Is SaaS Dead? Box CEO Aaron Levie on What AI Changes
Box's CEO argues the next software layer is not another app but a context layer: governed access to enterprise data that thousands of agents can query. Here is what that means for tokens, pricing, and open models.
Every few years the enterprise software industry is told it is about to be replaced. This time the claim is that agents make applications unnecessary, and Aaron Levie — Box's co-founder and CEO, speaking with Matthew Berman — has an unusually specific answer: the application does not disappear, it becomes the thing that decides what the agents are allowed to know.
That reframing is the useful part of the interview. It explains why a document management company is building an autonomous workflow designer, why its CFO customers are nervous about token bills, and why Levie keeps describing himself as "bitter lesson pilled."
The premise: many more agents than people
Levie's working model of the next few years is that organisations will run on the order of ten to a hundred agents per employee — "30, 50 billion agents roaming around processing information." If that happens, the bottleneck stops being intelligence and becomes grounding.
Something has to keep the ground truth on the guardrails of what those agents are doing. Something has to say what data the agents can access.
That "something" is what Box calls the context layer: access controls, data governance, and metadata rich enough for an agent to know which system holds the answer. Levie is blunt that this is harder than it sounds. You cannot drop enterprise content into one giant blob and hope retrieval sorts it out; the separation between what different agents may touch has to be enforced deliberately.
In his framing, systems of record become systems of intelligence — still holding the data, still giving humans an interface, but now also brokering agent access.
Box Architect: letting the model design the workflow
The concrete announcement is Box Architect. An enterprise describes an outcome — contract management, invoice processing, supply chain document review — and the system designs the workflow or app, using whichever frontier model is best at the time, on top of Box's existing building blocks.
The origin story is the most telling detail. An engineer built a demo where a model designed a full solution in about thirty seconds, work that customers normally spend two days whiteboarding. Levie's conclusion was not "we should automate that step." It was that the models might soon design processes better than the people who currently document them — so Box should let them, rather than automating today's process exactly as it looks.
There is an honest caveat in the interview: they are not sure where to draw the line between what Box builds and what the model does. Their rule is to avoid building anything that a better model will simply erase.
"Bitter lesson pilled"
This is the strategic discipline underneath the product decisions, and it is the part most companies get wrong.
Levie says Box has scrapped unreleased capabilities because a new model made them pointless, and he expects that to keep happening:
You have to be very unsentimental and just be comfortable throwing away anything that you have created that the model is going to do better.
What survives, in his account, is the substrate a superintelligent system would still need: strong security, data governance, high-throughput infrastructure, and schemas clean enough that an agent can create objects in your system without a human translating.
That is also an argument about interfaces. If a growing share of interactions never show anyone a screen, a product has to produce value without one — the same logic as Amazon's API-first mandate, applied to agents instead of developers.
The token bill arrives
The interview's most practical section is about money. Levie's prediction is a thousand-fold increase in token consumption, not because models get more expensive but because agents attach to almost every digital interaction: a profile rebuilt before every sales call, an agent reviewing each login anomaly, agents checking each code push and each supplier document.
He is careful not to argue that raw token volume wins. Application matters more:
I'm not a fan of token maxing. Being judicious about where you apply force really does matter.
Both things are true at once — spending compute on the right problems becomes a competitive advantage, and spending it indiscriminately wastes money. Levie acknowledges the uncomfortable implication: if compute is the advantage, capital concentrates, and concentration tends to reduce innovation.
For pricing, he expects a hybrid. Seats remain because buyers want predictability; consumption grows because agent work is variable and cannot be forecast per task. "It'll just become a line item that you decide to increase or decrease," he says, like marketing spend. Until run rates are understood, CFOs will keep flinching.
The technical response is efficiency. Box announced a context retrieval engine built with Nvidia that pre-indexes unstructured content so agents fetch only the chunks they need. Less irrelevant context, fewer tokens, faster answers.
Open weights as the escape valve
Levie is committed to open weights, and his reason is political as much as technical: if a handful of labs gatekeep the tokens, they effectively decide which companies can compete.
He is also realistic about the business model. Training open models is expensive, so whoever releases weights needs to monetise inference — which is why Meta, Google, and Nvidia can play and most startups cannot. His observation about the gap is the reassuring part: if open weights stay three to six months behind the frontier, they still absorb the vast majority of use cases, because most work does not need the newest model.
And open and closed grow together rather than replacing each other. Cheap open models make it affordable to process every document or transaction, which in turn frees budget to use a frontier model as the planner above them. His analogy is databases: open source databases won, and Oracle still became a $300 billion company.
What this means if you build or buy software
Three takeaways worth keeping:
- Context is the product. Whatever you run, the value increasingly sits in governed, well-described access to your own data — not in the interface on top.
- Assume your scaffolding expires. Build the parts that stay valuable when models get better; expect to delete the parts that only compensate for today's limitations.
- Budget for tokens like a channel, not a licence. Expect variable spend, route between cheap and frontier models, and measure cost per useful outcome.
The thread running through all of it is that agents need a reliable supply of context. That is also the case for RSSFlux: the product turns sites, feeds, and social accounts that lack a usable feed into structured input that automation and agents can consume. If you are wiring context into a workflow, see how the feeds work — or read how we handle X/Twitter feeds now that they require a paid provider.
Source: Is SaaS Dead? Box CEO Aaron Levie on What AI Changes — Matthew Berman.
Frequently asked questions
- What is Box Architect?
- An autonomous system Box announced that designs, builds, and later optimises agentic workflows on top of Box. An enterprise describes a process — contract management, invoice processing, supply chain document review — and the system proposes the full solution using the leading frontier model plus Box's building blocks.
- Why does Box care about a 'context layer'?
- Because agents are only as good as the data they are allowed to read. The context layer holds the access controls, governance, and metadata that decide what an agent can see and where it should look, which is what makes thousands of agents safe to run against enterprise content.
- Are token costs going to fall?
- Unit costs are falling, but total spend is rising because workflows are attaching agents to far more interactions. Levie expects token usage to become a normal budget line, managed like marketing spend, with routing between cheap and frontier models.