monday.com: From Systems of Record to Context

At the AI Engineer World's Fair, Omri Bruchim and Tomer Ast, engineering managers at monday.com, discussed the company's strategic shift from being a 'system of record' to a 'system of context'. This evolution is central to providing more intelligent and truly helpful AI assistance to users.

The Limitations of Current AI Assistants

Bruchim and Ast highlighted a common frustration with current AI assistants: their inability to provide truly actionable insights, even with access to vast amounts of data. They posed the question, "What should I focus on right now?" and noted that while AI can generate lists of tasks or paragraphs, these often lack genuine understanding of the user's priorities and context. This leads to generic or irrelevant suggestions, such as an AI advising a user to go to the gym when they're facing a critical work issue.

The core problem, they argued, is not a lack of data but a lack of 'understanding.' "The problem was never the missing of data, the retrieval. The problem is like the missing understanding," Ast stated. He emphasized that understanding, not just context or memory, is the key differentiator.

The Agent Gap and the Meaning of Records

The presentation identified three key challenges in building context-aware AI:

• The Agent Gap: Agents are sharp at executing tasks when the problem is clearly defined, but they struggle to identify what the actual problems or priorities are.
• Records vs. Meaning: Logs and records capture events, but not the underlying 'why.' For instance, a line of code in a repository tells you what happened, but not the intent or the problem it solved.
• Runtime vs. Precomputation: Understanding needs to be built ahead of time, not generated on the fly when a question is asked.

To address these challenges, monday.com is developing its 'Monday World Model.' This model aims to provide context that follows the user's work, understanding who they are, what they are working on, what actions are needed, and what should be avoided. It's described not as a larger prompt or longer context window, but as a fundamentally different approach.

The Monday World Model: Two Engines, One Understanding

Tomer Ast explained the architecture of the Monday World Model, which relies on two engines operating on different time windows:

• The Slow Engine: This engine operates over a long time window, mining user activity for patterns to build a durable profile. It learns the user's persona, routines, work rhythm, collaborators, and goals, reinforcing the profile over time.
• The Fast Engine: This engine focuses on a short, recent window to recompute live signals about the user's current state, identifying what's overdue, urgent, or who they are actively collaborating with.

This dual-engine approach is inspired by concepts in neuroscience ('complementary learning systems') and data infrastructure ('lambda architecture'), where a fast layer processes recent data, and a slower layer processes historical data, with both merging into a unified view. The model is built to be resilient, with isolated sources, and to degrade gracefully if a data feed fails.

Key Behaviors and Future Growth

The model's design yields two critical behaviors: resilience and the ability to understand urgency. It knows when to be proactive and when to remain silent. Crucially, the model is designed to compound its learning; as more data is captured daily, the profile sharpens, and the model understands more. Adding new data sources is intended to be cost-effective, continuously expanding the model's understanding.

Ast acknowledged the inherent challenges, such as the model always trailing the live world, the 'cold start' problem for new users, and the biases that can be built into signals. However, he stressed that the architectural design allows for continuous enrichment, testing, and improvement.

Answering the Core Question

The presentation concluded by returning to the initial question: "What should I focus on right now?" The speakers demonstrated how the Monday World Model, by processing a user's entire work history, from boards and emails to Slack messages and meeting notes, can now provide a much more relevant and contextualized answer, moving beyond mere data retrieval to genuine understanding.