The Teamwork Graph
The World Model Alternative: A Ledger for the Human Dimension of Work
In the era of AI, we are seeing a split in how companies approach intelligence. One path is the “General Intelligence” route: training massive models on the entire internet. The other—and arguably more valuable for an enterprise—is the **Contextual Intelligence** route.
This is the path Atlassian has chosen. By building the **Teamwork Graph**, they aren’t trying to build the smartest AI; they are building the smartest **map** for an AI to read.
### 1. What is it? (The “World Model” Alternative)
In physics, a “World Model” allows an entity to predict the outcomes of actions. In a corporation, the Teamwork Graph does the same for productivity.
It is a **Semantic Knowledge Graph** that sits underneath every Atlassian tool (Jira, Confluence, Loom). While traditional databases store “dead” rows of data, this graph stores “living” relationships. As of 2026, it maps over **23 billion work objects** and **84 billion unique relationships**.
* **Standard AI (RAG):** Searches for keywords. It’s like looking at a pile of loose puzzle pieces.
* **The Teamwork Graph: (with GraphRAG)** Understands the picture on the puzzle. It knows that a Slack message *resulted* in a Jira bug, which was *fixed* by a specific line of code, which was *demonstrated* in a Loom video.
### 2. The Architecture: How the “Brain” is Built
The complexity of this project is staggering. It is built on three core pillars that move it beyond simple software:
* **The Unified Data Model (UDM):** Atlassian’s “Universal Translator.” It normalizes data from disparate tools into a single language. A “Task” in any tool becomes a standardized **Work Node** in the graph, sharing the same DNA as a document or a person. This is a similar concept used by Palantir for it ontology.
* **The Secoda Layer (The 2025 Power-Up):** With the acquisition of Secoda, Atlassian added **Data Lineage** to the graph. This allows the AI to connect human conversations to raw numbers in Snowflake or SQL databases. It bridges the gap between *what people say* and *what the data shows*.
* **Activity Stream Ingestion:** This isn’t a batch process; it’s a “wiretap.” Every “Like,” comment, and code commit is ingested in real-time. The graph doesn’t just store the current state; it stores the **velocity** of the work.
* **Other Components:** The Relationship Engine, the Permissions Engine, Graph Storage, Vector Storage, the CDC technology are all complicated pieces that create the Atlassian moat.
### 3. Who is it for?
* **For the Individual:** It acts as a “second brain.” You don’t have to remember where a file is; you just ask “What was that thing Jamie was talking about in the meeting?” and the Graph follows the relationship from *Person* → *Meeting* → *Document*.
* **For the Executive:** It provides a “God View.” They can see not just *if* a project is delayed, but the exact “social friction” causing it (e.g., a specific team is consistently blocked by another).
* **For the AI Agent (Rovo):** This is the most important “user.” The Teamwork Graph gives AI the **Company IQ** it needs to act like a teammate rather than a chatbot.
### 4. The Moat: Why it’s (Nearly) Irreplicable
This is where the “investor logic” comes in. Why can’t a startup or a competitor like Asana just copy this?
1. **The “Lindy Effect” of Data:** Atlassian has decades of historical work patterns. You can’t simulate the nuance of how 300,000 companies actually work; you have to observe it over time.
2. **The Permission Wall:** Mapping data is easy; mapping **permissions** is a nightmare. The Teamwork Graph is “Security-Aware” at the traversal level. Building a system that can follow billions of links while ensuring a user only sees what they are allowed to see is a barrier to entry that takes years of engineering.
3. **The Integration Gravity:** Because the graph already contains the “State of Work,” every new tool you connect (Slack, GitHub, Salesforce) makes the graph exponentially more valuable. It is a **Network Effect for Data.**
### Summary: The Protocol of Teamwork
Just as Ethereum provides a secure, queryable ledger for financial state, the Teamwork Graph provides a secure, queryable ledger for **Intent and Execution**. It transforms a company from a collection of people using tools into a single, cohesive, and queryable “Organism.”
In the future, the companies that win will have a corporate “World Model” that they can query and that will proactively prompt them (e.g. did you know that the migration project is delayed due to human friction?). Large companies like **Block** and **Coinbase** built their own “World Models”, via dedicated in-house language models. This is a valid approach but requires large data projects to train the custom model and it needs to be constantly retrained. A “World Model” is a concept not an approach so the Team Graph (with GraphRAG) paired with a frontier LLM is also a “World Model” but without the training. That is, companies don’t need to build their own in-house world models, they essentially get the same benefits from a general LLM and the Atlassian Team Graph (with ROVO).