When a screen failed to take today's content, someone restarted it by hand. It worked every time.

That's why nobody found the problem.

At Creative Realities, the platform pushed content out to a network of about 500 screens. Some of them didn't take the push. Someone would go restart the ones that failed, and the updated content would land.

Nobody was being careless. Each failure looked like its own small incident, the restart took seconds, and it always worked. We were exiting beta testing and heading toward go-live and scale, where everything is urgent and a fix measured in seconds is a good trade.

But that labor scaled with the screen count. Nobody had written it down anywhere, so nobody had noticed it wasn't a cost — it was a ceiling. There's no number of people that makes manual recovery work across thousands of screens.

Years earlier, in a graduate lab, I'd worked on a video system with a similar constraint. The refresh couldn't read the entire memory block in a single cycle. Rather than trying to make it read faster, we toggled — first half on one cycle, second half on the next, and back again, faster than anyone could perceive. We had the clearest display in the class because we designed the conflict out instead of fighting it.

The push problem had the same shape. Every screen was reaching for content at the top of the hour. Many couldn't reach the server, or reached it and didn't get the content.

I recognized the pattern, proposed a fix, and directed the team that implemented it. We broke the push into blocks distributed across the hour. Same content, same screens, same hour — arriving in sequence rather than all at once.

The part that mattered came next. New screens get assigned into different groups as they're added, so the staggering holds as the network grows. A fix you have to redo at every new scale isn't a fix. It's maintenance you've agreed to forever.

That network went from about 500 screens to more than 12,000.

The stagger didn't create that growth. It removed the thing that made it impossible.

What I keep taking from it: a workaround that works is one of the most expensive things in an operation, because it stops the clock on diagnosis. The tell is labor. If the standing answer to a recurring failure is a person doing something, the cost is hidden inside someone's day, and it grows with your volume.

In a smaller company, nobody owns that question by title. That doesn't make it someone else's — and the cost is always calculable, once someone decides to look.

The expensive part was never the fix. It was every restart before it.