AI-Powered System Documentation for a Cloud ERP
How Deepico AG built an AI-powered system documentation for its Cloud ERP in the food industry within just a few weeks – from chatbot experiments to the successful deployment of Notebook LM.
English edition — originally published in German as KI-gestützte Systemdokumentation eines Cloud-ERP.
This blog post is a continuation of the specialist article from the book "AI in Project Management" by the Swiss Project Management Association (spm).
Specialist Article: Extract from the Abstract
The specialist article shows in practical terms how Deepico AG built an AI-powered system documentation for its proprietary Cloud ERP for the food industry within just a few weeks – with a focus on improvements in support and training.
Value (Goals) of System Documentation
- Efficient answering of support queries about system usage
- Sustainable training of customers and new employees
- Resource-efficient documentation of the dynamic evolution of the system
First Approach: The Chatbot in the ERP
Our first idea was obvious: consolidate existing but fragmented information sources and make them accessible via a chatbot directly in our Cloud ERP application. Users should be able to simply ask questions during their daily work and receive an immediate answer.
What we learned: The approach was promising but hit three limits:
Speed versus Quality
With a powerful "Deep Thinking" language model, response times were 15 to 20 seconds. That is long – scientific findings confirm that such waiting times noticeably interrupt the workflow. At the same time, our initial user testing clearly showed: response quality is decisive for whether a chatbot gets used at all. A fast but inaccurate answer helps nobody. Reducing the response time to perhaps ten seconds would be technically possible. However, the effort would not be proportionate to the benefit.
No Room for Customer-Specific Configurations
Our Cloud ERP serves various branches of the food industry – from bakery and confectionery to fruit and vegetables to meat processing. The different processes have led over time to various module versions and base setups. The chatbot could not reflect these differences: it did not know which customer works with which modules and versions. Solving this cleanly would have meant manually indexing all data sources and linking them to the respective customers – a considerable additional effort.
Source Code as a Difficult Data Basis
The original idea of using the source code as a foundation and translating it into understandable language via generative AI turned out to be more complex than expected. For the article master data alone, around 40 code files are needed. On top of that come linked files with reusable components. The language model can read all of this, but the context becomes so large that response accuracy noticeably suffers. Furthermore, the source code contains much information that is irrelevant for using the application. It would simply be difficult to instruct the model to filter out only what users actually need in their daily work. An efficient, automated comparison during further development was hardly feasible and would have led to considerable ongoing maintenance effort.
New Approach: Notebook LM as an Internal Knowledge Source
As an "AI first" company, we continuously monitor which new technologies truly move us forward. In doing so, we came across Notebook LM – and were quickly convinced.
The strengths are obvious:
- Versatile information access: In addition to the chat interface, content can also be prepared as video, podcast, mind map or flashcards. The flashcards in particular are an exciting feature for targeted knowledge training.
- Easy for everyone: Even employees without a technical background can use Notebook LM without difficulty. Information sources can be uploaded, updated or removed as in a simple web interface.
- Transparent sources: Notebook LM references the underlying source with every answer. If something is no longer accurate, the source can be directly adjusted and saved again – without detours.
- Multimedia data basis: Notebook LM processes not only documents but also reliably interprets videos and web links.
Today, the Deepico team works with Notebook LM. We have stored all system and module descriptions, instructions, instructional videos and the most important update notes as sources. Whether for customer onboarding, training or support inquiries: we use Notebook LM to create tailored instructions – in writing or as video, depending on the customer's preference. Currently, we use the tool internally, carefully review the outputs and then enrich the instructions with customer-specific screenshots and data.
In-App Documentation and Intuitive User Guidance
In parallel, we are investing specifically in documentation directly within the system. Functions are increasingly explained via context menus. Even more importantly: we place great emphasis on users being guided through the application itself – with a well-thought-out user interface that shows where functions are and how they work. Good UX is the best documentation.
AI as a Driver – Internally and Externally
This project exemplifies how we work at Deepico: we continuously use AI ourselves, gather real experiences and let this knowledge flow directly into our customer projects. With foresight, courage and the willingness to learn from individual steps, we help companies not just use AI, but create real value with it.
Author: Claudia Eigenmann, Project Manager at Deepico AG (a venture of Deep Impact AG)