Twelve months into deploying Fund I, a pattern has emerged across every diligence call, every portfolio conversation, and every market map we've drawn: the enterprise workflow automation stack is more fragmented, more technically immature, and more expensive to maintain than the vendors operating in the space would have you believe.
This is not a criticism of the vendors. The existing workflow automation platforms — the RPA tools, the iPaaS layers, the no-code builders — solved a real problem for a specific set of enterprise customers at a specific point in the market. But the problem has changed. Enterprise workflows are no longer about connecting two SaaS applications with a trigger-action model. They're about orchestrating multi-step processes that involve API calls, database queries, conditional logic that changes based on real-time data, and increasingly, AI steps that require non-deterministic handling.
What changed: the introduction of AI steps
The existing automation platforms weren't designed for workflows with AI in them. They assume that each step is deterministic — that if you trigger an action, you get a predictable output. AI steps break this assumption at the architectural level. An LLM call might succeed but return a response that needs human review before the next step proceeds. A classification model might return a probability distribution rather than a binary decision. A generation step might produce output that needs to be validated against a business rule before being routed downstream.
Building enterprise automation infrastructure for AI-in-the-loop workflows requires a different architecture from the ground up: durable state management, built-in retry logic with backoff, human escalation paths that don't break the workflow state, and observability that captures not just whether a step succeeded but what the AI output was and why. We are at the beginning of this rebuild.
Where we're investing
Orchestration infrastructure — step-function-style execution engines that treat AI steps as first-class citizens, with built-in durability, replay, and fan-out. This is the most foundational layer of the stack and the one where we've seen the least mature existing tooling. It's also the one where we're most excited by what's being built.
The teams that build this layer correctly will be underneath every AI-native enterprise application built in the next five years. That's the kind of infrastructure bet we're designed to make — and it's where we expect to spend the majority of our remaining Fund I capital.