How often do you run into limitations with AI models in your domain?

Models seem to be able to do anything these days, but I am curious if you see any cases where it doesn't work and can't seem to match up or hit a wall.

The benchmarks show a pretty picture, but the real-world performance in areas such as healthcare, legal analysis, finance, and scientific research has not been objectively quantified.

So, I am curious about which domain your work is in, and within your workflows, where does the current AI fall short?