Providing the right document does not guarantee that an AI system will use it to answer a question. Seekr, an enterprise AI software company, reported that two large language models sometimes gave incorrect answers even when the correct financial record was available to them.
The finding came from Seekr’s study, “Retrieved Does Not Mean Used: Measuring Evidence Utilization in Grounded Enterprise AI.” Researchers tested Llama 3.3 70B and Qwen 2.5 72B on 80 questions drawn from historical U.S. Treasury financial statements, comparing how accurately the models answered under different conditions.
When given only the correct document, Llama answered 95% of the questions correctly and Qwen 92.5%. When researchers added similar records from other years alongside the correct document, accuracy fell to 85% for Llama and 87.5% for Qwen.
The correct information was still available in both tests. The difference was that the models had to distinguish it from other records containing similar information.
Where the errors came from
Accuracy dropped much further when the models had to rely on a retrieval system to select documents for them. Llama answered 37.1% of its attempted questions correctly, while Qwen managed 30.4%. Those figures excluded questions the models declined to answer. Llama answered 70 of the 80 questions, and Qwen answered 69.
Seekr attributed most of this larger decline to retrieval failures, where the correct document never reached the model. The smaller drop in the controlled test showed the separate problem: errors that occurred even when researchers had already supplied the right evidence.
An earlier 40-question pilot also distinguished retrieval failures from errors made after successful retrieval. Of 26 errors recorded across the process, seven occurred despite the correct document being retrieved.
The financial records help explain how such mistakes happen. The report explains that a question about expenditures in 1903 might bring up Treasury statements from both 1903 and 1913. The tables and headings in those reports can look nearly identical, making it possible for a model to select a figure from the wrong year.
Finding evidence versus using it
To examine which records influenced the answers, Seekr used context attribution, a method that estimates how supplied documents affect a model’s response.
Once the correct record reaches the model, context attribution can help assess whether that record influenced the answer. In an earlier blog post, Seekr argued that retrieval performance and the use of retrieved evidence should be measured separately.
The study was limited to two models and a small set of historical financial questions. Its results cannot establish how frequently similar errors occur in enterprise applications involving contracts, regulatory filings or financial reporting.
For businesses using AI to analyze large document collections, the tests identify a specific limitation in how accuracy is measured. A system can successfully retrieve the correct file without producing an answer based on it. Retrieval performance alone cannot establish whether the model used the right evidence.