Box Architect: AI That Designs Your Enterprise Workflows for You
Box CEO Aaron Levie explains Box Architect, an AI system that autonomously designs agentic workflows on top of Box's unstructured data layer.

What is Box Architect?
Box Architect is an autonomous AI system, announced by Box at BoxWorks, that designs, builds, and runs enterprise workflows and agents on top of Box’s platform. Instead of a company hiring consultants or assembling a stack of separate tools to build something like a contract management system or an invoice processing pipeline, a user can describe the process they want and let Architect design the full solution, using building blocks already inside Box such as its workflow engine, no-code app builder, and document management system. Box CEO Aaron Levie describes it as the product of letting frontier AI models, not just human engineers, design the actual business process rather than simply automating a slice of it.
TL;DR
- Box Architect autonomously designs full enterprise workflows (contract management, supply chain review, invoice processing) using Box’s existing workflow tools, app builder, and document system as building blocks.
- Levie says an internal prototype built roughly six to nine months before launch did in 30 seconds what customers normally spend two days whiteboarding and architecting by hand.
- Box frames the future of enterprise software around two layers: a huge population of autonomous agents and a ground truth context layer that governs what data those agents can access.
- Levie says Box deliberately avoids building anything that improving frontier models would soon make obsolete, a stance he jokingly calls being “bitter lesson pilled.”
- Box plans to have Architect eventually optimize workflows over time based on real failures and outcomes, not just design them once.
- The company says it remains partner-centric and API-first, aiming to let unstructured data flow into other systems (Salesforce, SAP, ChatGPT, Claude) rather than locking customers into a single interface.
- Levie acknowledges Box has thrown away unreleased capabilities multiple times because newer models made the scaffolding they built around them unnecessary.
Why is Box building this now?
Levie points to two things converging this year: tool use and computer use in frontier models got dramatically better, and benchmarks measuring real-world business tasks (he references GDPval) started showing models performing well on enterprise workflow-style problems. Box had already built the component pieces, a workflow tool, a no-code app builder, a document management core, so testing whether a model could stitch those into a working solution was a short experiment rather than a new product line. According to Levie, an engineer ran that experiment with the era’s leading model and the result did in half a minute what customers typically spend two days designing on a whiteboard.
That result, repeated as models like Opus, Astra, and others improved, convinced Box leadership that the design step itself, not just the execution of a workflow, could be handed to AI. Levie frames this as a philosophical shift: rather than automating today’s business process exactly as it exists and dropping an agent into it, an AI system might design a better process from scratch, because it has broader pattern knowledge of how similar workflows get built across many organizations.
How does the context layer fit into this?
Levie’s framing for where Box sits echoes a prediction he attributes to Microsoft CEO Satya Nadella: software splits into an agent layer and a ground truth context layer. Levie adds a number from Nvidia’s Jensen Huang to illustrate scale, the idea that organizations could eventually have 10 to 100 times more agents than employees, putting enterprises in the range of tens of billions of agents moving through their systems.
For that many agents to operate safely, something has to control what data each agent can touch: which CRM records, which ERP data, which files. Levie argues this can’t just be dumped into a generic database. It requires access controls, metadata, and governance so an agent knows not just what information exists but where to route a query and what it’s allowed to see. Box’s pitch is that it already functions as this kind of system of record for unstructured content (contracts, documents, supply chain files), and Architect is a way of extending that system upward into the design of the workflows that consume that content.
What decides whether Box builds a feature or leaves it to a model?
Levie says Box tries to “ride the curve of intelligence” as far as possible, meaning the company avoids building anything that a better frontier model would simply make redundant. The exception is scaffolding, temporary structure needed because current models can’t yet do something end to end. He describes Box as heavily influenced by the “bitter lesson,” the idea (originating in AI research) that general methods leveraging more compute and better models tend to beat hand-engineered, domain-specific approaches over time. Levie jokes he’s “in the upper limit” of being persuaded by that argument.
Practically, this means Box focuses on what it calls timeless guardrails: data governance, security, high-throughput infrastructure, and schemas that let an agent interact with Box’s underlying objects easily. Those pieces, Levie argues, stay valuable even if the AI models doing the reasoning get a thousand times smarter. He compares the posture to Amazon’s early internal decision to go API-first with what became AWS: build for an unknown number of future consumers of your service, including ones that never touch a visible interface.
Is Box Architect worth watching if you’re not a Box customer?
Even outside Box’s own product, Architect is a useful signal of where enterprise workflow tooling is heading. Levie points to comparable moves elsewhere: customer support platforms like Decagon, Sierra, and Fin have each built agents whose job is to help a human operator design the support agent that will talk to customers, rather than just running a fixed script. Architect applies the same idea more broadly, to any internal process a business wants to automate, not just customer support.
For people building AI products, the relevant takeaway isn’t the specific Box feature but the pattern: companies with deep domain data (documents, support tickets, workflow history) are starting to let frontier models design the automation on top of that data, instead of only executing predefined steps. Levie frames this as inevitable given how fast tool use and computer use have improved in recent models, and Box’s bet is that owning the unstructured data and governance layer gives it a durable position even as the agents themselves get swapped out for whatever the newest model happens to be.
How often does Box scrap work because of model improvements?
Levie says it happens constantly. He describes situations where Box built unreleased capabilities that were made obsolete before shipping, simply because a newer model came out that could do the same job without the custom scaffolding Box had built around an older one. He frames this as an accepted cost of trying to stay at the frontier rather than over-investing in solutions that are really just patches for a model’s current limitations. The implication for Architect itself is that its underlying design logic is expected to keep changing as “whatever the leading model is at a given time” changes, with Box swapping in newer models (he names Astra, Opus, and Fable as recent examples) as they become available.
Frequently Asked Questions
What problem does Box Architect solve?
It replaces the manual process of designing an enterprise workflow, such as contract management or invoice processing, which Levie says customers often spend around two days whiteboarding and architecting by hand, with an AI system that designs the solution using Box’s existing tools.
Does Box Architect require other Box products?
Yes. Architect works by assembling building blocks that already exist inside Box, including its workflow engine, no-code app builder, and core document management system, into a complete workflow or application.
Will Box Architect keep improving on its own?
Levie says the plan is for Architect to eventually optimize workflows over time by analyzing failures and outcomes, moving from a one-time design tool toward an ongoing autonomous system for a given business process.
Is Box Architect locked to one AI model?
No. Levie describes it as using “frontier AI,” whichever leading model is available at a given time, and names several (Astra, Opus, Fable) as examples used during development, implying Box swaps models as better ones emerge.

Nothing gets lost.
How does Box Architect relate to Box’s data governance focus?
Levie frames Architect as an extension of Box’s role as a governed system of record for unstructured data. The same access controls and metadata that govern what human users can see are meant to govern what the agents designed by Architect can access and act on.



