Why Architecture Determines the Future of AI Growth

Enterprises today face a structural information problem. The majority of new business data is unstructured — documents, emails, contracts, and multimedia scattered across fragmented systems.

IDC estimates that 90% of the data generated by organizations in 2022 was unstructured. A study by Bar‑Ilan University researchers shows that employees lose meaningful portions of their workweek to information searches, with more than 10% spending over a full workday each week on avoidable retrieval waste.

As this sprawl grows, organizations struggle to maintain consistent taxonomies, permissions, and governance. NIST SP 1800‑39 exposes how these inconsistencies break Zero Trust Architecture and prevent AI models from being trained on authoritative data. Slowing decisions, increasing exposure, and limiting the value enterprises can extract from their information.

The bottleneck is no longer data scarcity but data sprawl. Generative AI raises the stakes: its value depends entirely on the quality, structure, and accessibility of enterprise data — and most organizations are not architected to support it.

In a recent interview, Daniel Fagella, founder of Emerj Artificial Intelligence Research, sat with Aaron Levie, CEO of Box, to discuss how enterprises can re-architect their systems to survive the AI race.

Their conversation highlights two critical strategic insights:​

• Modular architecture as an AI advantage: A services‑oriented platform lets enterprises plug in new AI models quickly while preserving governance and avoiding vendor lock‑in.
• AI‑ready data organization as a prerequisite for value: Normalizing unstructured data and permissions across systems enables AI to retrieve authoritative information and support document‑heavy workflows reliably.

Listen to the full episode below:​

​

Guest: Aaron Levie, CEO, Box ​

Expertise: Enterprise Architecture, Data Governance, SaaS Innovation

Brief Recognition: Aaron Levie is the co‑founder and CEO of Box, a leading enterprise cloud content platform. He has guided Box’s transformation into a cornerstone of secure, AI‑driven content management and is widely recognized for his leadership in SaaS innovation and enterprise AI strategy.

Modular Architecture as an AI Advantage

Levie frames the model ecosystem as a moving target, one that shifts faster than enterprise systems can adapt. In his view, the only sustainable response is to design for change itself. That means treating AI models as interchangeable components rather than architectural commitments.

Box’s platform reflects that philosophy. Instead of binding workflows, permissions, and data flows to a single model, Box routes everything through a middleware layer that abstracts the model away entirely. The result is an environment where adopting a new model is an operational choice, not a multi‑quarter project.

“You can’t predict which model will be best six months from now. So the architecture has to assume you’ll change your mind. We built the system so we can plug in a new model the same day it becomes available — without rewriting everything underneath.”

— Aaron Levie, CEO of Box

This approach shifts the economics of AI adoption:

• Model evaluation becomes continuous rather than episodic.
• Governance remains stable even as the underlying reasoning engine changes,and
• Vendor lock‑in becomes a strategic option rather than an inevitability.

For Levie, modularity isn’t a technical preference — it’s the only way to keep pace with an ecosystem defined by rapid, uneven innovation.

AI‑Ready Data Organization as a Prerequisite for Value

Levie’s view is that most enterprise AI failures trace back to one issue: the model is reasoning over disorganized, permission‑inconsistent content. When documents, messages, and records live across dozens of systems with different structures and access rules, AI cannot determine what is authoritative or what a user is allowed to see, and the output becomes unreliable.

Box addresses this by creating a federated content layer that normalizes metadata, permissions, and relationships across repositories. The goal is not to centralize files, but to give AI a consistent, permission‑aware map of the enterprise’s unstructured data.

As Levie explains:

“Getting all your content into one place isn’t the hard part. The hard part is giving the AI enough context to understand what that content means — who owns it, how it relates to other documents, and what the user is actually allowed to see. Without that structure, the model isn’t being intelligent; it’s taking a guess.”

— Aaron Levie, CEO of Box

A unified content layer provides AI with the context needed for document‑heavy workflows. In practice, it enables:

• Accurate retrieval: The model surfaces the correct version of a document or policy.
• Permission‑aligned responses: Access controls remain intact across every query.
• Cross‑repository visibility: Teams can work across systems without migrating data.
• Lower error rates: Consistent metadata reduces hallucinations and incomplete answers.
• Faster deployment: New AI use cases plug into the existing content layer without re‑indexing.

For Levie, data readiness is the foundation of trustworthy AI. Without a consistent, permission‑aware content layer, enterprises are effectively asking models to reason over noise. With it, AI can support high‑stakes workflows with far greater reliability.

.

.