Evolving workplace records are now a measured failure point for enterprise agents. In MMEnterprise-Bench, released on 20 October 2026, the strongest model tested scored 83.3% on questions that needed a single text record and 32.5% on temporal conflicts, where the same fact had been revised over time. Correct retrieval did not produce a correct answer.

The benchmark, from researchers at KAIST and Korea University, holds 1,613 questions across 30 simulated collaboration projects. Its retrieval corpus holds 7,140 records, 5,725 textual and 1,415 visual, spread across Slack messages, Notion entries, email, meeting notes, PDF pages, slides and standalone images, with 287 record versions preserving how the workspace changed. The authors evaluate 10 proprietary and open-source multimodal models on three tasks: evidence reasoning, conflict resolution, and recognising when evidence is insufficient to answer. Source: https://arxiv.org/abs/2610.27911v1 and the project page https://mlvlab.github.io/MMEnterprise-Bench/

What evolving workplace records actually test

A cross-source conflict is two sources stating different versions of the same fact at the same time. A temporal conflict is one source stating different versions at different times, with the earlier version still in the corpus. Both are routine in a real company. A deadline moves in email, stays old in a slide deck, and is corrected again in a meeting note.

On cross-source conflicts the top model reached 36.3%. On temporal conflicts it reached 32.5% with 10 retrieved records. Across all 10 models, conflict resolution ranged from 11.8% to 36.3% on cross-source questions and from 13.2% to 33.5% on temporal questions, according to the paper. Source: https://arxiv.org/abs/2610.27911v1 ; secondary reporting with the same top-line figures: https://www.eweek.com/news/ai-agents-changing-workplace-records/ and https://aiweekly.co/alerts/gpt-56-sol-tops-mmenterprise-bench-at-68-multi-image-at-275

The visual gap is as large. The same top model fell to 27.5% when a question needed multiple images and to 25.6% on mixed text and image records. On unanswerability, recognising that the evidence does not support an answer, Gemini 3.8 Flash led at 87.4%, ahead of GPT-5.6 Sol at 65.5%. A model that cannot say the record is insufficient will answer from a superseded version instead.

More retrieval helps recall, not judgement on evolving workplace records

With 10 retrieved records, GPT-5.6 Sol scored 52.6% overall. With 50 records it scored 68.0%. Temporal accuracy rose in that larger budget, reported by eWeek at 56.1%, but a large share of temporal questions still failed, and the paper states the gain is not consistent across models.

"even the strongest models struggle when evidence contains cross-source or temporal conflicts or spans multiple images" and "increasing the retrieval budget or employing a larger retriever improves evidence recall but does not consistently translate into higher answer accuracy."

That distinction matters for buying decisions. A larger index, a larger retriever, and a larger context window raise recall. They do not give a record a version history, a source, an expiry date, or a rule for which statement supersedes another. Without that structure, the agent receives more versions of the same fact with no reliable way to rank them by time or authority.

The limit of the study should be stated plainly. MMEnterprise-Bench is a preprint built on simulated projects grounded in 12 research papers and 18 government publications. It tests question answering, not live transactions, so its scores do not establish production failure rates. Source for that caveat: https://www.eweek.com/news/ai-agents-changing-workplace-records/

In production, the wrong version becomes the wrong record

The production cost of that weakness is already reported separately. In The State of Agentic AI: H2 2026 from Nylas, a survey of 1,030 United States professionals who build, buy, or influence AI systems, fielded 26 August to 12 September 2026, 54% said an agent had done something wrong at their organisation in the past 90 days. The most common failure, at 31%, was triggering a workflow the team had to reverse. A quarter, 25%, had an agent update the wrong record. Only 53% could produce a full audit log of what their agents did in the past week. Sources: https://www.nylas.com/blog/the-state-of-agentic-ai-h2-2026/ and https://www.businesswire.com/news/home/20261020854203/en/54-of-Companies-Report-an-AI-Agent-Error-in-the-Past-90-Days-Nylas-Survey-Finds

The incident rate rose with scale, from 45% among organisations running one to five agents to 70% among those running more than 20. Nylas chief executive Jeff Koets put the shift in one line, linked to the Business Wire release above: "The category has moved from ‘will this work’ to ‘what do we do when it doesn’t.’"

An agent that updates the wrong record is often not hallucinating. It is acting on a version that was once true. The failure sits in the record it was given, not only in the model that read it.

That is why Jilu treats the company record as a record first and a retrieval corpus second. A fact that joins the shared record needs its source, its date, the statement it replaces, and a review step before people and other agents rely on it, so a correction propagates instead of competing with the version it corrected. Retrieval can then work over a record that already knows which version was true, and when. The benchmark does not settle which product design is right. It settles a narrower point, with numbers: if the underlying record cannot tell time, a stronger model will still give the old answer, confidently, at scale.