Box’s CEO reminds us that 88% of the economy runs on different rules, different budgets, and different constraints than the VC world…
If you want to understand why most enterprises are still stuck in AI pilot purgatory, you can do a lot worse than listen to Aaron Levie. The Box CEO recently delivered one of the most lucid, unsparing conversations I’ve heard on what agents actually mean for large organizations, and almost none of it matched the breathless narrative coming out of the labs and the VC echo chamber. He had to practically shake his interviewer, Harry Stebbings, to get some of it across.
So here is what enterprise leaders really need to hear about agentic AI…
Stop measuring AI’s impact on the 12% of the economy that already has it. The other 88% is where the transformation happens.
“Everybody is so myopic about this,” Levie said, and his point is that the tech industry represents 10-12% of GDP, while the banks, pharma companies, manufacturers, and industrial firms that make up the other 88% are not sitting on engineering surplus but are chronically under-resourced. AI-assisted development does not threaten Silicon Valley jobs but opens engineering-grade capability to the rest of the economy for the first time, meaning CS graduates no longer go to Google but to John Deere or Eli Lilly, and the profession expands rather than contracts.
The same logic applies to legal, because AI-generating contracts do not reduce the workload but flood the system with more work requiring qualified human review, and the bottleneck was never generation but courts, regulators, and approvals, which means there will be more lawyers in five years, not fewer. The real casualty is the junior pipeline, because when AI does the apprenticeship work, you lose the mechanism by which the next generation learns the craft, and every bank and law firm that built its talent model around that apprenticeship structure is facing that reckoning now.
The workflow needs to be redesigned for agents, not for people
Every large enterprise that wants meaningful returns from AI automation has to go through genuine change management. Data is fragmented across decades of employee-brought tooling, legacy document management systems, and network file shares. Agents will find the wrong contract, the wrong document, the wrong piece of customer data, not because the models are bad, but because the underlying data state was never organized with machine consumption in mind. People could navigate ambiguity by knowing where to look, which agents cannot. For example, a healthcare organization is using agents to automate patient referrals, which sounds like a win, but if the next available specialist appointment is still 18 months out, nothing has changed for the patient. The bottleneck was never the paperwork, but was the shortage of doctors as automation makes the queue move faster, but does not make the queue go somewhere. Every enterprise has a stack of constraints like this sitting behind the process they just automated, and agents surface them faster than any previous technology. You can automate the intake, but you cannot automate the shortage of doctors. Automation does not eliminate constraints, but reveals them faster.
Professional services firms are not being replaced by AI. They are about to get busier than they have ever been
Agentic is creating years of structural work for transformation partners. If a Fortune 500 company wants an agent to identify contract renewal risk, it might encounter ten systems holding contracts, half incompatible with modern API architecture, which means getting that data organized and context-aware is not a model problem but a decade of work for Accenture, Cognizant, or the next generation of specialized implementation partners. And there is one more dynamic that guarantees the services layer persists, which is that someone has to be accountable when it goes wrong, and as Levie put it, “you are not going to be able to blame Anthropic,” which means the moment you need liability you need ownership, and the moment you need ownership you need people, a structure that does not dissolve with better models.
Hire the agent operator now. Up to a million of these roles are coming and enterprises that wait will not catch up.
Levie floated a job title the industry is still workshopping, which is the agent operator, someone technical enough to understand MCPs, CLIs, and agent configuration files, but also fluent enough in business process to go into a marketing team or legal operations function and translate AI capability into workflow leverage, redesigning the process around the agent rather than the person who used to do it.
The second a new model drops your workflow probably breaks because the prompting syntax changes, which means this is not an IT role but a standing discipline sitting at the intersection of technical acumen and operational transformation. Levie puts the job creation at 500,000 to a million roles, with enterprises that build this capability internally seeing their AI returns grow and accelerate over time, because each workflow redesigned makes the next one faster and cheaper, while those that treat agent deployment as a one-time project will find themselves stuck in an endless cycle of pilots that nobody is equipped to sustain or scale.
Move AI spend out of the IT budget and into operating expense. Your CIO’s ceiling will never fund the transformation your business needs.
For the first time, technology providers can walk past the CIO and approach the line of business directly, offering a productivity return that justifies a share of operating budget, which is new, given that IT budgets have always been the ceiling. Levie’s view is that this probably doubles global enterprise technology spend by unlocking a new category of investment tied directly to workflow productivity. But what Silicon Valley keeps ignoring is that enterprises have EPS commitments and annual planning cycles, which means every dollar of AI spend has to justify itself against earnings targets.
So the practical discipline is to identify the 5-10% of your workforce doing the highest-value work, give them the best models with no capacity constraints, apply more efficient and cheaper models to the next tier of use cases, and let general productivity run on commodity AI, because treating every employee as equally deserving of frontier model access sounds inclusive but blows your budget and your earnings commitments in the same quarter.
Avoid single-vendor dependency in your AI stack. The enterprise AI platform race will end like cloud: multiple winners, massive market.
When asked to call the race between OpenAI and Anthropic, Levie gave the most pragmatic answer available, which is that in 2010 AWS had $500 million in revenue, Azure had just launched, and GCP had barely launched, yet today it is a multi-hundred-billion-dollar ecosystem where all three won, which is likely how AI ends up, with enterprises running multi-model stacks because no serious organization will accept single-vendor dependency in infrastructure this consequential.
Take your AI security risk seriously. Agents create attack surface faster than any human team can review it.
The moment you start generating code with AI you produce it at a volume that outstrips any human team’s ability to review it, with every feature shipped representing a potential vulnerability the agent introduced without anyone catching it. Meanwhile, the offensive side has exactly the same advantage and can scan for vulnerabilities at machine speed, leaving you with more surface area and faster attackers simultaneously. This is why Levie concluded that “agents are the solution to the problem that agents have caused,” a recursive dynamic that makes agentic security one of the few genuinely new infrastructure categories in this cycle.
Audit your platform stack for API depth, not UI richness. Agents do not care about your human interface… they care about your business logic.
Software built around dense interfaces designed for human clicking is genuinely at risk, because if agents are doing the work that humans used to do, that UI value simply evaporates. But software that embeds proprietary business logic in its API layer is a completely different story, because an ERP encoding two decades of supply chain rules is not just a database, and a platform enforcing FINRA retention at the API level is not just a file store. Agents actually need those capabilities more than humans did, which means the right question when evaluating any platform in your AI stack is not how good the interface is, but how deep the business logic runs and whether it was built to govern agent behavior rather than just enable human navigation.
Bottom line: Stop waiting for the technology to eliminate the hard parts. It will not.
As Levie put it, “we haven’t removed humans from the loop, we’ve just changed where they enter the loop,” and that single sentence should be on the wall of every enterprise transformation team. The real constraint is not model quality but workflow redesign, data readiness, change management, and accountability for when something goes wrong, and that work does not get automated away but gets more important with every agent deployed.
Enterprises that understand this will build a durable advantage, while those still waiting for the technology to eliminate the hard parts will still be waiting in five years as their competitors do the structural work that actually sticks.
Posted in : Agentic AI, AGI, AI layoffs, Anthropic, Artificial Intelligence, Automation, Google, OpenAI
