Multimodal AI in Production Struggles With Messy Real-World Inputs and Compliance

Demos of multimodal AI typically use clean, high-quality images, while production systems must handle scanned invoices, rotated pages, and low-resolution files. The article argues that the vision model is often the simplest component, with ingestion, pre-processing, routing, and human review posing the greater challenges. It notes that real deployments frequently need OCR fallbacks and that end-to-end latency can rise substantially once pre-processing and validation are included. Audio and video inputs add consent, retention, and redaction obligations that text pipelines rarely face. The author recommends building evaluation sets from actual production failures and designing deletion and fallback paths into the system from the start.