Language models can give a wrong answer in a confident tone. In an enterprise the real risk isn't the wrong answer but that nobody notices it.
01. Hallucination
A model can invent something it doesn't know. The most effective control is grounding the answer in a document and showing the source. An answer that cannot cite a source shouldn't be accepted for enterprise use.
02. Bias
Tendencies in training data can surface in output. In HR, credit and customer assessment that's direct legal exposure. Consistency of output across groups should be measured regularly.
03. Silent Drift
When a provider updates a model, behaviour can change without notice. Running a fixed evaluation set regularly is the only practical way to catch drift early.
04. Overconfidence
After a while users stop checking output. Making uncertainty visible in the interface and requiring a check step in critical fields balances that habit.
05. Prompt Manipulation
Text arriving from outside can carry content that behaves like an instruction to the model. Systems processing user documents or web content must be tested for this.
06. Acceptance Criteria
Define a numeric threshold for an AI feature to reach production. Releases made without defining "good enough" are hard to walk back when something goes wrong.