I'm not sure that you can influence them so much. After all, % of non-standard incoming data is estimated. Analytics is based on quantitative analysis. If any analyzing system (including AI) would redraw all the output data by obtaining 1 non-standard package (for example), then this would not work as a whole. It simply puts it off as a pattern with a certain degree of probability of its occurrence in the future.
I'm not saying that you are wrong, this is just my guess :)
As far as I know AI works differently. There are many types of networks, for many specific tasks - image recognition, text predictions (translations), generational adversarial networks, etc.
Very simplified: it takes a whole bunch of input, it sets a desired outcome, then it tries to minimize a cost function, by finding the best next match that will minimize this function. It does this by training the model. This training involves multiple passes over specific neuron layers, back-propagation, gradient descent and quite a few other interesting calculus operations applied. At the end of this, though, there is still a certain threshold that is applied to the results of the cost function. Meaning it's still numbers that you can be on top or below. These numbers are extremely low, 0.00000000000001, but it's still a number that can be affected by the input.
The term AI is misleading, I use it mostly because it's popular, the more appropriate one would be Machine Learning, it's closer to what's actually happening: a machine which learns possible paths, based on parameters.