An LLM hallucination is when a large language model produces a statement that is fluent and confident but factually wrong or unsupported, inventing a number, a fact, or a source that does not exist. It happens because the model generates the most plausible-sounding text, not the verified truth, and plausible is not the same as correct.
In finance, hallucination is the disqualifying risk. A model that confidently states a wrong figure, an invented invoice total, or a non-existent contract clause is worse than useless: the output looks authoritative. Any AI touching financial data has to be built so it cannot simply make things up.
The defense is design, not hope. An AI that verifies against source documents and exposes its reasoning cannot hallucinate a result the way a free-generating chatbot can, because every output must point back to real data.
This is how Phacet is built. Its agents do not generate figures from general knowledge; they check the company's documents and report what they find. The agent that controls supplier billing compares invoices to real prices, the three-way matching agent ties each one to its order and delivery, and the agent that verifies invoices against contract terms checks against the actual contract. Every conclusion is backed by a native audit trail showing the source.
Hallucination is the risk of AI that asserts without grounding. Phacet counters it by verifying against source data and exposing the reasoning, so every output can be checked, not just trusted.
