Not just tokens
IMHO it's not just about the tokens. It's whether tokens are "wasted" or not.
It's about how good an outcome you get - i.e. how good the "AI" is at providing a solution..
If you give an "AI" a coding problem to solve and it manages it, then great. It was X tokens but did the job (you can tehn analyse if those X tokens were a saving compared to developer time & money)
However, issue starts when (as is often the case) the "AI" does not do a good job of providing a solution - iit may be totally useless, it may have some good stuff amongst the dross, but your X tokens have been wasted to some extent and you either now need to solve it by a human, or try "AI" again , burning more tokens.
I have used "AI" coding tools & a definite trade off - the more granular, well defined a task you gibe, then the higher the chance of "AI" doing a good job first time, but even then it often needs some fine tuning afterwards. However easy to stray into level of input being so well defined you may as well have just written the code yourself as you are essentially going to "code level" in describing the requirement.
Occasionally you get surprised by good results with a "loose" input query, but such instances are rare.
Bosses would be far happier if they only had to pay for tokens if result of those tokens was actually fit for purpose.
e.g. If I took car to a garage to have a new tyre fitted and tracking adjusted if needed & mechanic wasn't paying attention and they checked the tracking but added new spark plug instead of a tyre then, I would not be paying for the non required spark plug & the mechanic would be going back to install the new tyre & recheck the tracking.
Until "AI" billing is similar (i.e. you do not pay if it makes a mess of things) then there are (IMHO) issues.
I have seen improvements in "AI" output, but it is still a long way off getting it right every time on coding tasks.