Capabilities

Menu

EN

Menu

EN

Smart Form AI – when forms can suddenly think

Smart Form AI – when forms can suddenly think

Robert Krämer, Technical Director, denkwerk
Robert Krämer, Technical Director, denkwerk

Robert Krämer

Robert Krämer

Technical Director

Technical Director

denkwerk

denkwerk

In an age where service speed and user-friendliness determine customer satisfaction and frustration, innovative solutions are in demand. This is exactly where our newly developed form assistant comes in—a smart helper that not only provides forms on websites, but also actively assists users in finding the right form and filling it out correctly step by step.

in a TechTalk at TechDay, the two denkwerk developers Sascha Zander and Dogan Teke presented their chatbot-based assistant.

The challenge: forms wherever you look

We are all familiar with the issue: corporate websites feature numerous forms, often with similar designs, for various purposes—from contact requests to complaints to job applications. Customers frequently fill out the incorrect form, information is sent to the wrong service team, or important fields are left blank. The result: time-consuming inquiries and frustration on both sides. Our vision: to use AI to make the form process so intuitive that users can achieve their goals in a targeted and error-free manner – while making internal processes more efficient and reliable for companies.

“We wanted to build a chatbot that guides users through the entire form process—from selection to submission. Forms should not confuse users, but rather guide them in a targeted manner and help them avoid incorrect entries.”

Dogan Teke, Software Developer at denkwerk

Intelligent user guidance via chatbot

At the heart of our approach is an AI-powered chatbot that addresses the user directly, asks about the situation or concern at hand, and then guides them through the entire form process. This ensures that only the relevant fields are queried. Incorrect entries or misunderstandings are detected early on and corrected directly in the dialogue.

Architecture & Technology: How our smart assistant works

Our solution is based on a modern, modular architecture (in a Docker setup):


  • Ollama as a language model (LLM): The “brain” of the system understands natural language and helps to correctly identify and assign user requests

  • ChromaDB: A vector database handles semantic searches. Users do not need to know the exact keyword—the AI suggests suitable forms even for imprecise entries

  • Central data storage: Form data and processes are maintained in clear, flexibly expandable JSON files (“single source of truth”)

  • Session management & gateway: A gateway acts as the “heart” of the system, controlling the entire process, storing sessions, and coordinating the individual system components.


A fundamental idea was to divide the process into different phases in which the form reacts differently, i.e., different code is used. The three phases range from determining the correct form to targeted queries and validations of individual fields to submitting the data with final confirmation. The key point is the validation of the data:


“We validate the inputs ourselves—this way, we avoid AI hallucinations.”

Sascha Zander, Senior Software Developer at denkwerk


After a free chat introduction, comparison with the ChromaDB to find a suitable form, and the supposedly correct form recognition, we step in and analyze the entries. To do this, we transfer them to structured fields using JSON. In a field-by-field dialog, we go through all fields until everything is filled in correctly. Validation is the most time-consuming part, but ultimately saves resources and leads to the desired goal.

Aha moments and challenges

As with any (tech) project, challenges arose during development, from which we learned valuable lessons. Here is an overview of our big and small “aha” moments:


  • AI models are not always reliable, e.g., in the case of unusual names or language mixes

  • A balance must be found between preprocessing (e.g., prefiltering with ChromaDB) and the computational effort of the language models

  • “Prompt engineering”: Shorter, targeted inputs to the AI model deliver better results than complex, nested prompts

  • Consistency in models, indexing, and queries is crucial to avoid unexpected behavior

  • Errors in history/prompts can confuse the AI; a clear structure is important.


Finding the right balance between performance and result quality was a particularly exciting challenge for us.
Last but not least, we once again realized how important optimized “prompt engineering” is: short, clear instructions to the language model deliver better results than complex commands.


“Prompt overload is real – the simpler and clearer the prompts, the better the quality of the response.”

Sascha Zander, Senior Software Developer at denkwerk

Conclusion: Versatile applications for Smart Form AI

The use of AI and chatbots solves a key pain point in digital customer service: instead of leaving users alone in a jungle of forms, we guide them directly and clearly to their actual goal. For companies, this means fewer errors, lower support costs, and a better service experience across the board. However, the benefits of our system go far beyond traditional form processes. It can be used anywhere where user interactions need to be automated, structured, and simplified – whether for onboarding, support requests, internal processes, or even system control. 

Share this Spark

Last Sparks