See pricing
The New Startup Playbook | Box CEO Aaron Levie
The a16z ShowStartups

The New Startup Playbook | Box CEO Aaron Levie

Aaron Levie argues the biggest startup shift is not just faster coding, but a move from doing work to managing AI agents. That gives very small teams a real chance to outrun incumbents, while making strategy, workflow expertise, and distribution even more decisive.

October 25, 2026 · 34 min listen · 13 min read · Aaron Levie
0:00
–:––

Context

Aaron Levie, co-founder and CEO of Box, argues that AI agents have changed the economics of building a startup more than any prior tech wave he has seen. His core claim is that a very small team can now produce the output of a much larger company, not because people work harder, but because they can orchestrate agents in parallel. For product managers, this matters because the constraint shifts away from raw execution and toward problem selection, spec quality, workflow understanding, and distribution. The episode is especially useful if you are thinking about AI-native product strategy, vertical AI opportunities, or how startups can attack larger incumbents.

The Big Idea

AI agents shift startup advantage from doing the work to directing the work. As execution gets cheaper, the winners will be the teams with sharper strategy, clearer specs, stronger taste, and better distribution.

Levie's point is not that AI simply makes existing teams a bit faster. He argues that it changes the startup playbook itself, because small teams can now deploy resources at a scale that used to require much larger companies, at least for some kinds of work.

Key Insights

Small teams gained scale

Levie says the startup model now looks fundamentally different from earlier eras because a 3, 5, or 10 person team can increasingly produce the output of a 50, 100, or 200 person company. In earlier startup eras, even if tools improved, the work still looked broadly similar. Founders typed code, wrote copy, managed infrastructure, and were limited by their own hours. His claim is that AI breaks that old limit by letting small teams deploy much more work in parallel.

Review became the bottleneck

  • What changed: Levie describes founders as becoming managers of work, with agents doing more of the underlying execution.
  • Why it matters: the bottleneck moves from creating output to reviewing, orchestrating, integrating, and correcting output.
  • Concrete example: Levie describes a founder who is reviewing thousands of lines of code a day produced by agents.
  • PM takeaway: when review is the scarce activity, leverage comes from making work easy to inspect and approve, not just easy to generate.

Specs now drive leverage

Levie argues that as coding becomes more commoditized, value shifts toward being extremely clear about what should be built. He points to a more spec-driven way of building software, where teams write detailed instructions, define how agents should interact with the codebase, and break work into clearer units. That raises the premium on product judgment. If a PM or product-minded founder can define the right problem, the right workflow, and the right quality bar, AI turns that clarity into much more output than before.

Startups have an adoption window

Startups do not just gain speed. They gain a temporary structural advantage because incumbents have more process, more workflow baggage, and codebases that are often not prepared for agent-driven work. Levie's argument is that larger companies may know AI matters, but still struggle to rewire documentation, tooling, engineering methods, and coordination patterns fast enough. That creates a window where startups can outrun them in categories that used to be protected by sheer scale.

Agents will go vertical

Levie thinks the biggest startup opportunity is not one generic AI company. It is a very large set of specialized agents across industries and job functions.

  • He suggests thinking in a matrix of industries x job functions.
  • He says that matrix could have something like 5,000 cells, each a potential startup idea.
  • His examples include agents for legal work, consulting, healthcare, life sciences, and regulatory review.
  • His broader claim is that nearly every kind of knowledge work can support a specialized agent company.

For PMs, this is a useful way to search for opportunities. Instead of asking "what AI app should we build," ask "which workflow, for which role, in which industry, is painful enough to automate deeply?"

Cheaper labor expands demand

One of Levie's most important claims is that AI does not just substitute for existing work. It can create demand that did not previously exist because the old human version was too expensive. His example is legal review. If high-end legal work costs $1,000 to $2,000 per hour, and an agent can handle some version of that work for $5 or $10 per hour, the market may expand far beyond current spend. Companies may start doing work they previously left undone because it always sat at the bottom of the budget list.

This is a strong product insight. TAM is not always fixed. In some markets, cheaper AI labor changes the size of the market itself.

Enterprise captures future gains

Levie thinks a large share of future model improvements will matter more in enterprise than in consumer. He cites Dario Amodei's framing that moving a model from "undergrad in chemistry" to "grad student in chemistry" may not matter much in everyday consumer use, but could matter a lot to a company like Pfizer. His conclusion is that many consumer use cases are already "good enough," while many enterprise workflows still have a large quality gap that better models can close.

That does not mean consumer is dead. It means the next layer of model progress may show up more clearly in high-stakes business workflows than in everyday personal tasks.

Distribution still decides winners

Levie is explicit that AI does not remove the need for go-to-market.

  • What he rejects: the idea that building a great agent is enough.
  • What he says instead: "if you build it, they will not necessarily come."
  • Why it matters: AI startups will still compete for attention, trust, budgets, and adoption.
  • His test: can you build and keep access to an audience through community, marketing, sales, PR, LinkedIn, X, Reddit, or other channels?

His warning for PMs and founders is simple. A great product without distribution is still vulnerable to a worse product with better reach.

Mental Models & Frameworks

Typist to orchestrator shift

Levie frames the startup change as a shift in the unit of leverage. In the old model, progress tracked how fast founders could type into a keyboard. In the new model, progress tracks how well they can direct, review, and combine the work of multiple agents. Use this model when deciding whether a workflow is truly AI-native. If a person is still acting mostly as a producer, the team may not have rethought the workflow yet.

Industry by function matrix

Levie's startup opportunity map is simple and powerful:

  • Pick an industry.
  • Pick a job function inside that industry.
  • Ask whether an agent could perform a meaningful part of that workflow.
  • Evaluate whether the market needs a vertical product, a horizontal product, or both.

Use this model when generating startup or product ideas. It forces specificity and pushes teams toward real workflows instead of generic AI features.

Reviewable work creates leverage

A workflow only becomes highly leveraged if one person can supervise output from many agents faster than they could have produced that output alone. That is why Levie focuses on reviewing code rather than just generating code. Use this model when choosing where to apply AI first. The best early targets are workflows where output can be checked quickly, errors can be caught cheaply, and approvals are clear.

Specialized agents beat super agents

Levie pushes back on the idea that one all-purpose agent will solve everything. He points to context limits and the need for specialization, whether through distinct tools or subagents inside a broader system. He gives Claude Code's subagents as one example of how teams create specialization without needing a completely different product for every task. Use this model when a generic assistant starts to fail under too much context or too many responsibilities.

Price reveals hidden demand

Levie's demand model is that a big drop in labor cost can uncover work that companies always wanted done, but could never justify paying for. That means demand may not scale linearly with price. A 10x cost reduction can unlock far more than 10x usage if it removes a budget bottleneck. Use this model when sizing AI markets. Do not assume today's spend is the ceiling.

Decision Principles

Principle: Pick known workflows

  • When: choosing a market, startup idea, or AI feature area.
  • Why: Levie argues that the strongest starting point is a workflow you already know deeply, a workflow your co-founder knows deeply, or a workflow you study closely enough to understand like an operator.

Principle: Design for review

  • When: introducing AI into product, engineering, or operations work.
  • Why: leverage comes from making AI output easy to verify and integrate. If review takes as long as doing the work yourself, the workflow has not really improved.

Principle: Attack rigid incumbents

  • When: deciding where a startup can compete against a larger company.
  • Why: Levie believes AI adoption is hardest where documentation, process, and coordination are already deeply entrenched. Those are the places where a small team may move faster than a much bigger rival.

Principle: Build distribution early

  • When: planning a launch or product strategy for an AI startup.
  • Why: Levie does not think AI changes the core reality that products still need repeatable distribution. Attention spikes are not the same thing as a go-to-market engine.

Trade-offs & Nuance

Speed comes with variance

Levie is enthusiastic about agents, but he is not presenting them as magic. He says agents still make mistakes and compares the experience to a slot machine, where teams may need repeated attempts before getting usable output. The upside is huge leverage. The cost is inconsistency, rework, and the need for strong review.

Specialization adds coordination cost

Specialized agents solve context problems, but they also create orchestration work.

  • A more specialized system can improve quality within a narrower task.
  • A more fragmented system can create integration overhead between tools, agents, and workflows.
  • Levie does not offer one universal architecture because he thinks teams are still figuring this out in different ways.

Enterprise beats consumer on TAM

Levie leans strongly toward enterprise because AI is fundamentally a productivity technology, and productivity gains often monetize more directly in business settings. But he does not dismiss consumer. His nuance is that consumer opportunities may come more from better packaging and new product experiences than from raw model improvement alone.

Cheap labor can fool sizing

Levie is bullish on specialized agents, but he also notes that some markets will still be TAM constrained and some startups will fail to find enough scale. Lower prices can unlock new demand, but not every narrow workflow becomes a venture-scale business. PMs still need to test market size, frequency, willingness to adopt, and expansion paths.

Common Mistakes

Mistake: Chasing attention, not channels

Levie warns against confusing a viral moment with distribution. A flashy launch video or burst of social engagement may create awareness, but it does not create a repeatable pipeline of users or buyers. The better approach is to build ongoing access to a specific audience through communities, content, product-led loops, sales, or partnerships.

Mistake: Building without workflow depth

Levie repeatedly comes back to knowing the job you want the agent to do. A team that only knows the technology may build a demo, but not a durable product. His advice, drawing on Jared Friedman's idea, is to do the job yourself, work with a co-founder who has done it, or get close enough to the workflow that you understand how the work actually happens.

Mistake: Trusting one general agent

Levie argues that teams should not assume one super-agent can reliably own every task. Context overload, unclear boundaries, and weak specialization can make quality worse, not better. The better approach is to define narrower responsibilities and clearer workflows for agents, especially in code or other high-stakes tasks.

Practical Application

Turn PRDs into executable specs

  • Write product specs with enough detail that an engineer or agent could act on them without a live walkthrough.
  • Include workflow boundaries, acceptance criteria, edge cases, and quality expectations.
  • Review whether the document clarifies what to build, not just why it matters.

This follows Levie's argument that clearer specs now create disproportionate leverage.

Map one workflow cell

  • Choose one industry x job function intersection your team understands well.
  • List the repetitive decisions, high-cost steps, and bottlenecks inside that workflow.
  • Ask which part could be owned by a specialized agent rather than a generic assistant.
  • Size the opportunity twice: once using today's budget, and once using a much lower AI-delivered price.

This turns Levie's abstract market map into a concrete product discovery exercise.

Audit agent review loops

  • Identify one workflow where people still spend most of their time producing output rather than reviewing output.
  • Test whether agents can generate a first draft that a human can verify quickly.
  • Measure review time, defect rate, and integration pain before scaling the workflow.

If review is not substantially faster than manual production, the workflow is not yet a good AI fit.

Pair builders with operators

  • Match a technical builder with someone who has deep workflow knowledge in the target domain.
  • Have the operator explain exceptions, judgment calls, and failure cases, not just the happy path.
  • Use those insights to define where the agent should act autonomously and where a human should stay in the loop.

This reflects Levie's view that strong AI products come from pairing agent-building skill with real workflow expertise.

Build audience before launch

Start building direct access to likely users before the product is ready. That can mean a niche community, a content stream for a specific role, founder-led sales conversations, or a steady presence where the target audience already spends time. Levie's point is that owning a relationship with an audience is part of the product strategy, not something to bolt on later.

Questions to Consider

  • Where is your team still limited by how fast one person can type, instead of by how fast one person can review AI-generated work?
  • Which specific industry and job-function intersection does your team understand deeply enough to build a specialized agent for that workflow?
  • If AI reduced the cost of one expert workflow in your company to one-tenth of today's cost, which currently unfunded task would suddenly become worth doing at scale?
  • What process, documentation, or architecture would stop a larger incumbent in your market from adopting background agents as quickly as your team could?
  • Do you have repeatable access to a target audience through community, content, product-led growth, or sales, or are you mistaking one-time attention for distribution?

Bottom Line

Aaron Levie's core claim is that AI agents give small teams a level of leverage that used to belong only to much larger companies. That does not make strategy or go-to-market less important. It makes them more important, because once execution gets cheaper, the hard part becomes choosing the right workflow, writing the right spec, and getting the product in front of the right users.

Notable Quotes

"They're more managers of work and the work is being done by agents." (Aaron Levie)

"If you build it, they will not necessarily come." (Aaron Levie)

"You have to own an audience." (Aaron Levie)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free