From my experience in construction, the works are usually divied up into packages or 'claim cards'. Then each supervisors are supposed to report how complete each package under their control is.
Problem is, this can sometimes be a little tricky and the supervisors often fudge the numbers to smooth out how their works look. Ie, they may have done 50% of the work, but will only claim 20% because they know that the last 50% worth of the claim card will actually take way longer. They don't want to look bad at the end of each week cause the estimators did a shit job. So they'll take their gains and average them out.
A system that was more automatic and had way more compliance built in would be cool to see. I doubt however, that an algorithm could solve what is predominately an issue of people and estimation.
If 50% takes longer than another 50%, the metric by which estimation is performed is, at least regarding time, flawed.
If a subcontractor were to report, "hey boss, we've gotten 8/10 items done this week, but the remaining two are going to take two weeks because $REASONS. The estimates appear to be flawed, perhaps because $X and $Y.", I suspect the entire project's management would benefit.
I suspect the opposite. The managers will start to panic and then micromanage everything, have crisis meetings etc. This will slow it down even further.
The only real solution to this problem is better long term estimates! The thing is that all estimates are partly BS because we haven't figured out how to predict the future yet. So the solution above is that the supervisor got a BS estimate and they are pushing back with BS progress. There probably is no other way.
Good managers will take the opportunity to learn more. Either there is a way to aid those who made the estimates, a way to help the people doing the work to get the job done, a need to relax the schedule, or some combination of them.
The goal is to get the work done on time and with the resources at hand. Micromanagement is often a hindrance to the goal, so good managers avoid it.
You've just identified why a system that's more automatic is risky to naively trust.
Here's my prediction: the AI is going to see a lot of jobs that progress linearly to completion and then stop dead at 90%. That having been said, it'll still be useful if metrics can be sampled out to guesstimate how much padding is needed to account for the 90% dead-stop time.
Problem is, this can sometimes be a little tricky and the supervisors often fudge the numbers to smooth out how their works look. Ie, they may have done 50% of the work, but will only claim 20% because they know that the last 50% worth of the claim card will actually take way longer. They don't want to look bad at the end of each week cause the estimators did a shit job. So they'll take their gains and average them out.
A system that was more automatic and had way more compliance built in would be cool to see. I doubt however, that an algorithm could solve what is predominately an issue of people and estimation.