Asana Is Turning AI Agents Into Company Teammates With Shared Memory

Sapo

The next phase of enterprise AI is not about building a smarter chatbot. It is about giving AI a place inside the organization where it can remember what happened, understand why the work matters, collaborate with other people and agents, and still know what information it is allowed to see.

That is the bet behind Asana's Agentic Work Management, or AWM. Instead of treating an AI agent as a private assistant attached to one employee, Asana is building something closer to a shared operating layer for human and AI workers.

The interesting part is not simply that Asana is adding AI. The interesting part is that it is using nearly two decades of structured work data to give those agents organizational context.

And that changes the competitive question.

The race is no longer only about which company has the best model. It is increasingly about who owns the workflow, the context, the permissions and the feedback loop around that model.

The Main News

At VB Transform 2026, Asana chief product officer Arnab Bose explained how the company built Agentic Work Management around its existing Work Graph architecture.

AWM is already being used by several customers, including FedEx. CoreWeave is also using the system for complex product launches involving multiple AI teammates and deterministic workflow rules.

The product is designed around a simple idea. AI agents should behave less like isolated chat windows and more like members of an organization.

That requires memory.

But enterprise memory is much harder than simply storing previous conversations.

An AI agent might learn something while working on a confidential acquisition project. That information cannot automatically become available to another employee who later interacts with the same agent.

So Asana's challenge is simultaneously giving AI more organizational context while giving it stricter boundaries.

That tension is at the center of enterprise agent design.

What Is New?

The important shift is from prompt-based AI to workflow-based AI.

A conventional AI assistant receives a request, retrieves some information, generates an answer and finishes.

AWM attempts to preserve the state of the work.

When an AI teammate completes a task, the system records metadata about the execution. It can understand whether the task changed a project status and how the work relates to broader organizational goals.

That makes the agent part of an ongoing system rather than a disposable interaction.

The distinction sounds subtle, but economically it is enormous.

If every AI interaction starts from zero, companies are effectively renting intelligence one prompt at a time.

If AI can accumulate useful organizational context while respecting permissions, every completed workflow can potentially make the next workflow more effective.

The system begins to compound.

Why It Matters

Asana's biggest asset may not be its AI model.

It is the structure surrounding the model.

The company has spent 18 years building its Work Graph, which connects tasks, projects, portfolios and company goals. A task can have an owner and deadline. Tasks belong to projects. Projects roll into portfolios. Portfolios connect to larger business objectives.

That creates what Asana calls the Pyramid of Clarity.

Consider a delayed design task.

In a basic productivity application, the delay is simply a late task.

Inside a connected work graph, that delay can potentially be connected to a project, a portfolio and ultimately a company-level objective such as revenue.

This gives an AI agent something most standalone chatbots do not have.

Context.

The agent can understand not only what someone asked it to do, but also where that work sits inside the organization.

That is a much more valuable position.

How the Product Works

AWM combines several components.

First is the Work Graph, which provides the organizational context.

Second is persistent memory. Completed work can contribute information about workflows and outcomes.

Third is access control. The system determines which information can become shared memory and which information must remain restricted.

Fourth is model routing.

A user does not necessarily have to decide which model should handle a task. For something relatively simple, AWM can use a lighter and faster model. For a complicated task, it can route the work toward a more capable frontier model.

Bose cited Anthropic's Opus and OpenAI models as examples of models that could handle more complex work.

This is an important abstraction.

The employee should not have to become an AI engineer before assigning a task.

The experience should resemble delegating work to another employee.

You describe the outcome. The system decides how much computation, context and model capability are necessary.

Where the Data Lives

The underlying advantage comes from the Work Graph.

Instead of relying exclusively on a user's prompt or a local collection of documents, AWM can operate against a shared company ledger containing relationships between work items.

That is fundamentally different from connecting an LLM to Slack, Google Drive or another application through a conventional integration.

MCP-style integrations can provide access to information. But access alone does not create organizational memory.

The deeper question is what happens after the AI uses that information.

Does the result become reusable knowledge?

Does the system understand whether the work succeeded?

Does it know which project the work affected?

Can another authorized employee or agent build on it later?

AWM is designed around those questions.

How the Business Creates Value

The business model reveals another important part of the strategy.

AI workloads are unpredictable.

One task might require a relatively small amount of computation. Another could require extensive research, multiple model calls and large amounts of context.

If customers were billed according to every underlying token, model or computational decision, enterprise users would have to think about AI economics every time they delegated a task.

That creates friction.

Asana instead designed AWM around a static cost per completed task.

The platform absorbs the complexity of model selection, token usage and computational variability.

This resembles an important transition in enterprise software.

Customers generally do not want to purchase infrastructure complexity. They want predictable business outcomes.

The simpler the pricing abstraction, the easier it becomes for an organization to delegate more work to the system.

Platform or Product?

This is where Asana's strategy becomes more interesting.

AWM could be viewed as another AI feature inside a project-management product.

That would underestimate what Asana is attempting.

The larger opportunity is to become an orchestration platform for human-agent teams.

The difference is similar to the difference between a feature and an operating layer.

A feature helps one person complete one task.

An operating layer coordinates many tasks, people, agents, permissions and objectives.

If Asana succeeds, the Work Graph becomes more than a productivity database. It becomes a control plane for organizational work.

That is a much larger market position.

Competitors and the Advantage

The obvious competitive problem is that Asana depends on the same frontier-model ecosystem that is increasingly building its own agent products.

Anthropic and OpenAI can provide powerful models while also developing their own agent experiences.

That creates a classic platform conflict.

The model provider controls the intelligence layer.

Asana controls the workflow layer.

The strategic question is which layer captures more value.

Asana's argument is that a general-purpose model does not automatically understand the operating procedures, permissions, workflows and institutional history of a particular enterprise.

A general agent may be extremely capable.

But capability is not the same thing as organizational context.

Asana's 18 years of workflow data and its industry-specific operating procedures potentially provide a moat that is difficult to reproduce simply by improving a foundation model.

The model can be replaced.

The workflow graph is harder to replace.

The Risk

The largest risk is data governance.

Shared memory is powerful precisely because it is shared.

That also makes it dangerous.

A confidential project cannot accidentally become organizational knowledge merely because an AI agent participated in it.

Asana therefore has to distinguish between memory that should persist and execution that should remain isolated.

There is another risk.

If AI agents become deeply embedded in the Work Graph, Asana becomes increasingly responsible for the quality of decisions and recommendations generated from that graph.

Bad data can produce bad recommendations.

Outdated workflows can become automated mistakes.

And excessive automation can turn a previously visible human decision into an invisible system decision.

The more important the workflow, the more important those controls become.

What This Means for Businesses

For businesses, the implication is straightforward.

The first generation of enterprise AI focused on individual productivity.

Employees used AI to summarize documents, write emails, generate code and answer questions.

The next generation is moving toward organizational productivity.

Instead of asking, "How can AI help this employee?"

The better question becomes, "How should humans and AI divide the work inside this organization?"

That is a much bigger transformation.

CoreWeave offers a useful example.

Previously, product managers filled out complex forms containing infrastructure requirements, parameters and costs. Humans then reviewed those forms and divided the work among finance, marketing and hardware teams.

Under the AWM workflow, a product manager can provide a standard Google document referencing product requirements.

A deterministic AI workflow creates the project structure and assigns tasks.

Specialized AI teammates then execute pieces of the workflow.

One can monitor the overall project and identify bottlenecks.

Another can forecast infrastructure costs and recommend approvals based on historical budgets.

Humans remain responsible for evaluating the results.

The value is therefore not simply "AI does the work."

The value is AI restructures the workflow so humans spend more time on judgment and less time on coordination.

What Businesses Should Do

Companies should not begin by deploying hundreds of autonomous agents.

They should begin with workflows.

Identify processes where employees repeatedly gather the same information, create the same documents, move work between departments or perform predictable analysis.

Then ask four questions.

What context does the workflow require?

What information is safe to turn into persistent organizational memory?

Which decisions must remain human-controlled?

How should success be measured after the AI completes the work?

The last question is particularly important.

AI deployment without feedback creates automation without learning.

A company needs to know whether an AI-generated task actually improved the project, reduced cycle time, lowered costs or increased output quality.

That creates the foundation for a real organizational learning loop.

The Bigger Market Shift

There is a broader lesson here for enterprise software.

The value of AI is moving upward through the stack.

At first, AI was primarily a model.

Then it became a copilot.

Now it is becoming an agent.

The next step is orchestration.

That means coordinating agents with people, business rules, data, permissions and measurable outcomes.

In that environment, the winning product may not have the largest model.

It may have the best understanding of how work actually moves through a company.

That is why Asana's Work Graph matters.

It is not merely a database of tasks.

It can become a representation of how an organization operates.

And once software understands the organization at that level, the boundary between project management, workflow automation and AI execution begins to disappear.

Conclusion

The most important idea behind Asana's Agentic Work Management is not that AI agents can perform more tasks.

It is that AI agents can potentially become persistent participants in an organization's operating system.

The Work Graph gives them context.

Memory gives them continuity.

Access controls give them boundaries.

Dynamic model routing gives them computational flexibility.

Predictable task-based pricing makes the economics easier for enterprises to understand.

Put those pieces together and Asana is making a much bigger bet than adding another AI assistant.

It is betting that the future of enterprise software will be built around teams containing both humans and machines.

The winning platforms in that world will own the context in which work happens.

Models will continue to improve. New agents will appear. Interfaces will change.

But the organization still needs to know who is doing what, why it matters, what happened last time, what can be shared and what must remain private.

That is the real battlefield for enterprise AI.

And Asana is positioning its Work Graph directly in the middle of it.

AI Agents Need Company Context to Become Truly Useful

SEO keywords

Asana AI agents, Agentic Work Management, Asana Work Graph, enterprise AI agents, AI shared memory, AI workflow automation, human AI teams, enterprise AI platform, AI agent orchestration, AI workflow management, persistent AI memory, enterprise AI governance, AI model routing, AI productivity software, enterprise automation