Time bars sample the market at a constant rate regardless of how much is happening, which means a sleepy lunchtime hour gets the same weight as the minute after a rate decision. Activity-based bars sample more finely when information is arriving.
The practical benefit is statistical: returns measured on dollar bars are closer to independent and much closer to normally distributed than time-bar returns, which makes downstream statistics better behaved. Bar counts per day then vary with activity, which is the point.
The cost is operational. Bars no longer align across instruments or with the clock, which complicates portfolio construction, and the threshold N is another parameter to justify. Use them where microstructure matters and stick with time bars where it does not.
Related: ohlcv-bar, resampling, tick-data, feature-engineering