The AI on Your Factory Floor Is Getting Worse, and Almost Nobody Notices

Most people don't know this, but all those AI systems running on physical devices — sensors, cameras, machines on a factory floor — the day they start is usually the best day they'll ever be. After that, they start slipping. A camera contends with changing seasons. A machine starts vibrating differently year after year. People use the system in ways slightly different from the group it was trained on. None of that appears on a spec sheet, and none of it is the model's fault — but the model isn't adjusting for it on its own. It just quietly gets a bit worse.

With the backing of a company called Rhizome Labs, which published new research in late June on a fix for exactly this problem, and a significant shift in EU AI regulation this summer, the calculation for anyone building this kind of technology has changed.

Estimates of the edge AI market range from $30 billion to $48 billion in 2026, depending on whose research you look at, growing at 22–29% per year. Take any single figure with a grain of salt — different analysts use different criteria — but the direction is consistent. The reason for the growth: implementing AI directly on devices rather than in the cloud delivers three things at once — privacy (data stays on the device), speed (no round-trip to a remote server), and increasingly, economics (smaller, specialized models often make more sense than large cloud-based counterparts).

The EU AI Act is shifting in ways worth watching, particularly for European companies entering the American market. Requirements that would have obliged large corporations to comply with stricter rules for high-risk AI systems from August 2, 2026 were amended shortly after adoption. As the EU Council announced in June, the transition period for large firms to document compliance was extended to December 2027 — and to August 2028 for AI embedded in regulated products such as medical devices or automotive systems. Transparency requirements, including the obligation to notify a person when they're interacting with an AI system, did come into force this month, but alongside those other changes.

Once you've deployed an AI model onto hardware in the field and it starts to go stale, your options aren't good. You can haul the data it's accumulating back to a central server, retrain the model, and push the updated version back down — but that defeats the purpose: the data the device was meant to keep safe now travels somewhere you didn't want it to go. Or you can let the model degrade until it can be swapped out, which isn't a great option either. Researchers call what happens when a model forgets old information as you try to update it with new information "catastrophic forgetting." Neither path is acceptable when the stakes involve medical or industrial infrastructure where a performance lapse or security breach has real consequences. It's not a niche problem — the volume of research specifically on on-device model updating has grown markedly in the last two years.

Rhizome Labs's pitch is straightforward once you strip away the technical language: build AI that can be trusted to keep working in messy, real-world conditions, and that a human can still understand and check up on. Their newest published research, from the end of June, describes a way for a model to quickly adapt to new conditions using just a single new data point — rather than a full, expensive retraining process. In plain terms: instead of shipping data back to headquarters or accepting gradual degradation, the model adjusts itself on the spot. They've also published research on predicting medical-style outcomes from limited data, and had a separate paper accepted at a major computer vision conference this year. Different projects, same thread: models that keep working, and remain explainable, long after active supervision ends.

One of our other portfolio companies, Almetra, is tackling a related version of this problem from a different angle: cameras and AI running directly on factory floors, watching assembly work in real time. Manufacturers including ABB and Viessmann have reported solid productivity gains from it. Different problem, same underlying conviction — that putting the intelligence on the device, right where the data is being generated, is the better long-term bet, not just a workaround until better internet connections arrive.

The easy question in edge AI has mostly been answered: yes, it is possible to shrink useful AI models down onto small, cheap devices. The hard question — the one few people are seriously asking — is whether those models are still any good six or twelve months later, after they've stopped being carefully monitored. That's the challenge we want to invest in solving, and it's why we're interested in seeing systems built using our open testbeds.

If you're working on continuously learning models that retain their usefulness long after initial deployment, on devices with limited connectivity to central infrastructure — we'd like to hear from you.