It is estimated that 85% of generative AI projects never reach the production stage. Sounds familiar? If you run a company investing in GenAI, that statistic should concern you. But rather than analysing yet another report on the potential of AI, let us talk about what actually determines the success or failure of such initiatives.
Looking at the technology market in 2025, we see a paradox. On one hand, we have language models reaching a level of intelligence that seemed impossible just three years ago. On the other, enterprises that enthusiastically launched into GenAI deployments are now grappling with systems that perform brilliantly in demonstrations but fail in business reality.
The problem does not lie in the technology itself. It lies in how we approach building agentic systems. And that brings us to the heart of the matter.
Context engineering instead of prompts
Over the past two years the industry has talked about “prompt engineering” - the art of writing queries for language models. That approach was sufficient when we used ChatGPT for one-off tasks. But when we talk about production systems that must run reliably for months and years, we need something considerably more sophisticated.
Cognition AI, the creators of the Devin system, introduced the concept of “context engineering.” This is a fundamental shift in thinking. Instead of focusing on how to write the perfect query, we focus on how to automatically and dynamically supply the model with the right context at every moment it operates.
Imagine you are building a system to automate customer service. In the traditional approach, you give the agent one large task: “handle the customer’s query.” The agent tries to understand the problem, find a solution and formulate a response - all in one long thought process.
In the context-engineering approach, we break that task down, but not in the way you might expect. It is not about simple step-splitting. It is about ensuring that every decision made by the system has the full context of all preceding decisions, while not being overwhelmed by irrelevant information.
Why multi-agent setups are a trap
When you talk to AI solution vendors, you will often hear about multi-agent architectures. It sounds logical - if one agent can perform a task, a team of agents should do it even better, right?
Unfortunately, reality is brutal. Multi-agent systems that operate in parallel almost always lead to decision conflicts. Imagine two developers working on the same code without communicating - the result would be catastrophic. The same thing happens with AI agents.
Cognition sets out two key principles that should guide the design of agentic systems:
Principle 1: Share context, but share the full traces of agents’ actions, not just individual messages.
When an agent makes a decision, that decision carries hidden assumptions. If the next agent does not see those assumptions, it will make decisions based on its own interpretation, which may conflict with the original intent.
Principle 2: Actions carry hidden decisions, and conflicting decisions lead to poor outcomes.
This is not theory. It is a practice confirmed by hundreds of deployments. Systems that ignore these principles fail in production, even if they look great in demonstrations.
Workflow or agent - a false dilemma
In its guide to building agents, Anthropic introduces a distinction between “workflows” and “agents.” Workflows are systems where the LLM and tools are orchestrated through predefined code paths. Agents, by contrast, dynamically direct their own processes and tool use.
Most companies make the mistake of choosing one or the other. Yet business reality requires combining both approaches. A system that is 100% agentic will be unpredictable and expensive. A system that is 100% workflow will be rigid and unable to adapt.
The optimal solution? A hybrid approach, where critical business decisions are controlled by workflows, while tasks requiring creativity and adaptation are delegated to agents. This is exactly the approach we apply at SNOK, combining the UiPath platform with GenAI models.
Deployment reality - the SNOK case
At SNOK we have been intensively working on GenAI projects for clients across different industries for three years. As a Gold Partner of UiPath, we have a unique perspective on how to combine process automation with generative intelligence.
“The biggest challenge in GenAI projects is not choosing the model or writing the code. It is understanding that we are building systems that must operate in an unpredictable business environment,” says Michał Korzeń, CTO SNOK. “We have seen dozens of projects that worked perfectly in a test environment but fell apart on first contact with production reality. The key is designing with uncertainty in mind.”
Our experience shows that the success of a GenAI project depends on three factors:
1. Proper context management. A language model, however intelligent, without proper context is like an expert deprived of information about the task at hand. At SNOK we developed a Dynamic Context Management methodology, which ensures that every model call has exactly the information it needs - no more, no less.
2. Integration with existing systems. GenAI does not operate in a vacuum. It must work together with ERP, CRM, document systems and dozens of other tools. This is where the UiPath platform shows its strength - it provides integration bridges that let AI models communicate safely and efficiently with enterprise infrastructure.
3. Safety and control mechanisms. In the corporate world, we cannot afford systems that “hallucinate” or make unpredictable decisions. That is why every one of our GenAI projects includes multi-level validation and control mechanisms.
Success story: public administration - synergy between UiPath Document Understanding and an LLM
An interesting case is our collaboration with one of the key ministries, which handles more than 200,000 applications and documents from citizens and businesses each year. The challenge was multi-layered: automating the analysis of diverse administrative documentation while maintaining full compliance with administrative procedures, and exploiting the synergy between UiPath Document Understanding and the Polish language model PLLUM.
“Public administration is an environment where precision of data extraction has to go hand in hand with a deep understanding of legal context,” explains Grzegorz Surdziel, Intelligent Automation Team Consultant at SNOK. “We receive documents in dozens of formats - from hand-filled forms, through scans of documents years old, to modern electronic forms. That is why we designed a system where UiPath Document Understanding handles intelligent digitisation and data extraction, while PLLUM performs advanced substantive and legal analysis.”
The system works in a complementary model, maximising the strengths of both technologies. UiPath Document Understanding handles the technical layer of document processing. The system automatically recognises document types, from grant applications to appeals against administrative decisions, extracting key data with accuracy reaching 97%, even from poor-quality documents. OCR technology enhanced with machine learning copes with handwritten annotations, stamps and multilingual attachments. The system automatically categorises documents, verifies the completeness of required documentation, and creates structured data summaries.
PLLUM takes over when deep substantive analysis is required. The model analyses the extracted content for compliance with applicable regulations, identifies which legal provisions apply in a given case, detects inconsistencies and gaps in documentation, and generates drafts of official correspondence in the appropriate linguistic register. Crucially, PLLUM was further trained on a corpus of Polish official and legal language, so it understands the specific constructions and terminology involved.
A key element was ensuring full data security. The entire infrastructure runs on-premise in a government data centre. Every operation is logged, every decision made by the system can be audited, and citizens’ data never leaves the secured state infrastructure.
The results of the deployment exceeded initial expectations. The time for initial application analysis fell from several days to a few hours. Errors in document categorisation dropped by 85%. Officials now receive fully prepared cases with extracted data, identified legal grounds and draft decisions. This allows them to focus on cases requiring expert judgement and an individual approach, rather than spending hours manually retyping data from paper forms.
Orchestration as the key to success
In its analysis of agentic frameworks, LangChain introduces a concept that perfectly captures our experience: success does not lie in choosing between agents and workflows, but in skilfully orchestrating both.
Imagine a conductor leading an orchestra. He does not play any instrument, but without him the music would be chaos. The same is true of GenAI systems - they need an orchestrator that ensures every element plays its part at the right moment.
In our implementations, UiPath plays the role of that orchestrator. The platform:
-
Controls the flow of data between AI models and enterprise systems
-
Provides retry and error-handling mechanisms
-
Monitors performance and costs
-
Implements business rules and compliance
-
Manages versions and deployment
-
It is precisely this orchestration layer that determines whether a GenAI system runs reliably in production or remains an interesting experiment.
Five steps to GenAI success
Based on our experience, we have developed a methodology for implementing GenAI projects that significantly increases the chances of success:
Step 1: Start from the process, not from the technology
Too many companies start with the question “what can we do with GenAI?” The right question is “which business process will deliver the greatest value once automated with AI?”
We identify processes that:
-
Require processing large volumes of text
-
Contain pattern-based decision elements
-
Are repeatable but not fully deterministic
-
Generate measurable business value
Step 2: Build a proof of concept, but think about production
A PoC is not a game with ChatGPT. It is a prototype that from the outset takes into account:
-
Integrations with source systems
-
Security mechanisms
-
Scaling and performance
-
Operational costs
-
Monitoring and error handling
“We have seen too many projects that worked great as a demo but needed a complete rebuild before production deployment,” says Kacper Wojciechowski. “That is why we build even our PoCs with the idea that they might become the foundation of a production system.”
Step 3: Start with small steps
Do not try to automate the entire process at once. Start with the most valuable fragment. Once that works and delivers value, expand the system.
In one of our projects for a financial company, we started by automating the analysis of loan agreements. Only once that element worked reliably did we add risk report generation, and then decision recommendations.
Step 4: Measure everything
GenAI is not magic. It is a technology whose effectiveness can and must be measured. In every project we implement dashboards showing:
-
Accuracy (response accuracy)
-
Latency (response time)
-
Cost per request
-
Error rate
-
User satisfaction
These metrics allow us not only to monitor the system but also to make informed decisions about its development.
Step 5: Plan for the long term
A model that works brilliantly today may be outdated in six months. That is why the system architecture must allow for easily swapping models, updating prompts, and adapting to changing business requirements.
Mistakes to avoid
Our practice shows that most GenAI project failures stem from recurring mistakes:
Mistake 1: Overcomplicating the architecture
“Simplicity is the ultimate sophistication” - this principle is particularly important in GenAI. Systems with dozens of agents working in parallel look impressive on a diagram, but in reality are a nightmare to maintain.
Mistake 2: Ignoring costs
GPT-4 or Claude 3.5 are powerful models, but also expensive ones. A project that costs 100 złoty a day in the PoC phase can generate bills of hundreds of thousands a month in production. That is why from the outset we design systems with cost optimisation in mind - we use smaller models wherever possible, implement caching, and carefully decide which queries require the most powerful models.
Mistake 3: Lack of control mechanisms
AI can “hallucinate” - generate information that sounds convincing but is untrue. In a production system that is unacceptable. Every one of our projects includes verification mechanisms, cross-checking and human-in-the-loop for critical decisions.
Mistake 4: Underestimating the importance of data
“Garbage in, garbage out” - this principle is even more important in GenAI. A model is only as good as the data you feed it. That is why we devote a substantial part of every project to data engineering - cleaning, structuring and enriching data.
Mistake 5: Forgetting about users
The best AI system is worthless if users do not accept it. That is why we involve end users in the design and testing process from the outset. Their feedback is invaluable.
The role of orchestration platforms
Anthropic rightly notes that most production agentic systems are a combination of workflows and agents. But how do you manage that? This is where orchestration platforms such as UiPath come in.
An orchestration platform is not just a tool for connecting components. It is:
A context management system - ensures that every agent receives exactly the information it needs, in the right format and at the right moment.
A security layer - controls data access, logs all operations, implements business rules.
An error-handling mechanism - when something goes wrong (and it will), the system must be able to recover from it.
A monitoring tool - lets you track what is happening in the system, identify bottlenecks and optimise performance.
An integration interface - connects GenAI with the rest of the enterprise’s IT ecosystem.
At SNOK we use UiPath as our main orchestration platform, though it is not the only solution. What matters is that the platform is:
-
Scalable
-
Reliable
-
Easy to integrate
-
Well documented
-
Backed by a strong ecosystem
The future belongs to hybrids
“Observing the development of GenAI and automation technology, I am convinced the future belongs to hybrid solutions,” says Jacek Bugajski, CEO of SNOK. “Pure automation without intelligence is blind. Pure AI without automation structures is chaotic. Only the combination of both produces systems that genuinely transform business. That is why our strategy of combining the UiPath platform with the latest AI models is not just a response to current market needs - it is an investment in a future where the boundary between automation and intelligence will become increasingly fluid.”
This vision is already materialising. In our latest projects, we see UiPath and GenAI reinforcing one another:
-
UiPath provides structure and reliability, GenAI adds intelligence and adaptability
-
UiPath handles deterministic processes, GenAI copes with ambiguity
-
UiPath integrates with systems, GenAI understands and processes data
-
UiPath controls the flow, GenAI makes the decisions
Practical guidance for leaders
If you run a company considering a GenAI deployment, here is some concrete guidance:
1. Start with a process audit
Identify processes that:
-
Involve highly skilled staff in repetitive tasks
-
Require analysis of large volumes of text
-
Contain decision elements based on experience
-
Generate organisational bottlenecks
These are ideal candidates for GenAI automation.
2. Build a cross-disciplinary team
Success in a GenAI project requires combining competencies from:
-
Domain experts who understand the process
-
AI engineers who understand the technology
-
IT architects who understand the infrastructure
-
Project managers who can manage it all
3. Set realistic expectations
GenAI is not a magic wand. It is a powerful tool that requires time, investment and patience. Expect the first results after 3-6 months, full value after a year.
4. Secure an adequate budget
The costs of a GenAI project are not just model licences. They also include:
-
System integrations
-
Data preparation
-
System development and maintenance
-
User training
-
Infrastructure (particularly for on-premise solutions)
5. Think about scale
A system that handles 100 queries a day is a completely different proposition from one handling 10,000 queries. Design with your target scale in mind.
SNOK as a partner in GenAI transformation
At SNOK we do not just implement technology - we build partnerships. Our philosophy rests on several foundations:
Deep business understanding - before we write the first line of code, we spend time understanding your business, processes and challenges. Technology is only a tool - it must serve the business, not the other way round.
A methodical approach - we have developed a proven methodology for implementing GenAI projects. From analysis, through PoC, pilot, to full deployment - every stage is carefully planned and executed.
A combination of the best technologies - as a Gold Partner of UiPath, we have access to the best automation tools. We combine them with leading AI models - both global (GPT, Claude) and local (PLLUM) - choosing the optimal solution for each case.
End-to-end support - we do not leave clients with a system and wishes of good luck. We provide full support - from training, through maintenance, to development and optimisation.
Security and compliance - we understand regulatory and security requirements. Our solutions are designed with GDPR, industry requirements and security best practices in mind.
“Our greatest asset is experience,” concludes Grzegorz Surdziel. “We have been through dozens of projects, made mistakes, and drawn lessons from them. Now we can help our clients avoid the pitfalls and reach success faster. This is no longer a time for experiments - it is a time for considered, strategic deployments that deliver real business value.”
Conclusions for the future
The GenAI market is at a turning point. The phase of fascination and experimentation is gradually coming to an end. A period of mature, considered deployments is arriving. Companies that now invest in solid foundations - the right architecture, processes and partners - will be tomorrow’s leaders.
Key conclusions from our experience:
-
GenAI success is not a matter of technology but of architecture - the best model without proper orchestration is just an expensive demo.
-
Context is king - context management is the difference between a system that works and one that impresses.
-
Hybrid wins - combining deterministic workflows with intelligent agents delivers the best results.
-
Simplicity is strength - the simplest solution that works is better than a complicated one that might work.
-
Measurability is fundamental - if you cannot measure success, you cannot manage it.
Looking ahead, we see a world where GenAI will no longer be a separate category of projects, but a natural part of every IT system. Where automation and intelligence will be so intertwined that we will no longer distinguish between them. Where systems will learn and adapt while retaining the predictability and control that business requires.
That world is closer than you think. Companies that are already building GenAI competencies today, that learn from others’ mistakes and invest in proven approaches, will be ready to fully capture the potential of this revolution.
At SNOK we are proud to be part of this transformation. Every project, every challenge, every success of our clients is a step towards a future where technology genuinely serves people and business. Where automation unleashes potential and intelligence strengthens decisions.
As Michał Korzeń puts it: “GenAI is not an end in itself. It is a tool for building a better, more efficient and more human business. Our role is to make that tool work reliably, safely, and deliver real value. That is not easy, but that is exactly why it is worth doing.”
Technological progress is not slowing down. Models are becoming more powerful, tools more sophisticated. But the fundamental challenges remain the same: how do we make technology serve business? How do we turn potential into reality? How do we build systems that not only work, but transform?
The answer lies in combining the best software engineering practices with a deep understanding of AI’s capabilities and limitations. In a partnership between human and machine, where each does what it does best. In an approach that combines innovation with pragmatism, vision with execution.
This is the future we are building at SNOK. A future where GenAI is no longer a buzzword but the foundation of modern business. Where every company, regardless of size or industry, can harness the power of artificial intelligence to create value.
Are you ready for this future? We are. And we are ready to help you get there.
This article was written as part of the “Tech Thursday with SNOK” series, in which we share our experience and knowledge about the latest trends in automation and artificial intelligence. SNOK, as a Gold Partner of UiPath and a leader in the GenAI space, helps enterprises achieve successful digital transformation.
Want to find out more about how GenAI can transform your business? Get in touch with us and discover the possibilities offered by combining UiPath automation with the latest AI models.