Conference Presentation, Fireside Chat, Interview
The Missing Layer: Why Agentic AI Fails Without a Work Operating System | Asana | RAISE 2026
- Mike Butcher, founder and editor of Path Founders, introduces the platform as a new technology media brand reporting on the AI revolution in Europe.
- AR, CTO of Asana, identifies the primary enterprise misconception as the "chasm" between unlocked individual productivity via AI and the difficulty of translating this into group or business productivity.
- The disconnect is attributed not to "shadow AI" or lack of control, but to the absence of mature platforms for "agentifying" business processes, which require system integration, multiplayer collaboration, and governance.
- Asana differentiates itself by leveraging its foundational structure for human coordination, which naturally aligns with the task-decomposition and checkpoint mechanisms required for AI agents to function effectively in workflows.
- Despite market volatility and competition from "AI-native" startups (e.g., PromptQL raising $136 million), Asana remains well-positioned with 185,000 customers, including 80% of the Fortune 100.
- AR argues that legacy platforms should not be discarded but can serve as the necessary "harness" for AI, a view informed by his previous experience managing the shift from on-premise to cloud at Microsoft.
- Echoing Alex Karp's concerns, AR warns that enterprises risk losing their "compounding asset" (IP) if they allow model providers to operate too deeply within the stack, advocating for a layer that keeps the firm in control of data, traces, and prompts.
- Asana aims to be the intermediary layer between Large Language Models (LLMs) and enterprise agents to protect intellectual property and maintain vendor optionality across different model providers.
- Strategic advice for enterprises includes experimenting with multiple models and harnesses to optimize for both quality and cost, noting that output quality depends on the data harness and cleanliness rather than the model alone.
- Asana explicitly rejects the "token maxing" business model, a decision AR credits for keeping engineering morale high while avoiding the pitfalls of overspending on API calls.
- In the context of geopolitical tensions regarding open-source vs. frontier models (e.g., export controls, Chinese restrictions), AR advises maintaining optionality to prevent organizational paralysis if access to specific models is restricted.
- Future enterprise AI infrastructure will likely feature platforms that aggregate diverse agents, providing CIOs with essential capabilities: strong identity management, audit trails, cost control, and debugging tools.
- Quality control for AI agents will rely on breaking workflows into stages interspersed with deterministic checks (e.g., linters, API specs) or human approval tasks, mirroring traditional project management checkpoints.
- AR emphasizes that while agents can run autonomously, business-critical processes require "clamping down" on execution through periodic verification to ensure safety and accountability.
- Engineers at Asana are under instruction to maximize AI utility without compromising engineering craft, ensuring top talent does not become "burned out" or desensitized to quality by reading "AI slop."
- Token pricing is expected to decrease, but enterprises must retain control over vectoring between models to prevent margins from being eroded by LLM providers.