Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
The most common complaint about enterprise AI tools is not that they are too slow or too expensive. It is that they forget everything the moment a conversation ends. Every new prompt is a blank slate, every completed task vanishes without a trace, and the organisation learns nothing. Asana’s chief product officer Arnab Bose spent a fireside chat at VB Transform 2026 with VentureBeat’s Sam Witteveen explaining why that architectural flaw is the real bottleneck, and how his team built something designed to fix it.
The product is called Agentic Work Management, or AWM. It treats AI agents not as one-to-one assistants that serve a single user’s prompt, but as coachable teammates operating inside a shared company-wide context. Bose confirmed that AWM is already in production, with several enterprise customers live on it, including FedEx, which has published its own case study on the deployment.
Eighteen Years of Workflow Data as an AI Foundation
To understand what AWM actually does differently, you need to understand the substrate it runs on. Asana built AWM on top of its Work Graph, a graph-based database architecture that has been accumulating enterprise workflow data since the company’s founding. The Work Graph organises information through what Asana calls the Pyramid of Clarity. At the base sits a task, with an assignee and a due date. Tasks belong to projects, projects roll up into portfolios, and portfolios connect to company-wide goals. The structure means the system can trace, in real time, how a delayed design task affects a corporate revenue target.
This is the key distinction from a standard AI copilot. A typical large language model integration is stateless and scoped to a single user’s session. AWM plugs the AI into the Work Graph, which functions as a live ledger for the entire organisation. As Bose put it, the agent is “not just looking at a particular prompt that you’re sending it or looking at a particular individual’s markdown file system on their local file. It’s working off of that shared ledger for the whole company.”
That shared memory is what makes the system genuinely different from connecting an LLM to Slack or Google Drive via a Model Context Protocol integration. Those connections let an agent fetch data to answer a prompt, but the execution happens in isolation. It benefits one person once, and creates no reusable workflow for the next person building something similar. AWM records metadata on every completed task, including whether the completion improved project status and how it moved higher-level company goals, creating a permanent institutional record rather than a series of disposable conversations.
The Hard Engineering Problems Nobody Talks About
Shared memory across an organisation sounds powerful until you consider what it means for confidential work. Bose was direct about the boundary problem Asana had to solve. If an executive uses AWM to build workflows for a sensitive merger and acquisition project, the agent’s updated memory from that work must not leak context to an employee who later interacts with the same agent but has no access to that project. “I shouldn’t be able to leverage that shared memory when I run the AI teammate if you created that memory using that same teammate on a project that is, let’s say, a secret M&A project that I don’t have access to,” Bose said. Asana engineered a system of access controls that governs what triggers the creation of a memory versus the simple execution of a task, keeping the two distinct at an architectural level.
The second engineering challenge is model routing. When a user assigns a task to an AI teammate, say drafting a job description for a general manager role, the system automatically cross-references public job postings, internal style guides, and product requirement documents. Complex tasks are routed to heavy frontier models, with Bose citing Anthropic’s Opus and OpenAI’s models as examples, while lighter tasks are handled by faster and cheaper alternatives. The user never chooses. “We don’t want the knowledge worker to have to think through what the best possible prompt, context engineering, and attachments are,” Bose said. “It should feel as if you were assigning the task to a human being.”
Dynamic model routing creates a third problem: billing unpredictability. If customers have to reason about variable credit burn rates depending on task complexity, they will start capping how often employees can run AI teammates, which defeats the purpose. Asana’s solution is to charge a static cost per task completion, absorbing the complexity of model selection, token counts, and run limits internally. Predictable pricing is not a marketing decision here. It is a prerequisite for enterprise adoption at scale.
CoreWeave’s Product Launch Overhaul as a Real-World Test
Cloud provider CoreWeave offers the clearest illustration of what AWM looks like in practice. Previously, CoreWeave product managers filled out detailed forms covering infrastructure, parameters, and costs. Human reviewers then manually evaluated those forms and broke them into specific tasks for finance, marketing, and hardware teams. The process was labour-intensive and slow.
Under AWM, a product manager writes a standard Google document pointing to their product requirement documents. A deterministic AI workflow reads the document, automatically creates the project structure, and assigns tasks. Specialised agents then handle execution. One agent monitors overall project status and flags bottlenecks. Another, working inside individual tasks, forecasts infrastructure costs and recommends approvals when the numbers align with historical budgets. Human attention shifts from processing busywork to evaluating the AI’s outputs, which is a meaningful change in how knowledge work is actually structured.
Why the Frontier Model Providers Are Both Partners and Rivals
There is an obvious tension in Asana’s position. The same companies whose models power AWM under the hood, Anthropic and OpenAI, are also shipping their own competing agent products. Anthropic’s Claude integration in Slack, referred to as Tag in the discussion, is a direct example. Bose did not shy away from the overlap. “I think that’s the reality that we all have to live in,” he said.
His argument for AWM’s durability rests on two things that raw frontier models do not have. The first is Asana’s 18 years of user-experience and workflow data. The second is prebuilt standard operating procedures for specific industries. A product like Tag can work well inside Slack, Bose acknowledged, but it requires a highly curated channel and its own separate credentials for every downstream application it connects to. “There’s a big difference between the power of the model plus a lightweight way to demonstrate its value, and something that’s pre-built for true end-to-end use,” he said.
That distinction matters beyond Asana’s competitive positioning. For enterprise technology buyers in Malaysia and Singapore evaluating where to invest in AI infrastructure, the AWM architecture illustrates a broader principle: the value of an AI deployment increasingly depends not on which model sits at its core, but on the quality of the organisational context that model can access and the governance controls that determine who can see what. Stateless chatbots are easy to deploy and easy to outgrow. Building the memory layer is the harder, slower, and ultimately more consequential work.
Read More: Your Analytics Dashboard Is Lying to You About AI Commerce
