On your first question, the pattern I trust is to let the task decide rather than the model. If the output can be checked by code, such as a classification, an extraction against a schema or a field update, a small model is usually good enough, and the check tells you when it is not. Anything that needs judgement about a customer goes up a tier.
"Fallback must be boring" is the line I agree with most. Normalizing outputs across tiers into one schema, and validating before anything downstream sees them, is what keeps formats stable when the free tier drops out mid conversation.
On trust inside a CRM, I would start with proposals rather than writes: the agent drafts the change, a person approves it, and every applied change is logged with the source that justified it. "Tool schemas are the product" fits that well, since narrow tools are easier to approve. How has Mato v1.0 compared with the free tier on your own tool calls?