Box CEO Aaron Levie on AI’s ‘era of context’ | TechCrunch

Box launched the developer conference box office on Thursday by announcing a new set of AI capabilities that will build proxy AI models on the backbone of the company’s products.

This is more than the conference’s product announcements, reflecting the company’s growing speed of AI development: Box launched an AI studio last year, followed by a new set of data-pull agents In Februaryand other search and in-depth research in May.

Now, the company is launching a new system called Box automation This is an operating system for AI agents that divides workflows into different market segments and can be used as needed for AI enhancement.

I talked with CEO Aaron Levie about the company’s approach to AI and the dangerous work of competing with foundation modeling companies. Not surprisingly, he is very optimistic about the possibility of AI agents in the modern workplace, but he is also clearly invisible about the limitations of the current model and how to use existing technology to manage these limitations.

This interview has been edited for length and clarity.

TechCrunch: You announced a bunch of AI products today, so I want to start asking about the big picture. Why build an AI proxy in cloud content management services?

Aaron Levi: So what we think about all day – what is our focus – how much work happens due to AI. The vast majority of the impact now is on workflows involving unstructured data. We have been able to automate any content that processes structured data in a database. If you consider CRM systems, ERP systems, HR systems, then we have automation in this field for many years. But where we never have automation is anything that touches unstructured data.

Consider any type of legal review process, any form of marketing asset management process, any type of M&A transaction review – all of these workflows involve a lot of unstructured data. People have to view this data, update it, make decisions, etc. We have never been able to bring much automation to these workflows. We have been able to describe them in software, but computers are not good enough to read documents or view marketing assets.

So, for us, what AI proxy means is that we can actually take advantage of all this unstructured data for the first time.

TC: What about the risks of deploying agents in a business environment? Some of your customers have to worry about deploying something like this on sensitive data.

Levie: What we see from customers is that they want to know that every time the workflow is run, the agent will execute more or less at the same moment in the workflow and nothing will go away. You don’t want the agent to make some more complex mistakes, and after they do the first few submissions, they start running wildly.

It becomes very important to have the right dividing point, the agent starts and ends the rest of the system. For each workflow, there is a question about what the deterministic guardrail needs and what can be completely proxy and non-deterministic.

What you can do with Box Automate is to determine how much work you want each agent to do before handing it over to other agents. So you might have a commit agent separate from the review agent, and so on. It allows you to deploy AI agents at scale essentially any workflow or business process in your organization.

TC: What kind of problems do you prevent by separate workflows?

Levie: Even in state-of-the-art fully proxy systems like Claude Code (such as Claude Code), we have seen some limitations. At some point in the task, the model exhausted the context window chamber and continued to make the right decisions. AI doesn’t have free lunch now. After any task in the business, you cannot have only one long-term proxy with an infinite context window. So you have to break down the workflow and use a child agent.

I think we are in the contextual era in AI. What AI models and agents need is the context, and the context they need to solve sits in your unstructured data. So our entire system really aims to figure out the context you can give AI agents to make sure they execute as efficiently as possible.

TC: The industry is more debated in the industry than smaller and more reliable models. Will this put you on the side of the smaller model?

Levie: I should probably clarify: There is nothing to stop a task from being arbitrarily long or complex with our system. What we want to do is create the right guardrail so that you decide which agent you want to have the task.

We have no special philosophy about where people should be in that continuum. We are just trying to design a future architecture. We have designed this in a way that as the model improves and the proxy functionality improves, you will get all these benefits directly in our platform.

TC: Another problem is data control. Since the model has been trained with such a large amount of data, it is indeed a concern that sensitive data will be refluxed or abused. How did this join?

Levie: This is where many AI deployments go wrong. People think, “Hey, it’s easy. I’m going to do AI model access to all the unstructured data, which will answer questions for people.” Then it starts giving you answers to data that is not accessible, or you shouldn’t. You need a very powerful layer to handle access controls, data security, permissions, data governance, compliance, everything.

So we benefited from decades spent in a system that basically solved this exact problem: How do you make sure that only the right people can access every data in the business? So when the agent answers a question, you know with certainty that it cannot borrow any data that the person cannot access. This is just something fundamentally built into our system.

TC: Earlier this week, Anthropic released a new feature for uploading files directly to Claude.ai. There is a long way from file management of boxes, but you have to consider possible competition from the underlying model company. How do you treat it strategically?

Levie: So if you consider what your business needs when deploying AI at scale, you need security, permissions, and control. They need a user interface, a strong API, and an AI model needs to be chosen because one day an AI model gives them a better use case than another, but may change and don’t want to lock it in a specific platform.

So we build a system that effectively has all these features. We are performing storage, security, permissions, vector embedding, and connecting to each leading AI model.