Speech recognition is wrong often enough that any app trusting it blindly will eventually put the wrong number into somebody's accounts. So Buhata doesn't trust it. Everything you say stops at a confirmation before a single record exists.
The ladder a sentence climbs
Your words go through several stages, and most sentences never reach the last one.
- The grammar on your own device, which handles ordinary commands with no network at all
- That same grammar again, after repairing a word against things you've written before
- A shared pool of phrasings, built from patterns lots of people use
- A cache of sentences you personally have said before
- Only then, the AI model
Nine commands in ten get answered before the model is ever consulted. That's what keeps the feature affordable, and it's why the simple things happen instantly.
Then it asks
Wherever the answer came from, you land on the same screen. Every action it understood, listed out, each with a tick you can clear and a pencil you can press.
The pencil opens the app's real editor with the entry already filled in. Not a preview of the form. The actual form, with every field and every check it normally has.
What it does when it's unsure
If a line is missing something it needs, a date say, it tells you instead of guessing. If it heard something it can't do at all, it leaves the line out rather than forcing it into the nearest thing it can manage.
Seven hours of sleep isn't seven glasses of water, and an app that quietly files one as the other is worse than one that admits it didn't understand.
And if it missed something
There's a line at the bottom to add whatever it didn't catch. Speaking the whole thing again to fix one missed item would be a bad trade for your time.