The AI layer
The AI in the platform runs on a layer that is separate from the worksheets. Which model does which task is a setting for the platform administrator, not a development job, and a new model only goes into use after it has first run alongside the current one.
A model list, in administration
The models the platform can choose from are kept in a list in administration, not in the code. For each model it is fixed what it can handle and what it costs, also per prompt length, and from which limit «thinking» can be switched off. A choice in administration cannot make an AI button fail on a setting the model does not support.
Model and thinking depth per AI button
The platform administrator can set a different model, a different thinking depth, a custom length or a time budget per type of task and per AI button. As long as nothing is set, every button works as before. The length allows for the thinking, so an answer is not cut off mid-sentence because the reasoning used up the space.
A fallback when a model refuses
If the chosen model refuses a request before any answer has been given, or is overloaded, the button falls back once to another model that is active: the button's previous model, otherwise the default model for that type of task. A model that has been phased out is never chosen as a fallback. The user simply gets an answer instead of an error.
Two routes, with the region as a setting
The AI can run via two routes. A platform administrator can set the AI to run in the EU only. Both settings are off until they are switched on.
Trialling a new model first: the trial bench
In the AI models administration screen there is a Trial bench tab. You choose the current model and a candidate, run the same fixed set with both and compare the results side by side: how many correct, how many cut off, the time and the cost. You also read a sample of ten texts side by side yourself and give your verdict on each text. If something cannot be measured, it says «not measured» and nothing starts. A comparison that could not measure anything turns red and says at the top why, instead of green with a table full of errors. The trial bench runs with your own login.
A model change as a process
A model change follows a fixed procedure: a new model first enters the list as a trial model and is compared with the current one on the trial bench; only then does it go into use. Every night a check is made that the model list still matches what the provider offers.
What this means for a user
Nothing changes unexpectedly. The consultant only notices a model change through better or faster proposals; the rules in AI support continue to apply: a proposal is a proposal, with its origin, until someone accepts it.