Ben Kus

Ben Kus is the chief technology officer of Box, leading its technology and AI strategy for turning enterprise documents into useful information without compromising security. His work brings coordinated AI agents to practical problems such as extracting contract terms, researching private company files, and connecting applications to governed enterprise content.

Kus studied computer science at the University of California, Berkeley, and was an early employee at BigFix. After IBM acquired the company, he became chief architect of mobile security. He subsequently co-founded Subspace, an enterprise security startup founded in 2013 that developed a containerized browser for accessing corporate information securely across personal and company-owned devices. Box acquired Subspace in 2015, bringing Kus and his fellow founders into the company. He later served as Box’s vice president of product management before becoming CTO.

• Agentic document extraction: Simple prompts can extract information from straightforward documents, but long contracts, unreliable scans, multilingual material, and interdependent fields overwhelm single-pass systems. Kus favors workflows that group related information, subdivide difficult tasks, compare OCR output with page images, and check results across multiple models. Keeping contracting parties together with their addresses, for example, prevents technically plausible but mismatched answers. His agentic-platform architecture treats extraction as coordinated reasoning instead of a single model call.

• Corrective AI evaluation: An LLM-as-a-judge should improve unreliable answers, not simply warn users about them. Kus combines model-generated criticism, iterative corrections, independent checks across providers, standard evaluation sets, deliberately difficult challenge cases, and customer feedback.

• Model flexibility over repeated fine-tuning: Maintaining custom-tuned versions of several rapidly changing model families creates substantial operational overhead. Kus instead favors adaptable prompts, prompt caching, and modular agentic workflows, allowing teams to replace models or repair individual processing steps without rebuilding an entire system.

• Governed access to enterprise content: Useful enterprise AI depends on finding relevant organizational information while respecting existing permissions and security boundaries. Kus applies that principle to document question answering, research across private files, agent APIs, and the Box Model Context Protocol server. He also teaches developers to build MCP-connected applications and coordinated agents.

Kus separates the intelligence required to interpret an individual document from the distributed infrastructure required to process content at enterprise scale. That architectural distinction lets teams refine verification, planning, and output quality independently while extending the same foundation across extraction, research, and other business workflows.