All five AI assistants gave good answers. Only one gave them fast.
We tested Zoom AI, ChatGPT, Microsoft Copilot, Claude, and Google Gemini across ten real knowledge-work scenarios: meeting summaries, action-item extraction, grounded Q&A, document summarization, project status, and cost modelling. Accuracy scores landed between 4.5 and 4.7 out of 5. Total time to insight ranged from 100 seconds to 365 seconds.
The finding
Answer quality is no longer the differentiator. Every leading assistant is built on a highly capable model, and across our test set the gap in accuracy was 0.2 points.
The gap that mattered was getting the model to the right context.
Where a tool had direct access to the meeting transcript, the document, or the email thread, it produced a usable answer in seconds. Where it did not, the user spent that time finding, retrieving, and pasting source material before the model did any work at all. That overhead is invisible in model benchmarks but unavoidable in real workflows.
10 real-world tests | 5 AI assistants | Up to ~5× faster time to insight.
What's inside
-
Per-test-case accuracy and completeness scores for all five tools, with evaluator notes on where each one fell short
-
Full time-to-insight data, including how retrieval overhead was measured and why
-
A connector reality check: which assistant can actually reach Teams transcripts, Zoom chats, SharePoint, Google Drive, and Outlook today, and which cannot
-
Cost comparison across licensing tiers, including the risk of duplicative spend
-
A complete prompt library. All ten test prompts, verbatim, so you can run this benchmark in your own environment
On methodology and independence
This paper was commissioned by Zoom. enableUC retained full editorial control over the analysis and conclusions.
We publish the methodology, the scoring rubric, the per-test-case results, and every prompt used so you do not have to take our word for it. Scoring was blind: panel participants rated outputs without knowing which tool produced them.
Results depend on licensing tier, tenant configuration, permissions, and connector availability, all of which change frequently. If you are using this to inform a procurement or standardization decision, validate the scenarios that matter most in your own environment. The prompt library exists for exactly that purpose.
About the author
Kevin Kieller is a globally recognized Unified Communications, Collaboration, AI, and technology thought-leader, strategist, and implementation leader. He has been honored as the top UC Today UC All-Star.
Kevin is part analyst and part consultant, which ensures he understand both the “big picture” and the real-world realities. He helps organizations plan for and effectively implement new technologies. He works with vendors to provide competitive intelligence, product strategy, and to help refine and amplify key messages to generate demand.
He has led the development of many technology strategies for medium and large organizations and managed the deployment of hundreds of thousands of collaboration seats.
A long time ago Kevin created an award-winning game for the Commodore 64 and ever since has been focused on using technology creatively to deliver effective value