Instinct reaches $10 billion valuation as AI agents target travel
Instinct's $10 billion valuation, Meta's rapid Muse adoption and a wave of agentic assistants put travel bookings, reservations and administrative work at the center of AI's next contest.

By Marcus Bennett
Instinct raised $1 billion at a $10 billion valuation in September with 14 employees, no subscription price and an invitation-only assistant launched only that August. Investors had valued the startup at $2.5 billion one month earlier. Its founder was 23.
Silicon Valley sees personal agents, not larger chat models, as the next prize. Meta, xAI, Anthropic, OpenAI and startups aim to operate across email, calendars, reservations, schedules and eventually money. The practical test is completion, not prompt response.
Bennett, who tested several products, said: “I send a message. It gets to work.” Instinct runs through text or email, remembers prior conversations and manages tasks without elaborate prompts or workflows.
One user reported that it found a document awaiting signature. Another used Instinct through a connected OpenTable account to change flights, secure a difficult restaurant reservation and add the booking to a calendar. Coordinated agents can exchange scheduling information and arrange dinner without repeated messages.
That resembles administrative labor, but blocked websites, misunderstood instructions and human intervention remain visible limits. A concierge still needs measurable error rates before operators can trust it with bookings.
Who has the distribution advantage?
Meta has the clearest distribution advantage. In June 2025, Mark Zuckerberg committed about $14.3 billion to acquire 49% of Scale AI and recruit its 28-year-old founder, Alexandr Wang, for Meta’s superintelligence effort. Meta recruited rival researchers, and former OpenAI researcher Shengjia Zhao became chief scientist.
Meta released the Muse Spark model in April 2026 and launched personal agent Muse in September. The product reportedly drew 2.5 million downloads in 13 days and passed 3 million by week three. WhatsApp, Instagram, Facebook and Messenger give Meta a distribution path startups must earn.
Anthropic introduced Claude Duet in October with access to more than 6,000 applications and tools. Duet prepares briefings, monitors work, coordinates information and handles financial analysis, reporting and software development. It remains in research preview, and Anthropic acknowledges limits in interpreting priorities and handling some tasks.
OpenAI’s Dots preserves relevant context, works across connected applications and continues assignments, particularly appealing to established ChatGPT users. xAI’s Grok Bot builds specialist agents with shared context, persistent computers, recurring workflows and participation in Slack and Microsoft Teams. A research bot can inform a production bot, which another bot can review.
Which products fit which workflows?
- Instinct: Connects to calendars and email, handles reservations, purchases and coordination, retains context, and its founder says travel accounts for more than 50% of transactions. The company self-reported nearly $1 billion in annualized transaction volume. The free beta has no recurring fee. The company has disclosed neither revenue nor transaction fees.
- Catch AI: Costs $99 a month and focuses on executive administration: meeting coordination, calendars, email follow-up and business calls that navigate phone systems. The price gives consultants and small businesses an expense to compare with administrative labor.
- Meta Muse: Organizes travel, researches purchases, makes reservations and manages tasks through an app or WhatsApp. It keeps working after a conversation closes and offers a free tier. Questions remain over privacy and human intervention in agent-supported calls.
- Claude Duet: Could lead the professional category by handling ongoing work and assignments requiring context from connected systems. Its research-preview status makes reliability claims provisional.
- Grok Bot: Lets users assign research, deliverable and review roles to coordinated agents. Its value lies in delegation across tools, but oversight and consistency with complicated requests remain unproven.
- OpenAI Dots: Builds on existing ChatGPT relationships and persistent project context. The central test is whether retained history produces completed work without more errors.
- Viktor: Works through Slack and Microsoft Teams to update business systems, prepare reports and manage recurring workflows. Its credit-based pricing leaves the monthly cost of frequent or complex assignments uncertain.
- Lindy: Configures agents for email, scheduling, CRM work and business applications. It suits repeatable processes, but setup, testing and maintenance can offset time savings.
- Martin: Uses voice and messaging for reminders, scheduling and general coordination. Its handling of complicated or unexpected requests needs testing.
- Must AI Assistant: Offers specialized agents for writing, learning, planning and creative work. It is a productivity suite, not a direct substitute for agents that execute real-world transactions.
What creates switching costs?
Useful assistants accumulate priority rules, communication preferences, hotel choices, contacts, delegated tasks and approved decisions. Connected accounts and histories raise switching costs. Meta has social reach, OpenAI has a ChatGPT audience, and Anthropic has professional software connections.
Smaller vendors must make their services habitual before funding and valuations can rest on durable demand, because switching costs grow as accounts and history deepen.
What happens when agents act?
Calendar access differs from authority over email, contacts, purchases and financial information. Meta uses a dedicated virtual environment for Muse and requires approval for certain purchases and outgoing communications. Other vendors also use virtual computers, protected credential storage and approval controls. Those safeguards do not prevent an assistant from misreading a reservation, an email or an instruction.
Developers should disclose what they retain, how they use it and whether users can revoke access without dense legal language, because reliability and privacy will determine whether these systems replace administrative work.
Six months earlier, Bennett would not have expected to assign real business and personal tasks to several assistants. That change helps explain extraordinary valuations, but it does not show whether the businesses make money. No product has earned a universal verdict. The next test is repeat bookings, lower administrative labor and fewer correction calls, not another polished demonstration.
More from Marcus Bennett
Show full bio
Market editor covering technology and product news in Restaurant and hotel operations at The Pass Brief.
712 articles


