Notes from the AGI Horizon Tokyo Roundtable
April 8, 2026 · Tokyo, Japan
On April 8th, the Silicon Valley Legal Tech Frontier Community co-hosted a panel at AGI Horizon Tokyo, an AI summit that brought together 500–1,000 AI builders in Tokyo. The event was organized by WayToAGI (the largest Chinese-language open-source AI knowledge community) and LinkLoud (a global AI founders' community), with our community as a co-hosting partner.
The panel — "Vertical AI: Where the Money Follows" — featured three builders working across legal, finance, and voice AI, moderated by Galen from LinkLoud. Our community's Asia-Pacific lead Koki represented the legal AI perspective alongside Yua, co-founder of Fonda AI (finance), and Shucho (William), co-founder of Reco AI (voice agents, Japan).
The framing was sharp: 2023–2024 was general AI's moment. 2025–2026 belongs to builders who have picked a vertical and started building. Here's what we took away.
The Legal AI Timeline Is Moving Faster Than Anyone Expected
Koki outlined a timeline that resonated across the room:
- Before 2023: "Does legal AI actually work?"
- 2024: "Does it work well enough for daily legal tasks?"
- 2025–2026: "Who bears the liability?"
The shift in the central question tells the story — AI has moved from the periphery into the core of legal services. The conversation is no longer about feasibility. It's about accountability.
Koki broke down the landscape into two segments. Law firms — more than 50% now use some form of legal AI, but adoption depth varies enormously. Data security and client confidentiality remain the biggest friction points; many firms use AI tools but avoid going deep. In-house legal departments, especially in Japan, face a different challenge: their data is scattered across ERP systems, knowledge management platforms, and contract management tools, making integration hard for any external vendor.
Two Moats That Big AI Can't Easily Replicate
When the moderator asked how vertical startups defend against foundation model companies adding legal plugins, Koki identified two structural moats:
Data. The public internet is the tip of the iceberg. The most valuable legal data lives inside court systems, law firms, and enterprise legal departments — and even when court data is technically accessible (as in the U.S.), turning raw data into AI-usable, labeled datasets requires legal domain experts, which is expensive and slow. Law firms treat their internal databases as core assets and won't share them.
Liability. General AI companies won't register as law firms or take legal responsibility for their outputs. But a new category — AI-native law firms — is emerging. Koki cited the example of Crosby (a Sequoia-backed startup) whose CEO has publicly committed to bearing legal liability for the firm's contract review outputs. The model: eight specialized AI agents handle the review, human lawyers make the final call, and the firm stands behind the result. That combination of low-cost AI delivery plus human-backed liability is something no foundation model company will match.
The Rise of the "Vibe-Coding Lawyer"
One of the most forward-looking observations from the panel: lawyers are becoming builders. Koki highlighted a Hong Kong-based lawyer in our community who published a contract review demo on LinkedIn that was viewed over 1,000 times and sparked a wave of imitation worldwide.
The implication is significant. In the previous generation of legal tech, lawyers waited for startups like Harvey or Legora to build products and deliver them. Now, lawyers themselves are developing demos, validating their own needs, and distributing solutions — in their own language, for their own workflows.
Koki's takeaway: the old B2B sales model may be outdated. The new distribution channel runs through the lawyers themselves. If you build tools that let lawyers create what they need — vertical code editors, legal-specific co-pilots, low-code platforms — you unlock a very high-value market.
This connects to our community's core vision: empowering every legal professional to become an AI-native lawyer. Maybe one day, the term "legal AI" disappears entirely — because every lawyer is an AI-native lawyer.
Across Verticals: Shared Patterns, Different Playbooks
The cross-vertical conversation surfaced several parallel themes:
Voice AI in Japan (Reco AI): William shared that Japanese call centers are struggling with labor shortages — AI isn't replacing workers, it's filling a gap that already exists. Reco AI built its own speech-to-text and text-to-speech models in Japanese because none of the major providers (Google, Microsoft, OpenAI) performed well enough. Their moat: 40–50% of what human operators do isn't written in any manual or FAQ — it's tacit knowledge buried in operators' minds and enterprise workflows. Extracting and encoding that knowledge is Reco AI's real defensibility. They're also building multi-agent systems with supervisor agents monitoring handling agents in real time to prevent hallucination.
Finance AI (Fonda AI): Yua explained why Fonda chose institutional investors over retail: professional analysts can absorb deeper research, and B2B pricing lets them capture value rather than competing on $10/month subscriptions. Their team is half ex-hedge fund professionals — they are their own target users. On moats, Yua pushed back gently on the "vibe-coding" trend: in finance, if an analyst can rebuild your product in a day, you have no moat. Fonda's edge is delivering insights users don't already have — for example, predicting which vertical Anthropic or OpenAI will enter next by cross-referencing hiring patterns, CEO podcasts, and market signals.
A shared concern — hallucination and trust: Both William and Koki emphasized that in high-stakes verticals, the cost of getting it wrong is existential. For call centers, a hallucinating agent damages client reputation. For legal, an incorrect contract review creates liability. Both are building human-in-the-loop safeguards — not as a temporary crutch, but as a structural feature.
AI Won't Replace You — But It Will Change Your Job Description
On the future of work, all three panelists converged on a similar view:
- William: Human operators will shift from taking calls to becoming prompt engineers and AI supervisors — monitoring, improving, and handling the emotionally complex conversations AI can't.
- Koki: Multi-agent systems are making AI-native law firms possible in 2025–2026. The shift is from passive copilot (you ask, it answers) to proactive agent (you upload a contract, it delivers a risk report without being asked). Some jobs will be automated, but the opportunity is to move into more creative, autonomous roles. Koki also highlighted an important counter-trend: an "anti-distill" movement pushing back against companies that force employees to encode their expertise into AI systems, raising questions about worker autonomy and labor rights in the AI era.
- Yua: Finance has always been a high-turnover market with or without AI. Fonda's positioning is to make each analyst cover more ground — more tickers, more alpha — with the same time and effort.