LinkedIn·Monday, 10 August 2026·16d ago

Successful deployment of industrial AI pilots is kind of like painting your house. You should spend 80% of the time on preparation. In May,…

Kasper De Smaele

Co-founder & CEO | Kasqade

Successful deployment of industrial AI pilots is kind of like painting your house. You should spend 80% of the time on preparation. In May, I spoke at the Embedded Vision Summit in Silicon Valley about why these vision AI pilots fail, and why it's rarely due to model performance. My experiences deploying vision AI across EU and North America have taught me that: 1. Before anything: measure a baseline and quantify the improvement you aim to see. Time saved, output increased, waste reduced... I've seen factories wanting to reduce defects, but they couldn't tell me how many of each defect occurred per batch. 2. Understand who interacts with, and who is affected by this new technology: a fantastic in-line visual quality control model fails when operators are not involved and start sabotaging the lighting and trigger. Yes, this happens in real life - I witnessed it. 3. Time to value is key: once 80% of the prep is done and you actually run a pilot, everyone should start seeing value immediately. Production teams are too busy and overloaded to wait for a few weeks or months to see results. If you can't show at least one point of improvement/impact by the end of week 1, your chances of failure just skyrocketed. Thanks Jeff Bier, Phil Lapsley, and Edge AI and Vision Alliance for the opportunity to speak in front of such an expert audience, and for creating a platform where engineers, researchers, and business partners can share thoughts on where machine vision is headed.