How Are Enterprises Bridging AI Workflow Gaps?
Aaron Levie says enterprises are redesigning workflows, data access and software architecture as AI agents move from pilots to production.
Enterprise adoption of AI agents is exposing challenges that go well beyond model performance, according to Box Co-founder and CEO Aaron Levie. As organisations expand the use of AI across internal operations, they are grappling with workflow redesign, data access, security, and uneven AI spending.
In a post on X last week, Levie said one of the biggest barriers to deploying AI agents is change management. Most business processes, he said, were not designed to work with autonomous agents and require significant changes across technology, data and human workflows.
He added that enterprises are prioritising efforts to make both structured and unstructured data accessible to AI agents, recognising that data quality and accessibility remain critical to successful deployments.
Levie said many organisations are also embedding engineers directly within business teams to accelerate AI adoption. The approach, similar to having an internal forward-deployed engineer, helps companies identify workflow bottlenecks early and avoid lengthy, unsuccessful AI experiments.
As AI agents begin working across multiple departments, enterprises are encountering new identity and permissions challenges.
Levie said agents require their own roles and access privileges because they often need information spanning multiple business functions. Managing those permissions securely is difficult, he noted, particularly because agents cannot independently secure or govern themselves.
He also said Anthropic's recently introduced Mythos models are exposing more sophisticated security risks by enabling attackers to chain together multiple vulnerabilities, prompting organisations to build larger patch backlogs.
AI spending varies widely across enterprises
Levie highlighted significant differences in enterprise AI budgets.
According to him, some organisations cap monthly AI spending for software developers at $1,000, while others allow spending up to $5,000, treating that threshold only as a notification rather than a hard limit. Investment in AI tools for non-coding knowledge work, however, remains considerably lower.
On infrastructure, Levie said more enterprises are building internal routing systems that dynamically direct workloads between frontier and lower-cost AI models to balance performance and cost.
He added that while interest in open-weight models continues to grow, most enterprise deployments remain experimental rather than production-scale.
Levie also noted that enterprises expect software to support headless, agent-friendly interfaces, enabling AI agents to interact directly with applications rather than through traditional user interfaces. Vendors failing to support this shift risk slowing AI adoption.
>
>
Key Takeaways
- Address workflow redesign to facilitate the integration of AI agents into enterprise operations.
- Enhance data accessibility and quality to ensure successful AI deployments across the organization.
- Embed engineers within business teams to quickly identify and resolve AI workflow bottlenecks.
- Manage identity and permissions for AI agents to secure sensitive information across multiple departments.
- Recognize significant disparities in AI spending among enterprises to optimize resource allocation.