Webinar "KUKAs Erfahrung mit LLMs: Das Pilotprojekt auf Basis von KUKA Xpert"

Business summary (LLM pilot at KUKA on top of KUKA Xpert / empolis Service Express)

Company context (KUKA / KUKA Xpert)

• KUKA is a global automation company (HQ: Augsburg), operating across markets such as:
  • automotive
  • electronics
  • consumer electronics
  • healthcare
  • e-commerce
  • logistics

Scale indicators mentioned:

• 100 locations, 50 countries, 15,000 employees
• >8,000 internal employees use Expert
• 63,000 customers access Expert
• Up to 29 languages
• ~20 editors update content daily
• ~2,000 new users/month (reported “for years”)
• >1,000,000 information objects (about 1.5M total)
• >10,000 service cases

Knowledge platform

• “KUKA Expert” (branded as Empolis Service Express)
• Digitized customer documentation for >5 years
  • No printed papers: customers use a QR code to reach the exact documentation.
• Role-based content
  • Internal-only vs external-customer-accessible content
• Content types include:
  • operating instructions
  • assembly instructions
  • spare parts information
  • release notes
  • solutions/symptoms
  • training material
  • examples/videos
  • (and other knowledge artifacts)

Motivation for the LLM chatbot

• Even with access to Expert, many customers still contact the hotline.
• Key driver/metric: about 20% of hotline calls/inquiries are already resolvable using information in KUKA Expert.
• Why users don’t self-serve in the platform:
  • Users want chat/interactive Q&A instead of navigation, filtering, and search.
  • The platform can be complex for new users.
  • Users often need answers immediately.
• Additional internal demand:
  • Employees from multiple departments requested a generative chatbot for Expert.

---

What was built (product + operations)

Initial pilot scope

• A chatbot integrated with Empolis Service Express and the existing KUKA Expert knowledge base.
• Goal:
  • Answer customer questions in natural language
  • while grounding responses in existing documents

Architecture / operating model (framework)

• Implemented as a RAG (Retrieval-Augmented Generation) approach:
  • Semantic and symbolic search plus filtering
  • Ontology/knowledge-structure awareness
  • Iterative retrieval to build the right context
  • Prompt templates to generate answers from retrieved sources
• No LLM fine-tuning required
  • Emphasis on using a capable multilingual LLM and relying on the knowledge base for correctness.

---

LLM evaluation principles & governance (playbook-like points)

Source-grounded answers

• Generated summaries are linked to the most relevant documents so users can verify.
• Rationale:
  • Answers may not be 100% certain, so provenance is essential.

Editorial / safety constraints

• Expert content is editor-reviewed and expert-checked.
• Special care for machine guidelines and safety instructions:
  • Target users include time-critical hotline staff (not only domain experts).
  • Source inclusion is mandatory.

Role-based responses

• Outputs adapt based on user roles and access rights:
  • customers vs internal employees

---

Concrete use cases shown (examples from the system)

Simple factual query

• Example: “What is our latest version of WorkVis?”
• Outcome:
  • Direct natural-language answer
  • Includes document references users can open

Guided clarification when input is ambiguous

• Example: user wants robot article info but provides a category (e.g., “LRB easy”)
• System behavior:
  • Asks follow-up questions to identify the exact model
  • because there may be no single correct answer initially
• Benefit:
  • faster than manual navigation/filtering in Expert

Part-number to documentation lookup

• Example: user provides an article number (e.g., “GMotion rent … from Axis 1”)
• Outcome:
  • returns the correct answer and relevant documents
  • sometimes avoids extra follow-ups

Step-by-step procedure assistance (task workflow)

• Example: “How can you change the gripper of a robot?”
• System approach:
  • asks which robot model is involved
  • then compiles/condenses the relevant procedure info
• Notably:
  • aims to provide tool/material/time planning guidance in chat
  • still points to documents for full step-by-step details (and assets like images/videos elsewhere)

---

Technical implementation highlights (what Empolis added/optimized)

Pilot implementation foundation

• Built on the Empolis Service Express knowledge base.
• Developed an intelligent question answering chatbot that iteratively refines:
  • semantic search
  • ontology/knowledge-model handling
  • incremental context building for the LLM

Agent behavior

• Uses an agent approach consistent with “Perceive–Reason–Act”:
  • decides whether more information is needed
  • triggers additional searches
  • asks follow-up questions until it has enough detail
  • can escalate to a human/tool conceptually if necessary

Empolis-specific strengths leveraged

• Search methods (semantic + symbolic)
• Preprocessing:
  • chunking documents
  • retrieving the right context for large content
• Handling large documents:
  • KUKA docs can be up to ~500 pages
  • sourced from multiple systems, e.g.:
    • Noxum editorial
    • knowledge models from SharePoint/Salesforce

LLM choice

• Tested SystemClow 2 on AWS Bedrock (with “Tropic Cloud 2” referenced)
• Selected for:
  • good response times
  • multilingual performance

---

Metrics / KPIs mentioned (and what they imply)

Operational effectiveness KPI

• ~20% of hotline calls/inquiries can be resolved using Expert documentation
  • used as baseline justification for chatbot impact

Knowledge/scale KPIs (context for difficulty)

• >10,000 service cases
• 1M+ information objects (about 1.5M total)
• ~20 editors updating daily
• ~8,000 internal users, 63,000 customer users
• 29 languages
• Documents up to ~500 pages

No explicit business targets (e.g., cost reduction, deflection rate, containment/CAC/LTV) were stated in the subtitles.

---

Challenges found + next steps (actionable)

Remaining challenges

• Coverage challenge:
  • initial tests didn’t fully cover the entire knowledge-base concept
  • later improved coverage to increase accuracy
• Governance challenge:
  • ensuring every response aligns with:
    • editorial/expert validation
    • machine guidelines
    • safety instructions
    • source citation
• Integration challenge:
  • connecting the chatbot to broader service processes (e.g., hotline workflows and internal service activities)

Next steps

• Expand requirements and integrate:
  • the AI chatbot into additional service processes
  • operational workflows (e.g., helping hotline users quickly draft responses using chat output)
• Continue close collaboration to refine pilots and rollout.

---

Promotional / commercial motion (high level)

• Empolis offered a two-day Generative AI Evaluation Workshop to:
  • evaluate readiness
  • identify high-value pilot starting points
  • frame the approach as “divide the elephant into smaller pieces” (start with targeted pilots to reach best ROI/path)

---

Presenters / sources

• Ramona Stammler (webinar host)
• Letizia Baumann (implied as “Ms. Baumann”, Product Owner Knowledge Management at KUKA)
• Erik Prender (Chief Product Officer at empolis)