I think it was John Gall who pointed out that all sufficiently complex systems operate in failure mode 100% of the time, meaning that some of the controls have been bypassed, or are being operated manually.
Because the automatic controls operate invisibly, the only visible work is that of humans overriding the manual controls, which will always seem chaotic and inefficient. In fact this could already be the status quo today in many organizations such as bureaucracies.
If this remains the case, then the AI-orchestrated workplace of tomorrow will look like the highly computer orchestrated workplace of today, with the computers operating invisibly in the background, and humans manually correcting the errors. To overcome this stage in our development as a civilization, the controls don't just have to be automatic, but error-proof.
When I was a young data engineer in 2014 at a very large telco in Canada the real time resource allocation of network bandwidth would work flawlessly from 6am to 11pm, but would have a random panic every day at 11.25 pm, where the cluster would enter a degraded mode until just after 6am the next day.
This was very weird, but I trace the issue to a bunch of NFS drives which started being used after 11pm. We figured some job somewhere needed to run then and there was nothing to do about it but migrate the hardware. After a couple of months, a lot of planning and tens of millions we moved the data from those drives to some shiny new beefy servers and the problems went away.
Years later I worked with someone else from that telco. Turns out he was in a team that did nothing but monitor the allocation of jobs on the compute cluster in real time and decide which machine each job ended up on. Their work hours were from 6am to 11pm.