Documents and ledgers need different tools
Retrieval over documents works well for policies, procedures and contracts: find the passage, cite it, answer. Business data is different. "What do we still owe this vendor for this month?" is not a passage in a document. It is a calculation over invoices, payments and credit notes that changes every day.
An assistant that answers that question from a search index will sooner or later answer from an old copy. It has to ask the system of record instead.
Ask the system, through typed calls
We give agents a small set of typed operations against the ERP or CRM, such as look up a balance, draft a purchase order or flag an exception, rather than open access. Each operation has a defined input and output, so the agent cannot invent a field or a status that does not exist.
Validate before anything touches the ledger
Every write is checked against the schema and the business rules before it is accepted: the vendor exists, the amounts add up, the account is valid, the approval limit is respected. Anything that fails goes to a person, not into the books.
Low-confidence cases follow the same path. A person approves them until shadow-mode results show the threshold can move.
Log every action
Every call an agent makes, and every record it changes, is logged with the input that led to it. That log is what lets finance trust the automation, and what lets an auditor follow it.