AI Agents Are Running Enterprise Workflows — But Nobody Knows Who's Responsible When They Go Wrong

Here's a fact that should stop you cold: AI agents are now autonomously handling millions of high-stakes enterprise tasks — without any framework for who they are, what they're allowed to do, or who gets blamed when something breaks.

This isn't hypothetical. Two companies on this year's list have already crossed extraordinary scale milestones:

The problem? These agents operate on enterprise systems but exist in a total governance vacuum. They have:

• No persistent identity
• No verifiable owner
• No scoped authority
• No audit trail tied to a principal

The companies stepping into this gap — and building what might be called Know Your Agent (KYA) infrastructure — are among the most strategically important on this year's entire list:

The companies building the AI agent rulebook:

Physical AI Is No Longer a Prototype — It's Scaling From Single Robots to Coordinated Fleets

Physical AI — the category of systems that perceive, decide, and act in the real world — just became a standalone category in the AI 100 for the first time ever. That's not symbolic. It means the infrastructure to deploy autonomous systems commercially has, for the first time, matured enough to compete alongside enterprise software.

The sector raised a record $78 billion in 2025. The 11 Physical AI winners on this year's list span autonomous warships, general-purpose robots, and industrial humanoids.

The emerging frontier: fleet coordination

Most enterprise robot deployments today are still single-unit. But the companies on this list are already solving the next problem: getting multiple heterogeneous robots to work together toward a shared objective.

The Vertical AI Winners Are Defined by Their Data — Not the Sector They're In

This is perhaps the most counter-intuitive insight from the 2026 cohort. You might expect the strongest vertical AI companies to be the ones with the deepest domain expertise, the best sales team, or the slickest product. The CB Insights analysis tells a different story:

The durable businesses are the ones with data that nobody else can replicate — regardless of what sector they serve.

Three distinct patterns emerge:

Pattern 1: Non-textual data that generic AI can't touch

Where the underlying data is molecular structures, CAD geometry, or materials properties, general-purpose models like GPT or Gemini cannot natively represent it. Companies in this space have to build their own models — which becomes the moat itself.

Chai Discovery built its own antibody design model and grew from a $150M to $1.3B valuation in just 15 months. Leo AI reports 96% accuracy on mechanical engineering questions versus 46% for generic tools.

Pattern 2: Switching costs baked into critical workflows

In financial services, companies like Bretton AI, Further AI, and Salient build on top of existing models — but they embed so deeply into compliance workflows and lending systems that replacing them becomes genuinely painful.

Pattern 3: Rare datasets that cannot be replicated

Where data is text-based but hard to acquire — regulated patient records, licensed databases, institutional knowledge — the dataset itself is the moat. Two standout examples:

• Atomic Canyon trained on the NRC's 53-million-page regulatory database — a corpus that no competitor can simply "also train on."
• Assort Health has encoded 125M+ patient interactions without building its own foundation model. The data differentiates, not the architecture.