1) The whole point of AI is that it is far less controlled than traditional programming. That's a big part where it's power comes from. There are no guarantees AI will not suddenly decide to call black people "monkeys" in full view of a US audience. But any program that comes with guarantees won't identify anything correctly.
2) This means that management and government's screams for guarantees are impossible. Any honest salesman would of course point this out.
One might point out that humans have the same problem: there's no guarantee a human will follow guidelines. Guidelines from management and governments invariably are impossible to properly satisfy. There's always "tension" between different rules and never any real guideline on how to resolve that tension, oh and there's always "unwritten rules" (what companies/governments do but are ashamed of and so don't write down). And then there's the real world, which imposes it's rules with brutal uncaring efficiency and doesn't care about any decision by anyone. For example, there is never any guidance on what to do if the real world intervenes and a truck breaks down.
P.S. it isn't even really possible to "dial" guarantees vs capability, neither in humans, nor in AI models (aside from choosing a model, and even that's a lie, because choosing a model is based on a benchmark result, not on real world realiability. If benchmark and real world differ enough, ...)
Of course, enterprise sales can respond to this conundrum in TWO ways:
a) honestly identify the problem and help people think about their projects realistically
b) charge more for "AI governance"
It's selling pretty well.
And even the government building in tension is not going to work, because once these models become smart enough to outsmart not just most humans but even groups of humans trying to trick it, the government will of course find it's laws completely ineffective and bypassed for the slightest of reasons.