What does a model actually know?
A language model knows patterns in language. It does not know that a customer called twice last week, that the roof inspection failed, or that the balance is 45 days overdue. Every one of those facts has to be supplied at the moment of the request.
That supply line is the CRM. Remove it and the AI is improvising.
Why is a confident wrong answer the real risk?
An empty answer is obvious. A plausible answer that references the wrong job, the wrong price or the wrong appointment slips through and reaches the customer.
This is why grounding matters more than model choice: an AI should be required to look information up rather than recall it.
What does good grounding look like?
- The AI retrieves the specific record before answering about it.
- Counts and totals come from a query, never from the model's own estimate.
- Identifiers are exact, so a follow-up refers to the same job.
- When data is missing, the AI says so instead of filling the gap.