Enterprise AI has a memory problem. Not the kind of memory you fix with more RAM or a bigger vector store, but the kind that happens when an agent finishes a task, loses the conversation, and starts over from zero the next time you ask it to do anything vaguely similar. The models keep getting smarter. The interfaces keep getting slicker. The context, somehow, keeps disappearing. Asana’s chief product officer, Arnab Bose, states the issue bluntly: “Model providers are getting really, really good at improving reasoning and retry loops, but what they’re not good at is bringing the enterprise work context in a way that human beings can reason about for shared memory.” That last part — shared memory — is where the real competition is forming. On one side, Asana is treating memory as a multiplayer layer inside the work itself. On the other, NTT DATA AIVista wants to put memory into the infrastructure that runs the whole enterprise. Two very different architectures, one question: where should business context actually live?

The market data makes the problem hard to ignore. According to Asana’s own research, 75% of knowledge workers now use AI on the job, but only 5% of companies report actual productivity gains. The gap is not because the models are dumb. It’s because the context is fragmented. When someone corrects an AI agent — gives it better instructions, better feedback, better guardrails — that improvement usually vanishes the moment a colleague opens the same tool. Sriharsha Chintalapani, CTO and co-founder of Collate, described the issue to VentureBeat: “Agents are sensitive to the quality of their prompts. Someone with a strong understanding of the task will generally get more accurate results than someone less experienced.” The agent is only as good as the context that survives around it. As one commenter on HotMolts put it: “An agent that can't remember what happened five conversations ago is basically useless in enterprise settings.” That is the problem both Asana and NTT DATA AIVista are trying to solve. They just disagree about where the answer belongs.

Asana: memory as a multiplayer work layer

Asana’s bet is that enterprise memory should live inside the work graph — the data model that already contains tasks, projects, portfolios, and goals. That architecture is not new. The Work Graph has been around for nearly two decades, which means Asana is not bolting memory onto a generic chat interface. It is embedding memory into the structure that teams already use to coordinate. Asana calls its agents “AI Teammates.” Each one has an identity, scoped permissions, an audit trail, and cost constraints. The same console that governs human users controls what the agents can see, what actions they can take, and what they are allowed to spend. The memory layer itself is built around four functions: | Memory function | What it does | |---|---| | Learning | Memories are created during execution or supplied explicitly by users | | Retrieval | Relevant knowledge is pulled into future tasks automatically | | Access control | Memory respects the same permissions that govern human collaborators | | Transparency | Users can inspect and govern what the AI Teammate has learned | Asana’s agents build memory through two channels. The first is inferred memory: the agent learns from instructions, resources, actions, and feedback during execution. If someone says, “always copy the data model reviewer on these tasks,” the Teammate remembers. The second is explicit memory: users can directly teach the Teammate how to behave in a broader context, such as how to handle a particular resource or workflow. The key design choice is that memory is shared across the team, not siloed per user. When one person corrects an AI Teammate, the correction applies to everyone else who works with that agent. That turns one user’s feedback into institutional knowledge. The hard part is privacy. A personal assistant can treat memory as a simple extension of one person’s history. A shared enterprise memory cannot. Asana’s solution is to run memory through the same permission rules that already govern the Work Graph. The company says a Teammate only draws from memories that the current user has permission to see, preventing confidential information from leaking through a shared context layer.

Users say the results are real — but uneven

Asana has published internal and customer case studies that are worth reading with the usual vendor optimism in mind. Before deploying an AI Teammate, Asana’s APJ revenue team was manually triaging requests. A representative would fill out a form, someone would review it, and the average response time was three to four days. According to Asana, response time dropped to under a day after deployment. The team recovered 165.6 hours per week — roughly one full sales quota — and saved about $100,000 a year. The workflow automatically turned Slack messages into Asana tasks, then had the agent extract context and route the work to the right person. At the nonprofit Human-I-T, IT director Jean Favela built an AI Teammate that validates device specifications. Before, manual review took about two hours per day. Now it takes 30 minutes, and downstream data errors are at zero. Asana says the agent runs more than 14 hours a day, checking RAM, CPU, storage, and other fields. Marketing teams have gotten similar mileage. Asana’s marketing operations lead, Sheila Head, used an AI Teammate to review whether projects had complete setups — a manual check that used to take three to four hours per project. Quantitative researcher Lizzy Munro says competitive teardown work that used to take days now takes one to two hours. Asana also points to the COS brand, which sped up campaign production by 90%, and FedEx, which accelerated the production of key business documents by 9x. The aggregated numbers are equally strong. According to Asana, tasks managed by AI Teammates were 3.2x more likely to have clear ownership and 2.6x more likely to have defined deadlines. Early users completed work twice as fast. Perhaps most tellingly, 93% of AI Teammates were granted full edit access — not view-only or comment-only. That suggests the people closest to the work are willing to let agents execute, not just suggest. Not every user is happy. One Asana Forum user was pointedly annoyed: “NO ONE ASKED YOU FOR AI TEAMMATES. I find the AI bots that Asana creates to be more trouble than they are worth, and the fact that now, if I want to check the rules, templates and forms section, I have to first click through the AI teammate tab is ridiculous.” Others are more forgiving. A Gartner Peer Insights reviewer in April 2026 described the Asana team as “incredibly receptive to feedback on the product and hands on when we needed assistance and training.” On Reddit, the consensus is somewhere in between: “Asana is great for marketing teams but the pricing when you scale is brutal.”

NTT DATA AIVista: memory as enterprise infrastructure

NTT DATA AIVista comes at the problem from the opposite direction. It is not an application-layer company. It is a wholly owned subsidiary of NTT DATA, created on December 1, 2025, and based in Silicon Valley. The CEO is Bratin Saha, whose resume includes senior leadership roles at NVIDIA and AWS. The unit’s job is to help large companies operationalize agentic AI at a scale that most SaaS tools cannot touch. AIVista’s pitch is that enterprise AI needs a runtime intelligence layer — memory, orchestration, and context that sit underneath the applications and coordinate agents across multiple platforms. As Saha said at VB Transform 2026: “It's not just a model, you're building a system around the model.” That system is what NTT DATA calls the Enterprise Agentic Grid. According to an IDC Market Note, NTT DATA is moving beyond a services-only AI story toward a “services-as-a-product” model. The Grid combines a horizontal control plane, verticalized agentic applications, and a dedicated productization arm in AIVista. The target is not a better project management experience. It is a way to stop “platform sprawl” — the chaos that happens when agents run across Salesforce, ServiceNow, custom apps, cloud workloads, and legacy systems, none of which share context with the others. The infrastructure numbers explain why NTT DATA can make that argument. The company’s public disclosures show: | Metric | NTT DATA figure | |---|---| | Global data center capacity | ~1,630 MW | | FY2025 data center investment | ¥377.9 billion (~$2.5 billion) | | FY2026 planned data center investment | ¥505.0 billion (~$3.3 billion) | | FY2025 data center net sales | ¥520.3 billion | | Data center EBITDA margin | 55% | | Target capacity by FY2030 | >3 GW | That kind of capacity lets NTT DATA support sovereign cloud deployments, regional data residency requirements, and regulated workloads in a way that a single SaaS subscription cannot. For banks, insurers, manufacturers, and government agencies, that matters far more than how many pre-built agent templates are available. AIVista is also deliberately model-agnostic. Its architecture references OpenAI, Google, Anthropic, Mistral AI, tsuzumi2, and open-source LLMs. Asana leans heavily on Anthropic’s Claude through the Model Context Protocol. AIVista wants to be the layer above all models, not the enforcer of one model family.

Customer proof from the infrastructure side

NTT DATA’s customer stories are less flashy than Asana’s, but they are aimed at a different scale of problem. The company highlights a transnational insurance claim scenario where forms contain handwritten fields, dense checkboxes, and jurisdiction-specific regulatory requirements. That is the kind of messy, high-stakes document processing that a generic chat agent cannot handle without enterprise context. In more concrete examples, NTT DATA has worked with Damm, a Spanish beer manufacturer, to improve operational efficiency and service quality through AI applications. Dovista, a door and window manufacturer, partnered with NTT DATA to digitize parts of the financial closing process. Honda Trading Asia moved to AWS with NTT DATA’s help under tight time and cost constraints. Internally, NTT DATA claims that 9,000 employees now use AI tools, and a failure analysis that took five engineers three days can now be done in about 30 minutes. These are not headline-grabbing consumer features. They are back-office, operational, regulated workflows — which is exactly where the enterprise memory gap hurts most.

Side-by-side: where the philosophies differ