Most RFID projects that disappoint were decided at the frequency stage, before anyone bought a tag. The three bands behave so differently that picking the wrong one cannot be fixed later by buying a better reader.
The short version
LF at 125 kHz reads a few centimetres, works near metal and water, and is slow. It is what old access cards and animal microchips use. Choose it when you have legacy readers to match.
HF and NFC at 13.56 MHz read up to about ten centimetres, one tag at a time, and every modern phone can read them. Choose it when a person is deliberately tapping something.
UHF at 865–867 MHz in India reads several metres and hundreds of tags per second. Choose it when you want to count things without handling them.
The question that actually decides it
Ask this: does a person deliberately present the item, or does the item pass by?
If a person taps a card on a reader, HF or NFC is right. The short range is a feature — it means only the card being tapped is read, not the one in the next person's pocket.
If items pass through a doorway, sit in a tray, or fill a carton, UHF is the only band that works. A HF reader cannot read a hundred tags in a trolley, no matter how good it is.
Where people get it wrong
The common mistake is choosing UHF for an access door because "longer range is better". It is not. A UHF reader at a door will read the card of the person behind you in the queue, and now your attendance record is wrong in a way that is hard to detect.
The opposite mistake is choosing HF for stock counting because the cards are cheap and familiar. Counting a shelf one tag at a time with an HF reader is slower than counting by hand.
Metal and liquid
Both detune a tag. UHF is affected most, HF less, LF least. This is why an on-metal tag exists — it has a ferrite layer that lets the tag work on a surface that would otherwise kill it. If your items are metal, liquid-filled, or people, you need a tag designed for it, and you need to test it.
What to do next
Before ordering any quantity, get samples of two or three candidate tags and read them on your actual product, in the actual orientation, at the actual distance. A read-rate trial takes a day and settles arguments that would otherwise last months.