Asana Built an AI That Remembers Your Team's Work Without Leaking Secrets

The Stateless Agent Problem

Enterprise AI deployments share a common frustration: chatbots that answer questions in isolation but retain nothing. A marketing manager asks an agent to draft a campaign based on last quarter's performance. The agent pulls data, generates output, and vanishes. The next person building a similar campaign starts from scratch. No workflow persists. No institutional knowledge accumulates.

Asana's chief product officer Arnab Bose frames this as the core limitation of current enterprise AI. Standard copilots operate in a vacuum, tethered to individual prompts with no connection to the broader organizational context. They can fetch information but cannot see how a delayed design task cascades into a missed revenue target, or how one team's workflow might benefit another.

The company's answer is Agentic Work Management, or AWM. The platform repurposes Asana's existing enterprise architecture to create what Bose calls "coachable teammates" rather than one-off assistants. These AI agents draw from a shared memory layer spanning the entire organization, updating their knowledge as they complete tasks and learn from human feedback.

An 18-Year Architecture Gets a Second Life

AWM runs on top of the Work Graph, Asana's graph-based database that has organized enterprise work since the company's founding. The structure follows what Asana calls the Pyramid of Clarity: tasks at the bottom, each assigned to a person with a due date. Tasks roll up into projects, projects into portfolios, portfolios into company-wide goals.

This architecture creates a real-time ledger of organizational activity. The graph can trace dependencies across layers, showing how a single delayed task impacts a portfolio milestone or a corporate objective. For AWM, this existing structure eliminates the need to build an enterprise context graph from scratch.

Because the AI agents plug directly into the Work Graph, they operate with full visibility into organizational priorities. An agent completing a task doesn't just execute the prompt; it updates project status, logs progress toward goals, and makes that context available to the next human or agent that needs it.

Bose describes the distinction clearly: the agent isn't looking at a single user's markdown file or a prompt in isolation. It works from a shared ledger visible across the company, creating a persistent state that accumulates knowledge over time.

Confidential Work and the Memory Boundary Problem

Shared memory introduces a technical challenge: data governance. If an AI agent learns from every interaction, what stops it from leaking confidential information to unauthorized employees?

Bose outlined the scenario Asana had to solve. An executive uses AWM to build workflows for a confidential acquisition. The agent completes tasks, updates its memory, and improves its understanding of the executive's preferred style and priorities. Later, an employee without clearance interacts with the same agent. If the system isn't carefully designed, the agent might inadvertently reveal context from the confidential project, even if the employee never directly accessed those files.

Asana engineered a system of access controls to govern memory creation and retrieval. The platform distinguishes between executing a task and creating a memory. When an agent completes work on a restricted project, it logs the task but isolates the memory behind the same permissions that govern the underlying Work Graph data.

This approach ensures that an agent's accumulated knowledge respects organizational boundaries. An employee querying the agent will only receive information drawn from projects and tasks they have explicit permission to view. The shared memory remains shared, but only within the appropriate access tiers.

Dynamic Routing and Billing Abstraction

AWM abstracts away the complexity of prompt engineering and model selection. When a user assigns a task to an AI teammate, the system determines the computational requirements and routes the request to an appropriate model.

A complex task like drafting a job description for a senior role triggers a call to a frontier model. The agent cross-references public job postings, internal style guides, and product requirement documents to generate output. Simpler tasks get routed to faster, cheaper models that can handle the workload without burning unnecessary credits.

This dynamic routing creates a third engineering problem: unpredictable costs. Agentic tasks vary widely in complexity. A user drafting a marketing plan might consume ten times the tokens of a user updating a status report. If Asana charged per token or per model call, customers would face volatile bills and might ration how often employees could use the system.

To solve this, Asana designed a flat-rate billing model. Customers pay a static cost per task completion, regardless of which model the system uses or how many tokens the task consumes. The platform absorbs the variance, making pricing predictable and removing the need for customers to throttle their own teams.

CoreWeave's Product Launch Overhaul

Cloud infrastructure provider CoreWeave adopted AWM to streamline its product launch process. Previously, product managers filled out lengthy forms detailing infrastructure requirements, cost estimates, and timelines. Human reviewers manually parsed these forms and distributed tasks to finance, marketing, and hardware teams.

Under AWM, a product manager writes a standard document linking to product requirement specs. A deterministic AI workflow reads the document, generates the project structure, and assigns tasks automatically. Specialized agents then handle execution. One agent monitors overall project status and flags bottlenecks. Another works inside individual tasks, forecasting infrastructure costs and recommending approvals when projections align with historical budgets.

The system automates the busywork while keeping humans in the loop for evaluation and decision-making. CoreWeave's teams now spend less time on administrative coordination and more time assessing the AI's recommendations.

The Frenemy Dynamic

The frontier model providers powering AWM are also shipping competing agent products. Anthropic offers Claude in Slack, branded as Tag. OpenAI is building similar integrations. These companies supply the underlying models Asana relies on while simultaneously trying to capture the same enterprise workflows.

Bose acknowledges the tension but argues Asana's advantage lies in depth, not raw model capability. The company has 18 years of workflow data and prebuilt standard operating procedures tailored to specific industries. A general-purpose chatbot in Slack can answer questions and fetch data, but it requires careful channel curation and separate credentials for every downstream application it touches.

AWM, by contrast, ships with end-to-end workflows already configured. The AI teammates understand project hierarchies, know how to update goals, and operate within the same permissions structure employees use. Bose frames this as the difference between demonstrating model power in a lightweight interface and delivering a production-ready system that handles real organizational complexity.

What This Means for Enterprise AI

Asana's approach highlights a strategic choice facing companies building agentic systems. One path is to integrate frontier models into existing communication tools like Slack or email, creating lightweight copilots that assist with individual tasks. The other is to embed agents into the operational systems where work actually happens, giving them access to organizational state and the ability to update it.

AWM chooses the latter. The platform treats AI agents as participants in a shared workspace, not as external assistants summoned for one-off questions. This design requires solving harder technical problems around access control, billing abstraction, and memory persistence, but it creates a different kind of product: one that accumulates institutional knowledge and improves as the organization uses it.

FedEx is among the early adopters already running AWM in production. The company published its own case study documenting the transition. Bose noted that several other customers are live on the platform, though he didn't name them.

The question for enterprises evaluating agentic systems is whether they want AI that assists individuals or AI that operates as part of the organizational fabric. Asana is betting that the latter category represents the bigger opportunity, even if it means competing with the same companies that supply the models underneath.