How to Manage AI Agents as Digital Employees

TL;DR

Treat AI as a digital employee that can analyze documents, use tools, and execute long-running tasks under human direction. Companies can gain more output by assigning AI well-scoped work, preserving security guardrails, and selecting models for particular domains, while employees benefit from learning to manage context and incorporate AI throughout their daily routines.

Transcript

[music] [music] [music] [music] [music] [music] [music] [music] [music] >> Okay, we have Aaron Levie, CEO of Box on the show. Thank you so much for coming all the way from the Bay Area. >> Thank you for having me. Appreciate it. >> Yeah. I heard you just landed. >> just a couple hours ago. Yes. >> A couple hours ago. Oh my god. Oh, so busy. >> Yeah... Read More

Key Insights

• Recent frontier AI models are strong at coding, advanced reasoning, tool use, computer operation, and specialized work in legal services, health care, life sciences, and finance. Their capabilities now extend beyond answering simple questions to executing long-running enterprise tasks.
• Most leading models deliver comparable performance for 80–90% of use cases. Meaningful differences appear in deeper domain-specific tasks, such as front-end development, back-end development, or legal work, because training data, post-training methods, optimization choices, time, and available compute vary.
• AI model improvement is not slowing down, according to Levie. Talent, capital, computing capacity, purchased datasets, data-labeling resources, and improved training methods continue to produce capability gains, with each new model version often showing measurable improvements over its predecessor.
• Software development gains depend on the code layer and its risk profile. Some Box teams ship two to three times faster with AI, while other teams improve by roughly 20%, particularly where sensitive back-end logic requires extensive human review.
• Front-end development is better suited to rapid agentic coding because interface improvements generally pose more contained risks. Letting an agent make broad changes across millions of back-end code lines could affect business logic, expose data, or alter permission structures.
• Enterprise AI creates value by processing unstructured information such as contracts, marketing assets, invoices, and financial documents. Agents can read these materials, extract key data, identify insights, and support workflows including client onboarding, insurance claims, contract intelligence, and due diligence.
• The SaaS apocalypse is mostly an overreaction, but interface-only software serving individual users faces genuine pressure. An agent can answer questions, analyze information, and generate interfaces, reducing the distinct value of products concentrated mainly at the presentation layer.
• Systems of record may gain importance in the AI era because agents need secure operational backbones. These systems protect critical organizational data, define business processes, preserve workflows, and prevent agents from freely accessing information or executing unauthorized decisions and tasks.

Questions & Answers

Q: What can the latest frontier AI models do for businesses?

The latest frontier models can perform coding, advanced reasoning, tool use, computer operation, and specialized tasks in areas including legal services, health care, life sciences, and finance. Instead of merely retrieving a simple fact, they can analyze large collections of documents, extract important details, combine insights, and execute longer-running tasks within enterprise workflows.

Q: How different are the leading AI models in practical use?

The leading models are close enough that Levie expects similar performance across 80–90% of use cases. Differences become clearer in demanding or specialized work. One model may be stronger in front-end development, another in back-end development, and another in legal tasks because each provider uses different training data, post-training methods, optimization priorities, time, and computing resources.

Q: Why are AI model capabilities continuing to improve?

AI capabilities continue improving because substantial talent and capital are focused on model development. The companies building these systems have large amounts of computing capacity, can purchase major datasets from data-labeling vendors, and are refining their training methods. Levie says the pace has not slowed, and industry discussion points toward continued breakthroughs throughout the year.

Q: How much faster can AI make software development?

The productivity gain varies by team, software layer, and risk. Levie says some Box teams may ship two to three times faster, while others move only about 20% faster. He expects that within the next year or two, most companies should be capable of delivering two, three, or five times more software than before AI.

Q: When is agentic coding safest to use?

Agentic coding is safest when the proposed changes have a contained risk profile, such as improving a front-end feature without altering core business logic. Greater caution is necessary in back-end systems, where unreviewed changes across millions of code lines could leak data, modify permissions, or disrupt essential logic. Human review therefore remains especially important for sensitive code.

Q: How can AI agents process enterprise documents?

AI agents can read large collections of unstructured materials, including contracts, invoices, financial documents, and marketing assets. They can extract structured data, identify important details, and generate intelligence from the information. Box sees applications in client onboarding, insurance claim processing, contract analysis, and document-based due diligence, where faster review can improve workflow efficiency.

Q: Will AI agents make SaaS products obsolete?

Levie considers a broad SaaS apocalypse an overreaction, although certain products face real risk. Software whose value is mainly a single-user interface may be vulnerable because agents can answer questions, analyze data, and generate interfaces efficiently. Software that stores critical records and governs business processes may instead become more valuable as the operational foundation for agents.

Q: Why do AI agents need systems of record and guardrails?

AI agents cannot safely access any information, make unrestricted decisions, or execute every available task inside an organization. Systems of record provide a controlled backbone that protects data, establishes permissions, and preserves required workflows. These guardrails let organizations apply agent capabilities while reducing the security and operational risks associated with unconstrained access and action.

Summary & Key Takeaways

• Aaron Levie says recent frontier models have become highly capable at coding, advanced reasoning, tool use, computer operation, and specialized work in fields such as legal services, health care, life sciences, and finance. Their defining advance is the ability to process many documents and complete longer, more complicated enterprise tasks.

• AI-assisted software development produces uneven gains because effectiveness depends on the task and its risk profile. Some Box teams ship two to three times faster, while others improve by about 20%. Front-end changes offer safer opportunities for agentic coding than sensitive back-end work involving business logic, data, and permissions.

• Levie rejects the broad idea of a SaaS apocalypse. Interface-only, single-user products face meaningful risk, but systems of record may become more important because agents require protected data, defined workflows, and operational guardrails. He expects people to manage AI subordinates, expanding what employees and even small organizations can accomplish.