loading...
RISING$500K · 21 Feb 2026 · 6 MIN
The Hard Part of AI Isn't Being Smart – It's Knowing When to Speak
Fern doesn't just transcribe meetings – it monitors the conversation in real time and whispers useful context exactly when you need it.
FERNishikilabs.ai ↗
LATEST ROUND$500K 15.01.2026
TOTAL RAISED$500K · 1 rounds
OPPORTUNITY SNAPSHOTbuild adjacent
ENTRY ANGLES
Contextually aware AI that monitors processes and decides when to intervene without user triggers · AI integrated into enterprise tools to alert employees before they make errors · Browser-based AI that surfaces connections between current content and past notes/information
VERTICALS
Enterprise software/workplace toolsChat and meeting applicationsBrowser and knowledge management
CAPABILITIES
Judgment calibration to determine when AI intervention is appropriate vs. noise, Real-time contextual understanding and process monitoring, Low-level sensorimotor/perceptual skills (per Moravec Paradox reference)
FERN FOUNDER
“In my Facebook and Instagram feeds I don't see my friends anymore,”
01 /The Concept
AI meeting assistants have proliferated to the point of saturation. Most of them are glorified stenographers: they transcribe the meeting, then summarize it so you can quickly review what was discussed.
Fern is also an AI meeting assistant, but its primary purpose is different. It's less a secretary and more a sharp colleague sitting next to you in the meeting – one who can whisper useful things in your ear as the conversation unfolds, and who can look things up or take action in real time.
Fern's first mode is explicit participation. Attendees can address it directly, ask it questions, or give it tasks on the spot – ask for its take on something, have it search the web for supporting information, draft and send an email, assign a task to a teammate, or schedule a follow-up meeting.
Its second mode is silent and invisible. Other attendees don't see Fern as a participant, but it still listens and analyzes everything – and feeds a private stream of commentary into a separate window: contextual background on what people are saying, follow-up questions worth asking, points you might want to make.
In both modes, Fern still does the standard meeting assistant job: it saves the transcript and generates a summary, enriched with its own conclusions and insights.
Fern has a Mac app for in-person meetings and for receiving its silent-mode commentary.
Pricing: 5 meetings per month free; 15 meetings for $15/month; unlimited for $30/month.
Fern was built by Ishiki Labs, currently going through Y Combinator acceleration. The startup published details about the assistant's launch two days ago.
02 /Why It Matters
Ishiki Labs' broader mission is building AI systems capable of "contextually appropriate" responses to incoming audio and video. "Appropriate" means the AI understands when to respond and when to stay quiet – which sounds obvious but turns out to be technically difficult.
Humans do this well, though not all of us. For AI, it's a genuinely hard problem. Today's AI systems are designed to respond when directly prompted – when asked a question or given an explicit task. Left to their own judgment about when to speak up, they largely can't.
Without that capability, AI can't be a true real-world assistant. An AI that watches you cook from a recipe and intervenes only when you start making a mistake requires it. An AI that monitors what goes into your shopping cart and flags something harmful – or notices you forgot an item – requires it. Currently, that kind of proactive, judgment-based assistance means prompting the AI every time.
The same technical challenge was identified by Continua ([related review](/review/a-tut-nuzhny-sovsem-drugie-ii-agenty)), which raised $8M in its first round last August. Continua is developing what it calls "Social AI" – an AI that can participate in group conversations, but only surfaces its suggestions at the right moments: helping participants recall something, offering advice, or surfacing information relevant to a decision – without disrupting the natural flow of conversation.
Continua's specific product is a group chat AI assistant that helps two people or a whole group plan an event, a trip, or any shared activity.
Last autumn, Digipals ([related review](/review/chtoby-vzletet-nuzhno-protiv)) – then going through YC – started building something similar. Its concept: an "AI-native social operating system" featuring specialized widgets that appear in chats exactly when the AI determines that participants could use help. Third-party developers can build those widgets, not just the startup itself.
The catalyst for Digipals was the migration of younger users from social networks into group chats. "In my Facebook and Instagram feeds I don't see my friends anymore," the startup wrote. "A third is ads, a third is content from strangers the algorithm decided I should see, and the last third is AI-generated junk optimized for views and engagement. My actual social life happens offline or in group chats, where it's just my friends."
The most recent example in this space is Alfi ([related review](/review/startap-na-tom-chto-molodjozh-uhodit-iz-socsetej)) – another YC graduate, another group chat AI assistant, built for the same underlying reasons as Digipals with similar functionality. Alfi's launch was announced less than a month ago.
03 /Opportunities
The idea of AI that is "contextually appropriate" – knowing when to speak and when to hold back – is more profound than it might appear. It's arguably the foundation for an entirely new generation of AI products, applicable across a very wide range of contexts.
The startups described today are building this kind of social AI for chats and meetings. But those are clearly just two applications. A contextually aware AI could surface inside enterprise tools – stepping in when an employee is about to make an error. It could sit in a browser and surface a connection between something you're reading now and a note you wrote six months ago and forgot.
The common thread: users wouldn't need to actively trigger these assists by pressing a button or asking a question. The AI would monitor the process and exercise its own judgment about when to intervene – finding the right balance between generating noise and surfacing genuinely important moments.
The catch: calibrating that judgment is trivially easy for a human, who resolves it in milliseconds through instinct and experience. For AI, it's anything but trivial. If you doubt it, try writing out the explicit rules an AI should follow to identify those important moments.
This brings to mind the Moravec Paradox, formulated in 1988. The observation: high-level cognitive tasks (chess, mathematics) are relatively straightforward to automate because they require modest computational resources. But low-level sensorimotor skills – perception, fine motor control, tasks a one-year-old handles effortlessly – require enormous computational resources and remain extremely difficult for AI.
Social skills, it turns out, may belong to the same category.
But "difficult" is not the same as "unsolvable." Social AI will get there. In fact, it's already getting there – as the startups here demonstrate. This is a real trend.
So the interesting direction is building platforms and tools with "contextually appropriate" or "social" AI built in. The remaining question is choosing the specific domain where this approach can deliver fast, obvious value at a quality level that's achievable today. Which domain fits that description for you?
funding raised$8M