Skip to content
a16za16z

Box CEO on the AI Adoption Gap | The a16z Show

Erik Torenberg, Steven Sinofsky, and Martin Casado speak to Aaron Levie, CEO at Box, about what happens to enterprise software when agents become the primary users. They discuss why coding agents succeed where other knowledge work agents struggle, what abstraction layers mean for the workforce, and how data access and systems of record must change in an agent-first world. Timestamps: 0:00—Intro 0:51—Building software for agents vs. humans 2:10—Can non-technical workers actually use AI agents? 14:31—CFO/CIO pushback: the real fear of agents doing integration 18:39—Treating agents like employees and why it breaks down 27:35—Diffusion gap: startups vs. enterprises 42:53—Wall Street's economics are off by an order of magnitude Read the full transcript here: https://www.a16z.news/s/podcast Resources: Follow Aaron Levie on X: https://twitter.com/levie Follow Steve Sinofsky on X: https://twitter.com/stevesi Follow Martin Casado on X: https://twitter.com/martin_casado Follow Erik Torenberg on X: https://twitter.com/eriktorenberg Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Aaron LevieguestSteven SinofskyhostMartin Casadohost
Apr 7, 202658mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Why AI agents spread slowly in enterprises, yet reshape software economics

  1. AI-native workflows will require software to be designed for agents at massive scale, but most workers lack the algorithmic thinking needed to reliably “drive” agents today.
  2. Near-term AI value will skew toward the consumption layer (finding, navigating, reporting) while write-actions and on-the-fly integrations trigger CFO/CIO fears about breaking systems of record and governance controls.
  3. Treating agents like employees (separate accounts, phones, cards) works for personal use but breaks down in enterprises due to oversight, liability, privacy, and prompt-injection/data exfiltration risks.
  4. Large incumbents and systems like SAP/Workday won’t be replaced by “vibe coding,” which slows enterprise diffusion even as startups move faster by having fewer constraints and less to lose.
  5. Wall Street and many CFO budgeting models underestimate AI’s opportunity and usage growth by an order of magnitude, while compute/token spend becomes a volatile new engineering budget line item until supply and efficiency improve.

IDEAS WORTH REMEMBERING

5 ideas

Agents will be numerous enough to force “agent-first” product design.

The group hypothesizes 100–1,000× more agents than people, implying that APIs/CLIs/tool interfaces and identity/access controls become as important as the human UI.

Non-technical adoption hinges on new abstractions, not better prompts.

Sinofsky argues most workers can’t describe workflows as flowcharts, so today’s agent orchestration is a “rocket scientist” task that must collapse into simpler domain-level capabilities over time (like spreadsheets did).

“Computer use” is a pragmatic bridge, but it doesn’t solve enterprise control.

Casado notes agents increasingly act like humans operating existing software UIs, which speeds early deployment, but write-capable autonomy still collides with governance, auditability, and safety expectations in systems of record.

Integration gets easier—and that is exactly what scares CIOs/CFOs.

On-demand runtime integrations (agents stitching systems 27 and 38) can multiply capability, but leaders fear uncontrolled integrations will corrupt records, create security holes, and generate unmanageable operational complexity.

Treating agents like employees works personally, but not cleanly at scale.

Giving an agent its own accounts/credit card/phone number reduces key-sharing, yet Levie argues enterprises need oversight and rollback, agents have no privacy rights, and liability remains with the human/org—making “agent as a person” an imperfect model.

WORDS WORTH SAVING

5 quotes

The diffusion of AI capability is gonna take longer than people in Silicon Valley realize.

— Aaron Levie

It's ridiculous. It's just absurd- to think you're gonna vibe code your way to- like SAP.

— Steven Sinofsky

Algorithmic thinking- is really, really, really hard for the vast majority of people who have jobs.

— Steven Sinofsky

'Cause you put people creating new integrations- and you just say, "Please break my system of record."

— Steven Sinofsky

Everybody is trying to figure out the economics of all of this- when they're off by at least an order of magnitude on how big the opportunity is.

— Steven Sinofsky

Building software for agents vs. humansAlgorithmic thinking and the next abstraction layerComputer-use agents vs. API/code-first agentsIntegration on demand and systems of record riskEnterprise governance: RBAC, identity, oversight, liabilityPrompt injection, confidentiality, and data exfiltration threatsEconomics: tokens, usage-based pricing, and underestimated TAM

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome