Boardrooms Say No to AI Notetakers — For Now

Boardrooms Say No to AI Notetakers — For Now

Legal teams spent the summer sounding alarms, and the warnings are getting louder. Cooley, Mintz, and White & Case all published guidance in 2026 urging companies to keep AI notetakers out of board meetings entirely [1][2][3]. The risks are concrete: a New York federal court already ruled that AI chatbot interactions don't enjoy attorney-client privilege protection, and AI's indiscriminate, overly literal note-taking can generate discoverable records far more detailed — and more dangerous in litigation — than a skilled human notetaker would ever produce.

All-party consent laws in states like California and Massachusetts add another layer of friction, and the chilling effect on candid boardroom debate is a recurring worry among general counsel. The emerging best practice: disable recording tools during privileged, HR, or strategy-sensitive discussions, full stop.

This isn't an anti-AI backlash — it's a call for selective, governed deployment. The tools that win in regulated or high-stakes settings will be the ones that let organizations toggle capture on and off by meeting type, keep data in clearly defined jurisdictions, and separate "transcript" from "legal record." Blanket always-on recording is quickly becoming a liability, not a feature.

The 800ms Problem: Why Voice AI Still Feels Like a Phone Call With Lag

New 2026 benchmarks confirm what anyone using a voice assistant already senses: most cascaded STT+LLM+TTS pipelines run 800ms to 1.6 seconds end-to-end, while natural human conversational gaps sit around 200–500ms [1][2][3]. One widely circulated benchmark found only 2 of 5 popular voice AI stacks stayed under the 300ms "feels seamless" threshold; everything above 800ms starts triggering interruptions or user abandonment.

The fixes are becoming standardized: streaming partial transcripts, speculative token generation, model-based endpointing in the 200–400ms range, and native voice-to-voice models (OpenAI Realtime, LiveKit+Gemini Live) that skip the cascade entirely and hit sub-300ms P95 [2][3]. The emerging engineering budget — roughly 150–300ms for STT partials, 200–400ms for LLM time-to-first-token, 100–200ms for TTS first audio — is becoming the industry's reference architecture.

For anything doing live transcription in meetings, this matters less for live chat-back features and more for real-time speaker attribution and on-the-fly summarization — the latency tolerance there is different, but the underlying streaming and partial-result techniques transfer directly.

Kyndryl's Luxembourg Bet: Agentic AI, But Make It Sovereign

Kyndryl inaugurated its first EU AI Innovation Lab in Bertrange, Luxembourg on October 1–2, targeting agentic AI for finance and government — sectors where data sovereignty isn't optional [1][2][3]. The lab projects 250 jobs by 2030, with founding partner Banque Internationale à Luxembourg already prototyping agentic AI for IT modernization. The framing throughout the launch materials is telling: "Forward-Deployed Engineers," "Human Systems Architects," EU AI Act compliance baked in from day one, human oversight as a design requirement, not an afterthought.

This is a direct signal to the Nordic and broader EU enterprise market: agentic AI adoption is accelerating, but only vendors who can demonstrate regulatory-grade data handling will get through procurement in finance, government, and healthcare.

What This Means For Your Meetings

Today's stories all point to the same tension: organizations want AI deeply embedded in how they capture and retrieve knowledge, but they want it on their terms — local when possible, governed always, and switched off when the conversation is too sensitive to risk. The Obsidian MCP wave shows what people actually want from a knowledge system: not a pile of transcripts, but a graph they can ask questions of, with retrieval that respects context and provenance. That's precisely the gap between "a transcription tool" and "a meeting intelligence system" — and it's why knowledge graphs, not search bars, are becoming the expected baseline.

The legal warnings around boardroom AI and the Kyndryl lab's compliance-first posture are really the same story told from two angles: enterprise buyers, especially in regulated Nordic and EU markets, are done treating "AI notetaker" as a single undifferentiated category. They're asking harder questions — where does the data live, can we turn it off per-meeting, does this survive an EU AI Act audit, does it create discovery risk we didn't have before. Vendors who can answer cleanly will win finance, legal, and government accounts; those who can't will get quietly uninvited from the boardroom.

And the latency research is a reminder that the "AI in the room" experience is still being engineered in real time — every millisecond between someone speaking and a system responding (or staying appropriately silent) shapes whether people trust the tool enough to let it listen to the meetings that actually matter.

Key takeaway: The future of meeting intelligence isn't "record everything and ask an AI later" — it's selective, governed, local-first capture feeding a real knowledge graph, with clear controls for when the AI should simply not be in the room.

Sources

1. https://glama.ai/mcp/best/obsidian
2. https://github.com/sweir1/obsidian-brain
3. https://raw.githubusercontent.com/eugeniughelbur/obsidian-second-brain/main/integrations/obsidian-mcp-server/README.md
4. https://www.cooleygo.com/ai-note-taking-in-the-boardroom-dont-do-it/
5. https://www.mintz.com/insights-center/viewpoints/2226/2026-09-22-hidden-risks-ai-note-taking
6. https://www.whitecase.com/insight-alert/ai-boardroom-privilege-and-recording-decisions
7. https://kenimoto.dev/blog/five-voice-ai-stacks-only-two-under-300ms/
8. https://www.twig.so/blog/voice-ai-agents-latency-budget-800ms
9. https://www.netguru.com/blog/voice-ai-low-latency-techniques
10. https://www.kyndryl.com/in/en/about-us/news/2026/10/innovation-lab-launches-in-luxembourg
11. https://www.luxtimes.lu/businessandfinance/kyndryl-opens-first-eu-ai-lab-in-luxembourg-in-move-which-could-create-250-jobs/162489929.html
12. https://en.paperjam.lu/article/kyndryl-opens-its-first-ai-laboratory-in-the-eu-in-bertrange