Why AI Agents and Unstructured Data Bring Opportunities for Startups
This article uses statements by Box co-founder and CEO Aaron Levie as a starting point to analyze "why startups win in the AI era" across four axes: corporate reality, data structure, business models, and product strategy. In particular, it substantiates the new market created by unstructured data and agents, as well as the shift in SaaS from seat-based pricing to outcome- or consumption-based pricing, using the latest examples and literature.
1. "AI Hype vs. Corporate Reality"—What is really changing?
Levie points out the reality that much of a company's time is spent on "necessary but non-strategic work." If AI agents can replace and automate these tasks, people can shift to areas of differentiation such as customer understanding, new feature creation, and proactive support. For startups, the opportunity lies in being able to tackle this "mountain of untouched work" (tasks that were left alone because they were not economically rational for humans to perform) with software for the first time. Box is doing exactly this by launching features that handle the extraction, automation, and governance of unstructured data (Box Extract, Automate, and Shield Pro) one after another, lowering the hurdles for implementation in the field.
1-1. No need to "persuade" infrastructure
In the early days of the cloud, it was necessary to persuade people about whether it was "truly safe," but with AI, social consensus has come first, and management has already experienced the "potential for use." The barriers have shifted to the implementation process, such as reliability, data connectivity, and operational systems. The trends at BoxWorks are precisely aimed at supporting this implementation phase.
2. Unstructured data is the biggest gold mine
80-90% of corporate data is considered unstructured (documents, emails, images, audio, slides, etc.). Conventionally, because systems were optimized for database-centric (structured) data, this massive asset has remained dormant as "Dark Data." Generative AI and agents enable document reading, summarization, extraction, verification, and workflow execution, transforming value from "things that couldn't be read can now be read" to "things that couldn't move can now move."
2-1. The "knowledge" layer that Box is targeting
Box has set its sights on "Intelligent Content Management," steering toward turning "unstructured knowledge within the enterprise" into assets using tools like AI Studio, which balances model selectivity with security. This is a design philosophy that integrates search, summarization, extraction, automation, and protection into a single, seamless flow.
2-2. A note on system design: Context degradation
In an interview with Business Insider, Levie pointed out "context rot," where performance degrades because a single large agent holds too much context. He argues that an architecture that combines purpose-specific sub-agents is effective.
3. Business model transformation—From seat-based to "workload" pricing
SaaS has long been capped by "how many people use it (seats)." However, because agents embody "labor," customers are more amenable to designs where they pay based on the amount of work (outcomes/consumption), such as "number of contract reviews," "number of inquiries resolved," or "processing volume." Levie himself has publicly stated that he is building hybrid models of seats plus consumption or outcome-based pricing. The key is the idea of charging a premium for "value-added software"—such as workflows, security, integration, and quality assurance—relative to the underlying token costs.
3-1. How to ensure revenue sustainability
Consumption-based pricing is highly volatile. The standard practice is to combine a base subscription (platform usage fee) with processing volume charges to achieve both recurring revenue and expanded usage. This mirrors the history of the cloud, where "costs go down, but gross margins are protected by software value."
4. A blueprint for startups to "defeat large enterprises"
4-1. Choosing where to fight with core/context thinking
In the context of Geoffrey Moore's "Crossing the Chasm," companies distinguish between core (the source of differentiation) and context (non-differentiating periphery). Since most companies do not want to build context areas like HR, accounting, or document processing in-house, excellent agent SaaS solutions are easily adopted there. Conversely, do not collide head-on with the opponent's core.
4-2. Find the unexplored "noun x verb"
Levie believes that obvious updates like "AI in CRM" will be captured by major players, and that there is a huge void in areas where agents can solve "work that could not be solved by software" for the first time. Hiring reviews, contract scrutiny, quality audits, regulatory reporting, bulk localization into multiple languages—the starting point is whether you can automatically run tasks that "could not be tested due to labor costs." Recent interviews and lectures around Box also depict a future where "hundreds of specialized agents stand side-by-side within a company."
4-3. Principles of Implementation Architecture
Organize unstructured data (classification, permissions, and provenance management) first. Governance determines AI quality.
Divide into sub-agents to control context window size. Maintain evaluation and audit metrics at the task granularity level.
Safety and superior experience (audit logs, rework detection, human-in-the-loop) build value in the "top layer" and avoid downward price pressure.
Conclusion
Use agents to handle "work that was previously unprofitable or impossible for humans to touch."
This simple principle provides leverage for startups in the AI era. The playing field is outside of structured data—it is about whether you can provide the end-to-end experience of "reading, acting on, and protecting" a company's unstructured knowledge as a safe and auditable product. Box's trends and Levy's guidelines are rooted in the reality of implementation. The window is not long. In the next 24 to 36 months, those who invent uncharted noun-verb combinations will redraw the map of SaaS.