Mixtral 8x22B model, the most powerful European open-source model, operates fluently in five languages and was trained on more than 140 billion parameters.
“A year ago, talking about running a local language model integrated with SAP sounded like science fiction. Today it’s a matter of choosing the right components and architecture. NVIDIA has shown that turning a data centre into an AI factory doesn’t require a revolution, only an evolution.” - Michal Korzen, CTO, SNOK
Orchestration: UiPath Maestro as the conductor of agents
With computing infrastructure and the ability to communicate with SAP in place, the remaining question is orchestration. How do you manage multiple AI agents? How do you ensure RPA robots, language models and employees work together coherently? This is where UiPath comes into play.
UiPath, known primarily for its robotic process automation solutions, underwent a fundamental transformation in 2025. The company moved from being a provider of RPA tools to being an agentic automation platform. A key element of this transformation is Maestro, an orchestration layer connecting AI agents, robots, tools and people into one integrated system.
Importantly, Maestro supports the Model Context Protocol, meaning it can manage agents from different platforms, including Google Vertex, Microsoft Copilot, Databricks and NVIDIA. An organisation can deploy different agents for different tasks. The finance department might run an agent specialising in reporting, while operations uses a different agent for demand forecasting. Maestro coordinates their work within a single process.
For on-premise environments, the key capability is the ability to deploy UiPath in self-hosted mode. The platform can be run on an organisation’s own infrastructure, retaining full control over data and processes. Daniel Dines, founder and CEO of UiPath, describes this philosophy as “people acting as conductors, while most of the operational work is carried out by agents and robots”.
TIME included the UiPath platform on its list of the best inventions of 2025, recognising its ability to manage distributed AI agents while providing appropriate human oversight. That last point is particularly important in the context of public institutions and regulated sectors, where every decision made by AI must be auditable and explainable.
The Polish context: why on-premise is not parochial
Discussions about digital transformation often feature a narrative in which organisations that remain on on-premise solutions are dinosaurs defending themselves against inevitable extinction. Reality, at least in the Polish context, is far more nuanced.
Public sector institutions operate under strictly defined data-processing regulations. Cloud Computing Cybersecurity Standards, the Government Cloud, GDPR requirements - all of this creates an environment in which the decision on data location is not an IT department whim, but a necessity driven by legislation.
According to research conducted by the Central Statistical Office for Informatics [Centralny Ośrodek Informatyki], most public institutions using cloud solutions prefer private clouds. The reasons are understandable: full control over infrastructure, the ability to adapt to specific regulatory requirements, and better risk management. We observe similar trends in the financial sector and healthcare.
This does not mean these organisations must fall behind in the race for artificial intelligence. Quite the opposite. 2025 showed that it is possible to build advanced AI infrastructure entirely on one’s own infrastructure, retaining all the benefits of data control while gaining access to the latest technological capabilities.
“Our clients in the public and financial sectors no longer ask whether to deploy AI, only how to do so without compromising on security. The answer is: you can have both. An on-premise agentic platform is not a compromise - it’s a deliberate architectural choice that responds to real business needs.” - Jacek Bugajski, CEO, SNOK
Implementation in practice: from theory to a working solution
What does a real-world deployment of an AI platform for an on-premise SAP environment actually look like? The process can be broken down into several key stages, each carrying its own specific challenges and decisions.
The first step is selecting and deploying a local language model. For organisations processing data in Polish, it is worth considering models optimised for European languages. Mixtral, developed by the French company Mistral AI, delivers excellent results with reasonable hardware requirements. The model can be run using tools such as Ollama, which simplify deployment to a minimum.
However, for organisations requiring the best possible understanding of Polish linguistic and cultural context, 2025 brought breakthrough home-grown solutions. Bielik, created by the SpeakLeash Foundation in cooperation with the Academic Computer Centre Cyfronet AGH, is the first Polish open language model available under the Apache 2.0 licence. Version 3.0 was presented in May 2025 at the GOSIM AI Spotlight conference in Paris, and in June the model was added to the NVIDIA NIM APIs catalogue, gaining global recognition. Bielik, trained on the Helios and Athena supercomputers, handles the nuances of the Polish language superbly and is available in versions ranging from 1.5 to 11 billion parameters. Notably, InPost launched the “Feed Bielik” [“Nakarm Bielika”] campaign, engaging fifteen million app users in the model’s development.
At the same time, the Ministry of Digital Affairs was developing the PLLuM project - the Polish Large Language Universal Model. This is an initiative of a consortium of six leading Polish scientific institutions led by Wrocław University of Science and Technology, with the participation of NASK, the Institute of Computer Science of the Polish Academy of Sciences, the Information Processing Centre, the University of Łódź and the Institute of Slavic Studies of the Polish Academy of Sciences. PLLuM is distinguished by particular adaptation to public administration terminology and was designed with applications in the state sector in mind. A family of eighteen models, using from 8 to 70 billion parameters, is already publicly available. Crucially for on-premise environments, PLLuM can be deployed locally, on an organisation’s own infrastructure, without dependence on foreign cloud providers.
The next element is computing infrastructure. NVIDIA AI Enterprise offers a validated technology stack combining hardware, software and networking into an optimised, AI-ready environment. For smaller organisations, an alternative is servers with RTX cards, which, with the right configuration, support models of a size suitable for practical business applications.
Integration with SAP requires deploying an MCP server. For ABAP systems, add-ons are available that allow the server to run directly in the SAP application layer. This means an AI agent can communicate with the system without the need for additional middleware components, simplifying the architecture and reducing the potential surface for issues.
Finally, there is the orchestration layer. UiPath in self-hosted mode allows the entire ecosystem to be managed from a single place. Process definitions, run schedules, performance monitoring and audit of all operations take place within the organisation’s controlled environment.
Use cases: where AI is changing everyday work
Theory is one thing, but where do agentic platforms actually deliver value in on-premise SAP environments? Practice points to several particularly promising areas of application.
Conversational interfaces to business data are the first and most obvious use case. Instead of navigating complex SAP transactions, a user can ask a question in natural language: “Show sales orders from the last quarter for customer X with payment delays exceeding 30 days.” The AI agent processes the query, interrogates the relevant tables and CDS views, and then presents the results in a readable form.
The second area is support for decision-making processes. An agent can analyse receivables data, detect patterns of payment delays, identify customers requiring particular attention, and suggest actions. According to UiPath research, such solutions can shorten the time needed to process receivables disputes and improve an organisation’s financial liquidity.
The third case is automating the creation of documentation and reports. An agent can generate a monthly report by pulling data from multiple sources, carrying out trend analysis, and formatting the results in line with the organisation’s standards. Importantly, the entire process takes place in the local environment, without sending sensitive business data to external services.
The fourth area is support for ABAP developers. The new AI SDK for ABAP, introduced by SAP, allows generative capabilities to be integrated directly into code. A developer can ask an agent to generate test code, explain how an existing program works, or suggest optimisations. According to SAP data, this approach can shorten coding time by as much as twenty per cent.
The SNOK experience: successfully connecting two worlds
SNOK has specialised in the SAP ecosystem for more than twenty-five years, combining deep technical knowledge with a practical understanding of clients’ business needs. When the first opportunities to integrate AI with SAP systems emerged, the natural next step was to explore how these technologies could serve organisations operating in on-premise environments.
Our practice shows that the success of an agentic platform deployment depends on several key factors. First, deep knowledge of the specifics of the client’s SAP systems. No two SAP installations are identical - every organisation has, over the years, adapted the system to its own processes, creating unique extensions and modifications. Effective AI integration must take this specificity into account.
Second, a realistic approach to expectations. AI is not a magic wand that will solve all of an organisation’s problems. It is a tool that, properly deployed, can significantly improve specific processes. Our methodology begins by identifying use cases where AI will deliver measurable value, and then builds the solution iteratively, learning at each stage.
Third, security as a foundation, not an add-on. In on-premise environments, data security is often the main reason for choosing this architecture. Every AI solution must respect existing security policies, integrate with authorisation mechanisms, and ensure full auditability of actions.
Through partnerships with SAP, UiPath, NVIDIA and other technology leaders, SNOK is able to deliver comprehensive solutions that combine the best elements of each platform. Our experience in S/4HANA conversion projects, automation deployments and building secure IT architectures allows us to guide clients effectively through the transition towards intelligent automation.
“The key to success is not the technology itself, but understanding how that technology fits into the client’s existing ecosystem. Connecting an AI model to SAP is a technical matter. Making that connection deliver real business value is an entirely different story.” - Michał Korzeń, CTO, SNOK
Challenges and pitfalls: what to bear in mind
The optimism surrounding the possibilities of agentic platforms should not obscure the real challenges facing organisations planning such deployments.
The first and often underestimated challenge is data quality. An AI model is only as good as the data it works with. If data in the SAP system is incomplete, inconsistent or outdated, no amount of artificial intelligence will fix that. Deploying an agentic platform often forces an earlier clean-up of master data, which is itself a significant project.
The second challenge concerns competencies. Maintaining local AI infrastructure requires specialist knowledge that goes beyond the traditional skills of SAP administrators. The organisation must either build internal competencies or secure support from a technology partner. Experts from Bloor Research warn that layering intelligent agents on top of outdated structures can accelerate dysfunction rather than resolve it.
The third challenge concerns the operating model. Introducing AI agents changes the way people work. Processes that previously required direct human intervention can now be partially or fully automated. This requires rethinking roles, responsibilities and escalation paths. It is not about replacing people, but about changing the way people and machines work together.
The fourth challenge is cost. AI infrastructure, even in an on-premise variant, requires significant investment. Graphics cards, servers, software, licences, training - all of this adds up to a budget that must be justified by return on investment. The good news is that the cloud model shifts capital expenditure into operating expenditure, but the total sum does not disappear - it is simply spread over time.
Outlook: what 2026 will bring
Looking at the pace of development in agentic technologies, a few predictions for the near future can be ventured.
SAP has announced full MCP support in SAP HANA Cloud from the first quarter of 2026. This means Joule agents will have direct access to the database engine, allowing them to answer complex questions drawing on broad data context. For on-premise environments, similar capabilities can be expected through partner and community solutions.
NVIDIA continues to develop the Blackwell architecture and has announced further generations of platforms. The upcoming Vera Rubin platforms offer even greater performance at lower power consumption. For organisations planning AI infrastructure investment, this is important information when building a multi-year strategy.
UiPath continues to develop Maestro’s capabilities, adding increasingly advanced case management functions, process apps for business users, and deeper integration with third-party AI platforms. The direction is clear: from automating individual tasks towards orchestrating entire business processes connecting people, robots and AI agents.
SAP is introducing SAP-RPT-1, a new AI model optimised specifically for prediction on tabular data. Unlike classic language models that predict the next word in a sequence, SAP-RPT-1 predicts the next field in a table row. This is a fundamentally different approach that may better address the needs of working with structured business data.
“The most important lesson of 2025 is that the future of enterprise AI is not binary. It’s not a choice between the cloud and on-premise, between SAP and its competitors, between automation and people. The future belongs to those who can bring these worlds together into a coherent whole.” - Jacek Bugajski, CEO, SNOK
A revolution that requires no revolution
2025 showed that the transition towards intelligent automation does not require abandoning existing investments and competencies. Organisations working with on-premise SAP systems now have access to the tools and architectures needed to build advanced agentic platforms without moving to the public cloud.
The Model Context Protocol democratises AI integration with corporate systems. NVIDIA provides computing infrastructure ready to work with local language models. UiPath offers an orchestration layer connecting all these elements into a coherent system. SAP is developing native AI capabilities within its products.
For Polish organisations, particularly those operating in regulated sectors, this combination opens the way to the benefits of AI without compromising on security or regulatory compliance. It is not an easy path - it requires investment, competencies and a well-considered strategy. But it is a path that can be walked today.
The question is no longer whether to build a local AI factory. The question is: when to start, and with whom.
SNOK Sp. z o.o. | SAP Silver Partner | UiPath Platinum Partner | www.snok.pl
Tech Thursday with SNOK | January 2026
Would you like to see this in practice or discuss a deployment for your company? Get in touch - we will respond within 48 hours.