Why startups win in the AI era, with Box CEO Aaron Levie

Executive overview

Most enterprise data is unstructured — documents, contracts, presentations — and software has never been able to automate work on it. AI agents change that, unlocking a vast backlog of work companies couldn't justify doing before.

The AI transition differs from cloud in one key way: no one needs convincing. The opportunity now is execution — building safe, reliable, data-connected agents for workflows that previously required expensive human labour.

Startups that find new "nouns and verbs" — categories of professional work with no incumbent software — will define the next generation of $5–20B companies.

Cloud to AI: what changed

• Box pivoted from consumer to enterprise in 2005, riding the mobile and cloud wave to build a content management business.
• Cloud required convincing IT orgs; AI doesn't — the ChatGPT moment made every business leader a believer.
• Enterprise adoption is still slow, but the blocker is change management, not scepticism about AI's potential.
• The core shift: unstructured data (files, contracts, invoices) was always inert. AI agents make it queryable and actionable.

The AI job paradox

• Most company time goes to necessary-but-not-strategic work: finding information, manual data extraction, routine reviews.
• AI agents free up that time, but the bigger unlock is the backlog — work that was never economically viable to start.
• A 50-person company acting like a 500-person company grows faster and hires more people, not fewer.
• Large incumbents (Amazon) may reduce headcount; small companies will expand into more markets and functions.
• The press frames this as job destruction; the reality inside companies is relief at offloading low-value work.

The startup opportunity: new nouns and verbs

• From 2008–2014, every consumer and enterprise "noun and verb" (food, music, CRM, payroll) got solved.
• That left startups competing against modern, well-funded incumbents with no clear attack vector.
• AI breaks this open: there is a long list of professional work categories — legal review, compliance, research — that software never touched because only humans could do it.
• The window is now: roughly 2024–2027. Hundreds of companies started in this window will reach $5–20B.
• Don't target "CRM with AI" — incumbents will do that. Target categories with no existing software at all.

New business models: from seats to consumption

• Traditional SaaS sells seats — one licence per human. This caps addressable market at headcount.
• AI agents decouple value delivery from human count. Three lawyers can get the work of unlimited lawyers.
• New model: charge per unit of work (contracts reviewed, documents processed), not per user.
• Pricing example: human review costs $5–10 per contract; AI does it for 10 cents; charge $2 and capture high margins.
• Risk: pure consumption revenue is lumpy. Maintain a subscription base to protect recurring revenue.
• Gross margins depend on software depth above the AI tokens — thin software compresses to 2x token cost; rich workflow software supports 80–90% margins.

Core vs context: why enterprises won't build their own software

• Jeffrey Moore's core vs context framework: every company must decide where to innovate versus just keep things running.
• Disney's core is IP and storytelling; HR software is context — they need it to work, not to be bespoke.
• AI makes bespoke software easier to write, but most companies won't maintain custom systems for context functions — liability, bugs, no recourse.
• Klarna-style self-build stories are interesting PR but won't generalise. Companies will still buy software for context.
• Custom software will grow for core functions — the parts that genuinely differentiate the business.

Competing with incumbents

• Incumbents will add AI to their existing install base. That still leaves millions of businesses they don't serve.
• Workday has ~10,000 customers; there are ~10 million relevant businesses globally. Go after the rest first.
• Find use cases outside the incumbent's natural scope — they won't build agents for workflows adjacent to their core product.
• Always assume competitor agents are excellent, then find the strategy that still wins from there.

Advice for founders

• Read: Crossing the Chasm, Innovators Dilemma, Blue Ocean Strategy — internalise these before building.
• Have a co-founder. Even non-technical. The grind is more survivable with someone alongside you.
• Only pursue markets genuinely transformed by AI — don't fight unnecessary headwinds.
• Ride tailwinds, build a big vision, go now. This window closes in two to three years.