Detecting LLM hallucinations in financial reporting
A hallucination in financial reporting is a figure or a causal claim with no basis in source data. Detection matters, but prevention matters more: the cheapest control is an architecture in which the model never produces the number.
Two failure modes must be separated. A hallucination is invented content. Incomplete retrieval is missing content presented as complete. They need different tests.
Octopus AI removes the first class of failure by construction and tests for the second continuously.
Questions and answers
How do you detect a hallucinated figure?
Recompute it. If arithmetic runs as deterministic code, any figure that cannot be reproduced from reconciled source data is rejected before display. Figures without complete line-item lineage are treated as failures, not as low-confidence answers.
How is a hallucination distinguished from incomplete retrieval?
A hallucination has no supporting source record. Incomplete retrieval has supporting records that were not retrieved. Coverage checks against the reconciled layer identify the second case, while lineage checks identify the first.
Can prompting alone prevent hallucinated numbers?
No. Prompting reduces frequency but provides no guarantee, and a generated number can differ between two identical questions. Removing arithmetic from the model is a structural control rather than a mitigation.
What about hallucinated explanations rather than hallucinated numbers?
Causal claims are handled by asking the accountable person instead of inferring them. Variance commentary is collected from the responsible budget owner and logged as evidence, so the explanation has a named author.
What is the last control before an output reaches a person outside finance?
An approval gate. Every outbound message is signed off by a named human, and the approval is logged with the content that was approved.
Proof
- Reference client: $523K total contract value over 3 years.
- Approximately 430 analyst hours saved.
- Forecast cycle cut from about 60 hours to under 6 hours (90% faster).
- Error rate below 0.1%.
- 2,500+ autonomous tasks completed in month one.