Visual AI models play an important role in autonomous vehicles. Developing and using these models often involves difficult and expensive data operations, so developing robust and scalable AI systems remains a challenge. We provide evidence that treating images as physical measurements (raw photon counts) improves the reliability of AI models. Access to the raw photon counts also enables the generation of accurate synthetic data that can be used for training, testing and tolerances determination. Size is often cited as a disadvantage of raw data, but we show that high levels of compression can be achieved with low latency and no artefacts. The compressed data is subject to clear quality assurances. Good coordination between industry players, from sensor manufacturers to AI scientists, is key to scalable, reliable and sustainable products.