As long as the participants are randomly selected, all of those factors drop out into noise. The law of large numbers guarantees you that as sample sizes go to infinity, the magnitude in difference between the i.i.d. variables in the samples goes to zero.
* We don't have infinity, or sufficiently large numbers of companies to do this
* We don't have the resources to test sufficiently large numbers of companies like this
* The dependent variable in question is going to be entirely dependent on the number of people infected at a company at the beginning of the study, which is unknowable unless you can test everyone... which we can't.
* Social habits of the company
* Number of infected people within the company
* Compliance with the protocol
* Age demographics
* Infection rates from places other than the company
* Weird selection bias resulting from the company needing to not lock down to begin with
Very, very difficult study