Enterprise AI assistants improve trust with evidence verification
EvidenT: Building Trustworthy Enterprise Assistants through Evidence Groundedness and Traceability
Artificial Intelligence
Summary
Enterprise AI assistants often struggle to provide answers that clearly show where their information comes from, which can cause mistakes or untrustworthy responses. The authors introduce EvidenT, a system that double-checks the evidence used before giving an answer, making sure it matches real documents without needing to retrain the AI. This helps reduce errors like citing wrong sources and makes it easier to track exactly where answers come from. They tested EvidenT on about 500 real work-related questions and found it made answers much more reliable by improving how often the right sources were cited.
What this means in practice
- •For enterprise ai developers: Improve AI assistant answers by verifying and aligning evidence extracted from mixed company data sources before response generation.
- •For enterprise data compliance teams: Ensure AI-generated responses cite only verified and traceable sources to meet audit and information integrity requirements.
Authors
Anubha Kabra, Katie Jooyoung Kim, Colin Zhiwei Kou, Helene Sajer, Yimei Fan, Radomir Cisar, Heather Greenhalgh, Gabriel Martinez Vidiri
Abstract
Enterprise AI assistants must produce responses that are verifiable and traceable to source evidence. However, retrieval augmented generation (RAG) over heterogeneous enterprise data can suffer from citation drift, unsupported content, and weak source traceability. We present EvidenT (T = Trust + Transparency + Traceability), a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, without model retraining. EvidenT combines structured passage extraction with deterministic lexical alignment to filter unsupported content, correct citation drift, and preserve source-span traceability. On approximately 500 real enterprise queries, EvidenT improves gold-source hit rate by an average of 29% over prompting baselines, produces no citations to nonretrieved urls, and achieves near-saturated answer-to-source lexical coverage.