Are SaaS Companies Cooked: Which Thrive & Which Die | Aaron Levie

AI Adoption and Enterprise Workflow Redesign

• Aaron Levy argues the current AI shift is a commercial and economic race rather than an existential "all-or-nothing" race against China, emphasizing that a month or two of advantage is insufficient to determine the total outcome.
• The workflow must be redesigned for AI agents rather than for human input; this shift requires a fundamental change in business process implementation rather than merely layering tools on existing human-centric workflows.
• Levy predicts there will be more lawyers in the next five years than currently exist, as the constraint on legal work is no longer content generation but the limited number of licensed professionals required to review, approve, and file work in regulated court systems.
• The ultimate constraint in industries like healthcare and law remains human capacity (e.g., the number of available doctors or licensed attorneys), meaning automation will reveal new bottlenecks rather than eliminating demand for human labor.
• A new job title, "Agent Operator," is expected to emerge with demand reaching 500,000 to 1 million roles within five years; these individuals will bridge technical knowledge (MCPs, CLIs, agents.md) and business processes to redesign workflows for automation.
• Enterprises face a significant hurdle where half of their data estate is fragmented, legacy, or unconnected, requiring years of work (potentially by professional services firms like Accenture) to curate and organize data for agent consumption.
• Liability and accountability remain human-centric; companies cannot shift blame to AI vendors (e.g., Anthropic) when errors occur, necessitating internal ownership and human oversight in high-stakes regulatory environments.

Software Value, APIs, and Security

• Jensen Huang's thesis that AI will "explode the amount of software needed" is validated by the shift from user-interface-driven value (clicking buttons) to API-driven value, where robust business logic embedded in APIs becomes the primary asset.
• Agent-based usage is expected to drive a 100x to 1,000x increase in API calls, making the "headless" nature of software a critical priority for platforms like Box, where agents consume and generate massive volumes of unstructured data.
• Software companies are advised to categorize their value proposition based on the density of embedded business logic in their APIs; simple database wrappers face obsolescence, while systems managing complex compliance and logic retain value.
• Cybersecurity risks are intensifying due to two factors: the sheer volume of AI-generated code (increasing the probability of vulnerabilities) and the ability of adversaries to scan for vulnerabilities faster using open models.
• The solution to AI-induced security risks is "agentic security," where agents are used to review, patch, and secure the very code they generate, creating a dynamic cycle of risk and defense.

Financials, Budgeting, and Market Trends

• Token budgets must transition from IT capital expenditure to regular operating expense (OPEX), allowing them to be traded against line-of-business initiatives (e.g., marketing campaigns) rather than competing solely with traditional software licenses.
• Levy estimates enterprise technology spend could rise from 10-12% to 20% of revenue as AI automation is funded directly through OPEX budgets rather than constrained by traditional IT capex limits.
• Companies are adopting token maxing and competitive allocation models (similar to venture capital pitch decks) to allocate compute resources to high-value use cases while limiting lower-tier productivity tools to cheaper models.
• Levy notes that while Wall Street is currently pricing public software companies at multiples like 3x free cash flow, the market is likely overreacting indiscriminately, and a correction separating winners from losers is expected within the next 1-2 years.
• The demand for AI adoption in the enterprise is projected to be a 20-year diffusion cycle (similar to cloud computing), far exceeding the 18-month hype cycles often assumed by venture capitalists.
• Levy predicts the total market capitalization for AI frontier rounds could eventually reach $2.1 trillion, with valuations potentially continuing to expand significantly beyond current $850 billion levels.

Competitive Landscape and Strategic Advice

• The AI ecosystem will inevitably become a multi-vendor environment; enterprises will adopt multiple models (e.g., OpenAI, Anthropic) to avoid single-vendor lock-in, API changes, or service outages, making "winning the race" less critical than in the early cloud era.
• Unlike the coding sector, the rapid adoption of AI for general knowledge work is overestimated; complex workflows require significant integration, data curation, and human oversight that cannot be replicated by current models in short timeframes.
• Levy suggests that professional services (like Accenture) are uniquely positioned to win the next five years by managing the "change management," data organization, and workflow redesign required to operationalize AI in regulated enterprises.
• New infrastructure categories are emerging, such as agent observability and evaluation tools (e.g., Braintrust), which are essential for enterprises to verify agent performance and maintain compliance across different model providers.
• Box is re-accelerating revenue growth by introducing new plan tiers focused on workflow automation and application development, demonstrating that agent capabilities can drive significant pricing power and upselling opportunities.
• Levy acknowledges that the speed of technological change requires a "tsunami of execution," forcing public company CEOs to remain hyper-connected to technical shifts (e.g., weekly model updates) while maintaining strategic stability for their stakeholders.
• In a final assessment of the industry, Levy states he is still "loading up on all of the frontier rounds," betting on the continued massive expansion of the compute market and the creation of new enterprise value layers.