An AI assistant today is no longer just a bot that replies with rigid, pre-programmed text after a user picks one of the options in a chat window. Conversations with intelligent AI assistants are now expected to be available across any medium, feel as natural as talking to a human, and at the same time remain safe, predictable, and aligned with business goals.
Creating modern AI assistants that meet these expectations is possible with KODA Intelligence – a module of our platform that not only recognizes user intent but also adapts responses to specific needs and context.
In this article, we explain how multi-level intent recognition works in KODA Intelligence: from rules that ensure full control, through machine learning that detects recurring patterns, all the way to generative AI embedded in safe frameworks for business. We also show how each of these levels is best suited for different types of tasks and for answering a wide range of user questions.
Multi-level intent recognition
Level 1: Rule-based model
At this stage, KODA Intelligence checks whether a message meets defined conditions and attempts to detect intent accordingly.
An intent will be detected if the message:
- Contains at least one predefined phrase;
- Contains a word starting with that phrase (e.g., the phrase “baggage” will detect “baggages” or “baggage claim”);
- Matches the phrase exactly;
- Contains required entities (with optional entities also configurable).
Entities are specific elements in text, such as names, dates, or numbers, that the system extracts as useful values for fulfilling intent (e.g., “tomorrow,” “check-in”).
Additionally, the rule-based model within the KODA platform includes lemmatization (reducing words to their base form), automatic typo correction, and multilingual support with separate sets of phrases for each language. Thanks to this, the system can process messages that aren’t perfectly formatted or grammatically correct, or include foreign expressions. With proper setup, the assistant can also understand questions in different languages and respond appropriately.
When the rule-based model works best
Rule-based systems excel in simple, predictable dialogues where the scope of intents and entities is limited and well-defined. They are ideal for handling frequently asked questions (FAQs) or queries related to complaints and returns. At this level, the assistant works in a precise, fully controlled way, with no risk of hallucinations.
But not all questions or intents can be anticipated. If the rule-based model is not sufficient to answer, the message can be smoothly passed on to another level.
Level 2: Machine learning module
Unlike the rule-based approach, the machine learning module can recognize intent even in previously undefined statements, based on what it has learned from training examples.
Training involves adding sample user inputs (training phrases) for each intent in a specific language. The entire configuration and training process is done within the KODA panel, giving clients convenient access to their training dataset and the ability to edit it in one place.
From the same platform, our team – or the client’s – can also configure NLU (Natural Language Understanding) test cases: sets of sample user inputs with the expected intents. This feature makes it possible to verify whether the model is correctly understanding natural language.
When machine learning works best
Machine learning algorithms are most effective when input data contains hidden patterns that are difficult to capture with rules or generative language models.
In internal process automation, they help predict and prevent problems in advance, support analysis and optimization, and in customer service, they enable personalized responses based on prior interactions and deliver high-level recommendation personalization. ML algorithms can also detect anomalies in data (e.g., unusual user behavior), enabling rapid response and crisis prevention.
If intent is not recognized by rules or ML, the message may advance to the next level: generative AI.
Level 3: Generative AI
Generative AI opens vast opportunities for customer service, but careless use can degrade the user experience. Without proper safeguards and configurations, it can also expose systems to risks.
The goal is therefore to adapt the capabilities of large language models (LLMs) to the client’s specific needs – in other words, to harness them for business use: ensuring maximum safety and accuracy in the responses delivered to end users.
The Generative AI module in the KODA platform is a comprehensive solution for testing, evaluating, and monitoring various LLMs that we use to build intelligent AI agents for our clients.
Real business value doesn’t come from simply using LLMs, but from tailoring the capabilities of the right model to a given business type, organization, and process to be automated. With a set of platform tools, our team builds safe frameworks for effective use of LLMs in business process automation.
The system enables automatic response quality assessment, function call management, prompt control, conversation tracking, and knowledge base management.
Soon, we plan to expand the Generative AI module with advanced evaluators, new testing and query-analysis capabilities – allowing us to further strengthen quality control and raise solution security. More on that in upcoming articles.
When generative AI works best
Generative AI is most effective in processes requiring flexibility, creativity, and real-time content creation – where rules are too rigid and machine learning too limited.
It enables responses to novel, previously unseen queries, supports multilingual communication through automatic translation, and in internal automation can generate concise conversation or documentation summaries for teams. It can also act as a co-pilot for employees – suggesting draft replies for consultants to quickly approve or edit.
Crucially, in KODA Intelligence, Generative AI always operates within controlled frameworks – responses are tested, monitored, and anchored in business context to minimize errors and hallucinations.
Flexible intent recognition sequences
The three-level sequence described above is just one example of how KODA Intelligence can work. Steps and order can be configured as needed.
For example, Generative AI can be used first to classify the query type, and then the machine learning module can be triggered to fetch precise data for a detailed response. Below you can see example possibilities for creating sequences from three modules available in the platform:
Everything depends on the client’s needs and the business process to be automated.
Benefits of multi-level intent recognition & response generation
The three-step process of processing queries and recognizing intent takes just milliseconds. The system instantly interprets a client’s message and automatically adapts the response method. Clients don’t need to decide themselves how the assistant should react – the system, configured by our team, does it in line with the conversation goal.
From the user’s perspective, conversations with the AI assistant feel as natural as talking to a human.
From the client’s perspective, the platform ensures fast implementation, easier maintenance, and safe scaling, thanks to:
- Centralized configuration – the system processes messages centrally while communicating via multiple channels (chat, voice, email).
- Testing environment – nothing reaches users before thorough internal testing.
- Response quality control – the panel allows monitoring whether the agent delivers top-level user experience.
- Security standards – enforced by the required configuration within the platform.
AI power, under control
AI is not a universal key that fits every lock. When creating business process automation solutions for our clients, we always seek balance between the latest generative AI advancements and deterministic models. The key is tailoring the solution to the process and the client’s specific business needs.
In this sense, generative AI can be compared to advanced GPS navigation. It’s excellent for guiding you through unknown terrain, but if you just need to walk a few steps through a familiar office, you don’t need a map app – just a couple of simple, proven directions.
Experience at the core, change in mind
Today’s AI assistants are expected to deliver smooth customer service and highly relevant responses in the shortest possible time. The user experience when interacting with a modern chat or voice assistant is at the core of how we build effective business process automation.
At the same time, alongside the perspective of end users, we prioritize the experience of clients who create and develop automation with us. In developing Platform 3.0, we focused on multi-level security, reliability, and readiness for change – both those driven by the rapid evolution of AI technology and by the growth of our clients’ businesses.
At KODA, we create AI assistants that skillfully recognize user intent and generate accurate responses – not just from the input, but also from broader context. KODA Intelligence is just one module of our platform that allows us to efficiently build and optimize automation for business – from customer service and marketing & sales processes to internal operations.