CHAPTERS
Why AI adoption speed matters: workflows, not models, are the bottleneck
Aaron Levie frames the next decade of enterprise AI as a race: AI will permeate enterprises, but the real limiter is how quickly people and organizations can change workflows. As human “typing speed” and manual tool use stop being constraints, jobs shift toward orchestration, review, and integration.
- •AI’s enterprise impact depends on workflow change management more than model progress
- •Removing human speed limits (email, code, marketing assets) reshapes roles and processes
- •Competitive urgency: enterprises want AI to happen to them before it happens to rivals
- •AI drives job evolution toward planning, auditing, and managing agent output
From pre-ChatGPT enterprise AI to consumer breakout—and back into the enterprise
Levie explains why AI used to be an enterprise-only technology (hard-to-use, bespoke models) and why ChatGPT’s interface enabled viral consumer adoption. He then outlines why enterprise adoption is slower despite high interest: legacy systems, governance, compliance, and data risk.
- •Pre-ChatGPT AI required custom models and high startup costs, favoring enterprise deployments
- •Chat interface + low learning curve + free access triggered mass consumer adoption
- •Enterprise friction: entrenched workflows, legacy data access issues, compliance and liability
- •Shadow IT resurges as employees bring tools like ChatGPT/Cursor into the workplace
Enterprise deployment reality: governance, liability, and “case law takes years”
The conversation emphasizes that enterprise rollout timelines are shaped by budgets, councils, and risk—especially in regulated industries. Levie highlights liability questions (e.g., financial advice), IP uncertainty, and the long arc of policy and legal precedent.
- •Enterprise AI adoption is slowed by governance councils, compliance reviews, and budgeting cycles
- •Liability and accountability for AI recommendations are unresolved in many domains
- •IP ownership and ongoing lawsuits create uncertainty for deployment decisions
- •Expect multi-year timelines for the biggest productivity gains to materialize broadly
CIOs are more bought-in than they were for cloud: the posture has flipped
Levie contrasts early cloud skepticism with today’s assumption that AI is inevitable in the enterprise. Unlike cloud, where CIOs resisted full migration, AI is viewed as a competitive necessity; leaders now focus on sequencing and execution rather than debating ‘if.’
- •Early cloud era: CIOs assumed cloud would remain a niche; many vowed “never” to migrate
- •AI era: CIO/CEO/CDO consensus that AI will take over enterprise workflows
- •The new question is operational: who deploys, how to manage change, and whether models are ready
- •High-profile examples (e.g., rapid SEC filing drafts) validate AI’s immediate value
SaaS incumbents vs AI-native challengers: why both can win
Levie argues incumbents have an advantage because agents are perfect API consumers—making AI feel like a sustaining innovation layered onto existing systems. At the same time, he expects major greenfield expansion into categories that previously lacked workable software due to unstructured work.
- •Agents + APIs: SaaS platforms can add automation without replacing the underlying system
- •For many use cases, deploying “the ServiceNow/Workday agent” beats rebuilding the category
- •Unlike on-prem→cloud (multi-tenant rewrites), AI often layers onto existing stacks
- •Startups still win where there was no true incumbent or software category before
Is AI just a consumption layer? Interfaces, dashboards, and domain workflows still matter
They explore the idea that AI could become the primary interface to enterprise systems, abstracting away traditional apps. Levie pushes back: users still want prebuilt dashboards and decision scaffolding, and vertical SaaS value is often the domain workflow knowledge—not just CRUD.
- •AI may act as a new consumption layer, but full app/interface abstraction is unlikely
- •Users don’t want to prompt from scratch daily; curated dashboards remain valuable
- •Vertical SaaS defensibility often comes from domain expertise and workflow design, not tech complexity
- •Expect a hybrid future: GUI + APIs + agents coexisting
Business models and the COGS of AI: from seats to usage (and the hybrid reality)
Casado raises the economic shift: AI introduces variable inference costs that don’t match classic SaaS margins. Levie anticipates mixed models—baseline seat pricing plus usage/overages—unless humans truly disappear from the workflow as “seats,” which would trigger deeper disruption.
- •AI adds a new COGS profile; inference/agent work can push SaaS toward usage-based components
- •Near-term likely model: seat licenses plus consumption overages (as in dev tools)
- •Big disruption would be “no human seats,” which could force a radical pricing rethink
- •Incumbents may see AI as TAM expansion by automating work that previously lacked users
New AI-first categories: unstructured work in legal, healthcare, finance, and services
Levie argues AI agents unlock markets where software historically underpenetrated because the work was too unstructured and ad hoc. He predicts spend expansion in areas like legal services, wealth management, and investment banking-style workflows that never fully digitized.
- •AI expands software spend in domains dominated by unstructured documents and processes
- •Legal agent spend could jump from a small software market to many billions
- •Finance has “digital islands” (trading/banking) while other areas remained manual
- •Startups can capture new markets where no platform previously fit the workflow
Box’s AI integration: turning enterprise content into queryable, automatable data
Levie describes Box’s evolution from file sharing to an enterprise content platform and why AI is pivotal: unstructured data historically couldn’t be queried or operationalized. With AI, Box can extract fields from contracts, answer questions over documents, and enable automation over content.
- •Box manages enterprise unstructured content across collaboration, security, and integrations
- •The long-standing gap: structured data is analyzable; unstructured content gets “forgotten”
- •AI enables Q&A over documents and information extraction into structured fields
- •Once content is understood, workflows like contract routing and automation become possible
Bespoke software vs packaged workflows: why “everything homebrew” won’t happen
Casado asks whether AI makes software so easy that bespoke, per-company apps replace SaaS. Levie argues most people don’t care to customize and prefer established workflows; however, AI will massively expand the long tail of scripts, prototypes, and internal tools.
- •Two extremes: fully standardized software vs daily-generated bespoke software—reality sits between
- •Most users prefer predesigned workflows (HR, ticketing) rather than continuous customization
- •AI will accelerate the long tail of internal tools and automation requests IT never got to
- •Core systems persist while custom ‘glue’ software grows dramatically
AI in leadership and decisions: earnings prep, board prompts, and research memos
They discuss practical decision-support uses: generating analyst questions for earnings scripts, improving narratives, and quickly researching topics once delegated to staff. The memo-oriented meeting model (à la Bezos) becomes easier when AI can draft high-quality background briefs—but raises questions about whether it reduces human clarity.
- •Levie uses AI to critique earnings scripts and predict analyst questions using public patterns
- •AI helps identify missing answers and strengthen messaging rather than replace judgment
- •AI “deep research” replaces many ad hoc staff research requests and expands exploration
- •Memo-based meeting prep becomes easier, though it may trade off with forcing human thinking
Enterprise budgets: why AI spend can hide in the noise of headcount planning
Casado probes whether enterprise AI spend is zero-sum due to fixed budgets. Levie argues AI tooling is tiny compared to knowledge-worker costs, so adoption often fits within normal attrition, raises, and hiring variability—then gets justified by productivity gains later.
- •AI licenses can be ~1% of an employee’s cost—small relative to headcount budgets
- •Spend may come from slower hiring, smaller raises, or reallocation rather than layoffs
- •Budget flexibility over 1–2 years can absorb AI costs without ‘massive disruption’
- •Even a small % of knowledge-worker spend redirected to AI can dwarf prior software spend
The future of engineers: AI raises the floor, shifts skills, and complicates entry-level paths
They outline how coding is evolving from autocomplete (Copilot) to agent-generated chunks reviewed by humans. Levie expects more people to learn programming due to lower frustration and faster feedback, but warns about over-relying on “vibe coding” that yields unmaintainable systems; the key skill becomes review, architecture, and judgment.
- •AI helps strong developers disproportionately; expertise matters more in reviewing output
- •Work shifts: humans fix the AI’s ~2–3% errors while producing far more output overall
- •Entry-level engineers may be AI-native and less able/willing to code without assistants
- •Risk: over-adoption can create messy, unmaintainable codebases—moderation and discipline needed
Small businesses and the next 5–10 years: cheaper capability, more output, same equilibrium
Levie argues AI gives small businesses capabilities once limited to large enterprises—marketing, translation, bug fixing, and ‘consulting-grade’ analysis. Over 5–10 years, AI becomes anticlimactically normal: companies simply operate faster and run many more experiments; the payoff shows up as better products and societal outcomes rather than a single clean metric.
- •AI dramatically lowers the cost of high-end capabilities (e.g., marketing assets, research, translation)
- •Primary internal metric: ‘do more/faster’ rather than cost-cutting; experimentation first
- •Macro impact may be hard to see in GDP alone; improvements show up in quality-of-life outcomes
- •Long-run forecast: AI becomes embedded, output per company rises, and new roles emerge around managing agents
