Fraud and AML teams are now being offered AI assistants from every direction. Some are grounded in the institution's own data. Some are thin wrappers over a public LLM. The two look similar in a demo and diverge sharply in production. For regulated financial institutions the choice is not aesthetic, it is a question of accuracy, defensibility, and legality.
What grounded AI actually means
A grounded AI assistant retrieves information from a specific, known data source before answering. In the fraud context the source is the bank's transaction records, rule configuration, case history, and uploaded evidence. The model reads those records, reasons over them, and produces an answer that cites them.
The accuracy of a grounded assistant depends on what it can retrieve. If the retrieval is good, the answer is backed by evidence an analyst can inspect. If the question is outside the data, the assistant says so.
What a generic LLM is actually doing
A general-purpose LLM has no connection to the bank's records. It answers from the patterns it saw in pre-training. For a question like 'has this counterparty appeared in prior fraud cases at my bank,' the model has no basis to answer, but it may produce an answer anyway.
This is not a flaw in the model, it is a mismatch between the model and the task. Public LLMs are built for broad language tasks. Fraud investigation is a narrow task with high stakes and specific evidence requirements.
Why tenant boundaries are not optional in financial services
When analysts type investigation questions into a generic model, they are sending customer data to a third party. Depending on the LLM provider's terms, that data may be used for training, logged for abuse monitoring, or stored indefinitely.
For a Canadian federally regulated bank, this is a model risk, a privacy risk, and a vendor risk. OSFI's model risk guideline treats third-party models in scope. Canadian privacy law treats cross-border data transfer as material. The combination means a generic LLM used by analysts is not a tool to be ignored.
How the two look in practice
Example question: 'show me every transaction above ten thousand dollars from this user in the last 30 days that triggered a rule.' A grounded assistant queries the records, returns the actual list, and the analyst can click into any row. A generic LLM invents a plausible-looking table.
Example question: 'what should I check next on this alert.' A grounded assistant looks at the alert, the user, the rule that fired, and historical patterns, and gives specific next steps tied to the data. A generic LLM gives a textbook answer that may or may not match the bank's workflow.
What this means operationally
If an institution wants an AI assistant for fraud work, it should be grounded in the institution's own data, run inside the institution's own tenant, and respect the institution's permission model. A browser tab open to a public chatbot is not a substitute, it is a different and incompatible thing.
UMCA Aurora's AI Navigator was built against these constraints. Every answer is backed by Aurora's own records. The entire conversation stays inside the tenant. The access model is the same as the rest of the platform.
