Generative Defect Simulation for SAP Visual Inspection
Executive Summary
In 2024, as part of SAP Business AI, I collaborated with the Visual Inspection product group to design a generative AI workflow capable of producing synthetic defect images for machine learning training. This replaced the traditional and costly process of manufacturing and photographing defective parts. The concept advanced through leadership reviews, secured support, and later became a shipped capability inside SAP’s manufacturing portfolio, enabling customers to accelerate visual inspection model training with significantly lower cost and effort.
Introduction
Visual inspection systems depend on high-quality training data to accurately detect defects in manufactured products. Traditionally, this required physically creating defective samples, photographing them, and manually annotating each defect. This process was slow, expensive, and difficult to scale across product variants.
In 2024, SAP Business AI explored a new approach using generative AI to produce synthetic defect images based on a single clean photograph of a product component. This breakthrough allowed manufacturers to train machine learning models with far greater efficiency while reducing reliance on physical defects.
I partnered with the Visual Inspection team to design the end-to-end experience that operationalised this idea and made it usable inside SAP’s ecosystem.
Context and Strategic Position
SAP Business AI functioned as a central strategic team responsible for shaping AI experience guidelines and enabling consistent adoption across SAP’s product landscape. My role involved providing UX leadership and consultation to multiple lines of business to ensure AI concepts were both strategically aligned and operationally feasible.
Visual inspection was a particularly strong candidate for innovation because:
ML adoption in manufacturing is slowed primarily by dataset creation
Manufacturers repeatedly spend time and money generating physical defects
Rare defects are difficult to capture, leading to gaps in model accuracy
Each new product variant required starting ML training from zero
Synthetic data could reduce time and cost while increasing accuracy
This project demonstrated how SAP Business AI could unlock value at the core of digital manufacturing.
Problem
Machine learning systems require thousands of high-quality images representing both defective and non-defective product states. Manufacturers faced several difficulties:
Operational difficulties
Creating defective samples required stopping or adjusting production
Gathering rare defect types often took weeks
Manual annotation required skilled personnel and significant time
Model retraining needed full repetition for every new product variant
Business pain points
Dataset creation costs were high
Onboarding new products slowed down adoption
Inconsistent real-world image quality caused ML setbacks
Scaling across product lines required significant recurring investment
The challenge was to design a radically more efficient process that reduced cost, improved quality, and enabled reuse across variants.
My Role
As the UX designer representing SAP Business AI, I was responsible for:
Understanding manufacturing workflows, defect taxonomies, and ML training processes
Collaborating closely with the intelligent systems specialist and Visual Inspection product owner
Designing the generative AI workflow for producing synthetic defect images
Enabling users to upload product photos, generate defects, refine them, and prepare datasets
Ensuring clarity, transparency, and quality controls for domain experts
Creating the UX that connected this workflow to Visual Inspection’s ML training pipeline
Shaping how the concept aligned with SAP’s broader AI design guidelines
Preparing and presenting the concept for leadership evaluation
What began as a conceptual exploration later moved into SAP’s roadmap and evolved into a fully shipped capability.
Solution Overview
The designed workflow uses generative AI to create realistic defect images by building synthetic flaws on top of a base product photo. Users can generate multiple variations for each defect type and customise them to match real production scenarios.
The experience allows users to:
Upload a photo of any product component
Select a defect category
Generate synthetic versions with realistic visual quality
Adjust severity, scale, location, and frequency
Compare original versus generated variations
Approve or regenerate results
Export the entire synthetic dataset directly into Visual Inspection
Use the dataset to train machine learning models immediately
This streamlined dataset creation process eliminates repetitive manual work, accelerates ML adoption, and reduces cost for manufacturers.
Why This Matters for SAP
This work advanced SAP’s strategic goals in multiple ways.
Strengthened SAP’s manufacturing portfolio
Synthetic defect generation significantly reduces the barrier for customers adopting ML-driven visual inspection.
Supports cross-industry scalability
The generative workflow adapts to any product category because users provide their own images.
Reinforces SAP Business AI as an innovation leader
Demonstrates practical, valuable applications of generative AI within core enterprise workflows.
Drives ecosystem consistency
The design aligns with SAP’s AI guidelines, promoting uniform adoption of intelligent capabilities across teams.
Becomes a reusable capability
The logic for synthetic defect generation can support multiple SAP manufacturing products.
Most importantly, the concept matured into a real product capability, confirming its strategic and practical value.
Estimated Business Impact
(Industry-based estimations)
Time savings
Reduction in dataset preparation time: 70 to 90 percent
Reduction in total ML model development cycle: 50 to 70 percent
Cost reduction
Reduction in physical defect-creation cost: 80 to 95 percent
Estimated 20,000 to 200,000 EUR saved per product line per training cycle
Annotation-related cost reduction: 60 to 85 percent
Annual savings for medium manufacturers: 50,000 to 150,000 EUR
Quality improvements
Accuracy improvement: 10 to 25 percent due to broader defect coverage
Greater consistency in lighting, angles, and defect visibility
Scalability
New product onboarding time reduced by 50 to 80 percent
Reusable model logic increases ROI across multiple product lines
Value for SAP
Supports AI-first manufacturing strategy
Improves competitiveness in the industrial AI market
Provides a demonstrable example of Business AI enabling Lines of Business
These estimations show how generative AI can significantly shift efficiency, capability, and cost structures for manufacturers.
Closing
This project reimagined how manufacturers prepare machine learning datasets for visual inspection. By introducing synthetic defect generation, the workflow reduces time, lowers cost, and delivers higher consistency without disrupting production lines.
The most rewarding outcome is that the concept progressed far beyond exploration. It received strong leadership support, evolved through validation cycles, and is now a shipped capability within SAP’s manufacturing product portfolio. Customers are already using this to accelerate defect detection training and improve product quality with far less effort.
