Aaron Levie: a chatbot saves 30% at best, the gain is in agents that work before you ask

Asked which enterprise agents really change the work, Aaron Levie put a ceiling on the usual answer: a chatbot that people use like Google saves at best 30% of the reading time, and every other step of the workflow stays a bottleneck. The gain, he said, comes from asking why the sales rep opened the chat at all, and having background agents prepare the research deck, the customer’s product usage and the org chart before the meeting. Most workflows are not wired for that yet, and changing them is a decade of work; to him, that slow diffusion is the reason there is still so much left to build. Levie, co-founder and CEO of Box, spoke for 54 minutes in a fireside chat and Q&A for the Turing Frontier Fellows at AGI House in Hillsborough on 8 October 2026, during SF Tech Week. I was in the audience and recorded it.

The short version

• A chatbot tops out at 30%. An agent is, in his words, “the parallelization of cognitive or probabilistic work”, so the process has to be rebuilt around it, not wrapped in a chat window (42:29).
• The slow rollout is the market. Silicon Valley used S3 and EC2 within about three days of launch; the rest of the economy took two decades just to agree the cloud was the future. He expects AI to take two decades too, with an applied layer to build in between (01:47).
• Adoption follows one person, not the market cap. The best predictor he knows is the X feed of one “super AI-pilled” person inside (04:24).
• His bet is more jobs, not fewer: life sciences hiring two to ten times more people, more lawyers, Bay Area air-conditioning repair booked out by data centers. Frontline support tickets for consumer apps he expects to be 100% automated (08:28).
• Efficiency gets spent on output. AI was supposed to automate sales engineers. Instead buyers now ask to see the whole system simulated for their own company, which takes more work, not less (38:37).
• Judgment still wins. A beginner with agents learns to code ten times faster, he said, but for a deep system he would bet every day on 15 years of engineering plus agents (22:56).

Jump to a chapter: where AI took off · a two-decade rollout · one AI-pilled person · evals · jobs · his workday · newcomers and experts · juniors and universities · Q&A: interfaces · Box and agents · sales demos · agents that change work · who pays for cheap AI · what stays valuable.

Where agents actually change the work

The audience question was blunt: of all the agents sold to go-to-market teams, which truly transform the work, and could he name three? He named none. His answer was about why most of them do not (41:44).

Most workflows, he said, are not wired for agents to take part. Deploy a chatbot to the company and people use it the way they used Google: ask a question, get an answer instead of a link. As one-to-one task execution that saves at best 30% of the time you would have spent reading, and you are still stuck at every other stage of the workflow (42:29).

The fix is to step back and ask why the sales rep is prompting a chatbot in the first place. A background agent could already have built the research deck for the meeting, pulled how the customer uses your product and your competitors’, mapped the org chart and who reports to whom, and noticed that the buyer was just at a conference. That, he said, is what parallel compute in a go-to-market process looks like (43:31).

His second example was recruiting. Today AI there does a little research on public pages, because LinkedIn is largely closed off by API. With unlimited compute you would see that a person is speaking at a conference, and email them 30 minutes later about the role. “Nobody has re-engineered their recruiting workflow to even enable an agent to go and do that” (44:16). He sees that redesign of most job functions as a decade of work, done while compute is still scarce; the exercise is to imagine it abundant and ask what you would throw it at (45:03).

The last one came from his inbox. Most of his vendors, banks among them, send him emails unrelated to anything he cares about, because he sits in an “interested in tech” bucket of their decision tree. The alternative is an agent per customer, always working out the most interesting thing to pitch that person, with the obvious tension that over-emailing kills it (45:41).

Sales: the demo is now the customer’s own system

A question from the floor recalled a post of his from a couple of years ago: a Box salesperson used to present a use case, and now could just code up what is possible. Did that come true? The debate at the time, he said, was that AI would automate the sales engineer by producing the deck. He saw the opposite. Whatever AI allows eventually becomes a de facto requirement, and a customer no longer accepts an architecture diagram: “show me like the entire working system simulated for my environment” (39:23). The tool that was meant to save the sales engineer time can take more of it.

In 2026, he said, there is no reason to show how the product works at Acme Corp. You show how it would work at JPMorgan, having simulated JPMorgan, and that takes capable people and costs compute (40:39). He called it a microcosm of the wider economy: the efficiency gain is spent on a much bigger output per unit of work, and he expects to see it in sales for many years (40:03). Box does it already.

Why he expects a twenty-year rollout

Coding and customer support have “gone fully vertical”. Lower on his histogram of job functions come legal, then investment banking, then accounting (00:08). If the whole economy used AI the way the Valley does, he said, there would be nothing left to build. It does not: models have to be brought into workflows, change has to be managed, and “an entire applied layer of AI” has to get built (00:50).

The cloud is his template. Silicon Valley was on S3 and EC2 within about three days of launch, and within a year or two no startup would rent a server again. It still took two decades for the rest of the economy to agree the cloud was the future, and large workloads are still on-prem (01:47). The closer you are to the research, the stranger that feels: incredible capability in your own work, while the rest of the world still uses basic chatbot interactions (02:44).

What surprises him is that he cannot predict who has adopted what. Startups work differently, as you would expect. Inside a big company, though, at Goldman Sachs or Citigroup in his example, one person who is super AI-pilled changes the picture, and that tracks their personality and how wired into the news they are, not the market cap, the org or the function (04:36). “If I could see their X handle before every meeting, I would mostly be able to tell you, like, where is the company on its AI journey” (05:37).

He has no single recipe for the companies that get it right. They follow where the industry is going rather than the vendor who walked in with a pitch, and they are not stuck on what their legacy infrastructure allows. That sense of the current decides which vendors to work with, which models to choose and which evals to trust (06:25). The simplest test is whether a company has evals for its own workflows: if you do not know how AI affects a workflow, you cannot automate it (07:43).

Jobs: his bet is more, not fewer

He called it “somewhat of a low stakes claim”: if he is wrong, there are bigger problems than him being wrong (08:28). The basis he gave is 100 to 150 years of economic and industrial-revolution research, plus what he sees at the companies that have rolled out AI most. The open variable is not model progress but how much AI can automate across all the systems it has to touch.

• The heaviest users keep hiring. In engineering, marketing and sales, the companies using AI most are hiring and often growing faster. Speed exposes a new bottleneck, and they hire for it, even when the new person’s job is to run agents (09:52).
• Life sciences. The field is limited by how fast human scientists can do research, test it, get through the FDA and bring drugs to market, and he sees no law that says it cannot run ten times faster. If AI speeds research 10- or 100-fold, he guaranteed that life sciences would hire two to five to ten times more people, to run the tests, the labs, the supply chain and the go-to-market (11:22). A health system 10 to 50% more efficient would need “three times more doctors” (15:33).
• Lawyers. He takes the other side of the bet: in five or ten years, more lawyers. AI will bring more lawsuits and more negotiation, with one side’s AI drafting the contract and the other side’s AI cutting it down, until both get on a call and ask “whose AI should we actually trust on this?” (12:55).
• Security. Agents will execute most of the work, and people will decide where to deploy them and triage what they find (13:35).
• What goes. Frontline support for a consumer internet product, refunds, account access, shipping: “we can basically assume that’s gonna be 100% automated”. Customer success as a job stays; the tasks inside it change (14:10).

His last example was his house. Getting the air conditioning fixed was hard, because the Bay Area’s HVAC technicians had been called in to work on data centers (16:19).

How he uses AI himself

He can often compress a week or a month of work into a single day (17:25). Most of it goes into product. A brainstorm used to run one to three months: a meeting on Monday, another on Thursday, a week skipped because people were out, then the designers pulled in. Now that fits in one session with agents, and he walks into the room with the 10 or 15 people who own the area carrying a product, asking whether to build it and in what shape (18:02).

The daily-productivity layer matters less to him. He thinks he is “not bad at email”, so he has not automated it; he likes Wispr Flow because talking into the computer is faster; an AI reading his calendar might “help me by 3%” (19:09). He builds no personal software. What he builds are product ideas for the company, which he describes as non-working prototypes (32:43).

Beginners, experts and the training gap

The host asked how two of his own claims fit: agents make it easier to break into a hard field, yet experienced people keep a large edge. He sees no contradiction (20:02). He learned to code about 25 years ago by trial and error and “tapped out” early, next to friends a hundred times better. Now, within an hour of trying, a beginner sees why people write software at all. Understanding stays limited, so you need the agency to inspect what the agent did; with that, he said, you learn ten times faster than by trial and error, and on another pass “10 to 100 times faster” (21:45).

The depth does not shrink. Box stores hundreds of billions of files in a highly distributed system that has to be very secure. Who should build that: someone who learned to vibe code six months ago, or an engineer with five to fifteen years who now has agents, knows how to steer them and can tell when the agent’s judgment is wrong? “Obviously every single day you’re gonna bet on that latter person” (22:56). On recursive self-improvement research, he bets on Andrej Karpathy using agents over himself vibe coding an agent (24:15).

Training juniors is a collective-action problem in his view. Each company could optimize for mid-level people with agents, and the economy would end up with a gap in the next generation, so companies have to keep hiring from college and training people even when day-one productivity is lower (25:21). He says he has not met a Fortune 500 CEO who plans to use AI to wipe out the workforce (27:00).

Universities are the harder problem, and he said he does not know how to solve it. “How do you charge $150,000, $200,000 to teach people things that are available on ChatGPT?” (28:13). His answer is partly back to basics: learning how to learn, adapting, rigorous logic, and offline study of hard topics, which feels like the least efficient way for a year or two and works as “boot camp for your brain” (28:56). He thinks education should cost about a fifth of what it does, and on this he is “more Bernie Sanders-esque”: fund it as a society (30:53).

Who pays for cheap AI

An audience member asked about pricing: the labs subsidize consumers and charge enterprises full freight, so how does it play out? A “multi-trillion dollar question”, Levie said, and a weirder one than it looks: money from endowments like Harvard’s flows through the LPs to the startups, the startups buy the tokens, and that is what makes Claude cheap on the Max plan (47:24).

He cited reports of some plans carrying $10,000 of subsidy. That has to come down, and the labs’ bet, in his telling, is that efficiency gets it to $500 of compute in two years, with consumer plans at break-even and a 30 to 50% margin on the API. That needs cheaper inference and a stable split between frontier and open-weight models (48:04). The open question for the labs is whether AI ends up a 10% or an 80% margin business. The most AGI-pilled call that the wrong question: they expect to sell proprietary access to a bio model and be paid a billion dollars by whoever uses it for drug discovery. He thinks it might split into 10% commodity inference for everyday use and that other end (49:11).

The Q&A on interfaces and Box

On the interface of the future he called himself “a boring pragmatist”. The final form, in his view, is conversation by voice or text, with a visual representation you can confirm or reject, like ordering in a restaurant: you ask, and it happens. It may live on a watch, a charm or a pod in the house (34:04).

On where Box sits: agents need context, and that context lives in enterprise data, the same unstructured files people work with, from research documents, roadmaps and code to presentations, financial records and contracts. Box does not much mind whether the agent is its own or an external one from Claude or ChatGPT. Its job is security, compliance, governance, context retrieval and a native agent harness (36:12).

What stays valuable

The last question asked what is scarce and worth keeping, and whether compute and work itself should be rethought from first principles. That is “way above my pay grade”, he said; he chose B2B to build good software and see what enterprises buy (51:04). What he does see is an “insanely great time” to be an account executive in AI, if you have strong interpersonal skills and are technical enough to see where things are going, because the thing that is not abundant “is your ability to get in front of a customer and compel them to care about what you’re doing” (52:25). He drew it as a bell shape: heavy automation for the technical work, and rising value for judgment, intuition, creativity and the interpersonal side. Raw intelligence stays the baseline, which is why he would not throw away foundational skills (53:14).

What a solo builder can act on

My reading of the talk, not his advice:

• Rebuild one workflow instead of wrapping it. If a chat window caps out at 30%, the product is the background agent that has the work ready before anyone asks. His recruiting example is an open brief, a talk spotted and the speaker emailed 30 minutes later, and he said nobody has built it. The tooling for running many agents at once is in conductor-parallel-agents.
• Write the eval for the workflow first. His test for a company works for a single skill: plugin-eval-ablation runs each case with and without the plugin, because an absolute score measures the model, not your contribution.
• Sell to the one person. If adoption tracks one AI-pilled person and not the company, that person is the customer. The investor panel at stanford-cleantech-hackathon-2026 said something close about utilities: asked how the rare startup sold to one, the only explanation offered was a champion inside.
• Bring a working system to the first call. If buyers expect their own environment simulated, a deck is below the bar. It costs work and compute; with agents, one person can still build it for a short list of target customers.
• Do not price on today’s subsidy. If a $10,000 subsidy has to fall to $500 of compute, a product whose margin assumes today’s plan price is riding the labs’ bet. The Max plan is not a developer credential anyway: claude-subscription-boundary.
• Read what the agent wrote. His tenfold learning claim depends on inspecting the code, and reading-agent-output is about the terminal as a tool for exactly that reading.

How this was put together

I recorded the session on a DJI camera from the audience. The audio was transcribed locally with faster-whisper (medium model), and the English subtitles come straight from that transcript, so a few names are misheard in them (“Chatchabee” for ChatGPT, “Quad” for Claude). The Russian subtitles were translated cue by cue from the English, and both tracks are on the video. The recording starts a few seconds into his first answer, so the opening question is missing. The event listing also named Tanay Kothari as a speaker; this recording covers Levie’s session only.

Quotes come from the transcript with the time in the video. Numbers are as he said them, and the predictions are his, not established facts. Audience members who asked questions are not named. Seven shorts were cut from the same transcript by a script, a hook on top, one or two English words at a time in the middle and the Russian line below; the two public ones are embedded above. The video was edited with project-life2film, and the transcription workflow is in creative-video-transcribe.

Related, from the same week of SF Tech Week: neon-personal-agents-life2film-director-2026 (4 Oct) and cyberdefense-hackathon-screenshield-2026 (9 Oct).

Later the same month: running-production-with-agents-2026, the SRE side of agents in production (27 Oct).