A company updates its records, but its AI agent may still rely on outdated information. New research shows that even advanced models struggle to interpret conflicting records and determine which information was accurate at a particular time.
Researchers from KAIST and Korea University introduced MMEnterprise-Bench on October 20, 2026, to test how AI models interpret changing workplace records. The findings expose weaknesses in resolving conflicting information and tracking revisions across documents, messages, and images.
As enterprises give AI agents more responsibility, outdated information could affect decisions involving project deadlines, financial reports, and customer records. Correctly retrieving a document does not guarantee an accurate answer.
Where AI agents struggle with changing records
The MMEnterprise-Bench study evaluated 10 multimodal AI models using 1,613 questions across 30 simulated workplace projects. Its 7,140 records included documents, messages, and visual materials.
Researchers tested information extraction, conflict resolution, and whether models could recognize questions lacking sufficient evidence.
With 10 retrieved records, the highest-performing model, GPT-5.6 Sol, achieved 52.6% overall accuracy. It scored 36.3% on conflicts between sources and 32.5% on questions requiring historical information.
Increasing retrieval to 50 records raised overall accuracy to 68%. However, performance gains varied across models, showing that additional evidence did not consistently produce better answers.
Visual information proved particularly difficult. GPT-5.6 Sol achieved 83.3% accuracy on questions requiring one text record, compared with 27.5% when multiple images were needed.
As enterprises move from AI copilots toward autonomous workflows, these weaknesses could become more consequential. Agents acting on outdated information may introduce errors into processes that previously required employee verification.
Enterprise AI needs better record validation
Organizations connecting agents to document repositories and business applications need to test more than retrieval accuracy. Systems must distinguish revised information from earlier versions and recognize when records contradict one another.
Dataiku's proposed cross-platform AI agent tracking reflects growing demand for visibility across deployments. Monitoring agent activity, however, does not verify the accuracy of the information those agents use.
Enterprise architects should test contradictory documents, historical timestamps, and mixed-format evidence before allowing agents to support consequential decisions. Existing AI governance and traceability concerns make those checks particularly relevant as organizations expand automation.
What eWeek Found: More evidence helps, but historical accuracy still lags
MMEnterprise-Bench exposes a measurable weakness in reasoning about changing records. With 10 retrieved records, GPT-5.6 Sol achieved 67.5% accuracy on information extraction but only 32.5% on temporal conflicts—a 35-percentage-point gap.
Expanding retrieval to 50 records increased overall accuracy by 15.4 percentage points, reaching 68%. Temporal-conflict accuracy improved to 56.1%, leaving a substantial share of questions answered incorrectly.
The results extend earlier enterprise-agent evaluations, including Argo-Bench, which examined business task completion rather than reasoning across revised information. The benchmarks assess different capabilities and cannot be compared directly by score.
MMEnterprise-Bench is a research preprint built around simulated workplace records. It evaluates question answering rather than live business transactions, so its results do not establish production failure rates.
For deployments involving financial, operational, or compliance records, validation should include historical accuracy and conflicting evidence—not simply whether an agent retrieved relevant documents.
Want to learn more AI tips, tricks, and prompting techniques? Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work.
Learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. eWeek readers get free 7-day access. Check out all the lessons here →