Should Software Companies Embrace AI or fight it? — With Asana Chief Product Officer Arnab Bose
TL;DR
Asana’s bet is that enterprise software won’t be replaced by vibe coding — it’ll become the coordination layer for humans and agents — Arnab Bose argues customers shouldn’t burn tokens rebuilding security, reliability, integrations, and 99.99% uptime when Asana has spent over a decade building that infrastructure.
The real moat is context and shared memory, not just model access — Bose says Asana’s “work graph” gives agents enterprise-grade context from tasks, projects, portfolios, and company goals, while shared memory lets one teammate learn from feedback any user gives it.
Asana wants AI to remove coordination work, not replace taste and judgment — in his marketing example, an AI teammate can draft a creative brief from past campaigns and web research, getting it to 80-90% so a creative director can focus on differentiation instead of boilerplate.
Asana just launched 21 pre-built AI teammates and custom agent creation is live now — the agents handle jobs like campaign brief writing, IT ticket deflection, and launch-plan risk analysis, and customers can also build their own by prompting a builder and granting access to relevant projects and portfolios.
Bose thinks most companies are still getting ‘high-velocity noise,’ not business outcomes, from AI — his critique is that strong models alone aren’t enough unless outputs are heavily grounded in enterprise-specific memory, otherwise humans just spend more time reviewing longer generic content.
On models, Asana is pragmatic and frontier-first — the company uses both OpenAI and Anthropic across Asana AI, launches AI teammates on Anthropic Opus 3.6, and avoids open source for now because being even 3-6 months behind frontier capabilities would be strategically costly.
The Breakdown
Why Bose thinks vibe coding misses the point
Alex Kantrowitz opens with the obvious anxiety: if software can be vibe coded, what happens to a company like Asana? Arnab Bose’s answer is blunt — why would a travel company, bank, or healthcare provider want to spend tokens solving security, permissions, reliability, deep integrations, and 99.99% uptime instead of focusing on its actual business?
Asana’s AI strategy: embrace customization, own the coordination layer
Bose says Asana isn’t fighting customized AI workflows at all — it’s leaning in. His pitch is that Asana’s decade of work on the “coordination tax” and its “pyramid of clarity” makes it the right substrate for enterprises to bring their own agents or customize Asana’s built-in ones.
The work graph is the whole game
When Alex asks what Asana actually does, Bose grounds it in cross-functional work like marketing campaigns, launch plans, and strategic ops. The key twist is that Asana doesn’t just track who does what by when — it also stores how similar work went in the past, what broke, and how teams fixed it, which Bose calls “catnip for agents.”
From annoying approvals to AI-written creative briefs
The most concrete demo is marketing: a team kicks off a campaign, assigns the first task to an agent, and the agent uses historical briefs, public web research, and live team feedback to build a research plan and draft a creative brief. Bose emphasizes the multiplayer aspect here — instead of one marketer privately prompting ChatGPT or Claude, everyone can see the AI’s reasoning, feedback, and approvals inside Asana.
AI should free up taste-makers, not flatten creativity
Alex presses on the obvious creative risk: won’t AI produce the “average of averages”? Bose says only if it lacks context; with company-specific historical work, the goal is to get to 80-90% quality fast so the creative director can spend their time on taste and judgment, not blank-page work. He even frames this through Jensen Huang’s line that using AI for layoffs “lacks imagination,” arguing the smarter move is to expand output and markets, not just cut headcount.
Why most AI deployments still disappoint
One of Bose’s sharper points is that companies aren’t getting exponential outcomes from AI yet because they’re producing generic content at high speed, then forcing humans to sift through “reams of text.” His product thesis is to turn that into enterprise-specific output that improves every time someone gives feedback, without requiring each employee to maintain their own little prompting playbook.
The launch: 21 AI teammates, custom agents, and guardrails
Asana’s AI teammates are generally available now, with 21 pre-built agents for jobs like campaign brief writing, IT ticket deflection, and launch-plan analysis. If a company wants its own “creative director” agent, it can build one by prompt, give it access to the right projects and portfolios, and let it request starter tasks — but Bose stresses humans still control approvals, assignments, and scope.
Why Asana picked Anthropic now — and why one master agent is a bad idea
On model choice, Bose says Asana uses both OpenAI and Anthropic across the product, but the AI teammate launch runs on Anthropic Opus 3.6 because it performed best in testing. He’s not using open source models right now, arguing frontier labs are moving too fast to justify falling 3-6 months behind — and he closes by saying the future probably isn’t one all-knowing life agent anyway, because mixing personal and work memory is exactly how sensitive information leaks.