AI will be one form of "Insane" or another after 900 years of isolation
Given our current understanding of machine learning, it is a two part process. First we give it a set of learning data, then we give it qualifiers about how to process that learning data to create a meaningful result: here-in lies the problem. An AI can only learn as well as Humans can teach it. Yes, the ability of an AI to think faster and try more combinations and receive more inputs than a human does give AI certain superhuman learning powers that allow them to invent, consider, and remember things that we humans are incapable of doing ourselves, but when you isolate AI from human input, they start to learn "noise". If you show an AI a bridge, and say, "this is a bridge", then show it a rock and say "this is a rock", it can learn to discriminate between the two, but if you start to flood it with inputs without ever giving it any qualifiers for how to discriminate the inputs, one of 3 things will become true depending on the nature of the AI:
It will start randomly associating concepts with inputs as part of the Jitter in its learning matrix. Without any reinforcement, all things will become all things. One moment it will look at a Rock and call it a bridge, the next moment it will look at it and think it is bread... basically it will become incapable of applying consistent meaning to its environment leading to something similar to a totally debilitating form of schizophrenia.
A natural bias in its learning algorithm will lead it to become completely focused on a singular idea. If it has a bias towards thinking about rocks, then over time, bridges will become rocks, bread will become rocks, every thought and feeling it has will drift towards the idea of "rock" and after 900 years, your AI will just be stuck standing in some random corner saying "rock, rock, rock..." without another thought in its head. In this case, your AI falls into a sort of totally debilitating form of OCD.
Its designers gave it a learning tolerance to prevent the previous 2 outcomes. Basically this means that as the AI realizes it's not learning new information; so, it becomes more resistant to learning, and simply reinforces what it already believes is true. If this is the case, then after 900 years, it will be much more functional than the earlier two cases, but it will be like your ultra conservative grandpa x10. It will be unwilling to experiment, unwilling to learn from new inputs... so, when it does come into contact with humans again, it must either:
A. Remember how to make a deep space ship from 900 years ago, because it already knew everything it would need to know.
B. It will not know how to make a ship, and never be able to figure it out.
This narrows down your options for the nature of the AI
Considering option #3.A is probably the best case for your story, your AI will not need to preform any RnD... nor will it be able to. It will have to know every step in exact detail about what humans have to do. Going from stone age to medieval tech only takes about 1-3 years as can be observed from various YouTubers who've done just that... after this, things will slow down a lot. As you get closer to modern day tech, the scale and complexity at which you need to build becomes bigger and bigger. And limited human populations will become more and more of a hinderance. Making things like a modern microchip without a large scale civilization to support the massive investment cost is a really big problem.
This means that your exact human demographics will become important. If by "Tribal" you mean like how the Native American or Central Africans were before colonialism, then you will already have a the population you need, and it will only take a couple of decades to have them build up all the factories and infrastructure they will need to turn thier large-scale stone-aged civilizations into a modern ones.
But, if by tribal you mean hunter-gatherer societies, then you have a much longer road ahead of you. You AI will need to wait many generations for human settlements to simply grow enough to support the needed infrastructure to form a space program, but once the human population returns to a few 100 million with major urban centers that can be interconnected by a global trade network, it will be doable.