Aaron Levie on AI Adoption and Enterprise Workflows | The a16z Show
Summary
Enterprise AI’s near-term bottleneck is organizational integration, not model capability. Aaron Levie says coding agents thrive because engineers are technical, autonomous, able to debug failures, and working on verifiable outputs; ordinary knowledge work instead spans less-technical users, fragmented data, legacy systems, and undocumented relationships. Diffusion from startups into large enterprises will therefore take “a number of years.”
Top-down AI mandates are producing misleading failure statistics and measurable theater rather than operational change. Martin Casado says the claim that 95% of large-company AI efforts fail is “clearly silly” because employees are already using ChatGPT effectively; he distinguishes that from centralized, consultant-led programs that lack operational alignment. Token-counting incentives make the distortion worse—Aaron Levie says he and his coworkers assign agents useless tasks because “you get whatever you measure.”
Integration, permissions, and change management remain the durable enterprise workload—and potentially a decades-long services market. Steven Sinofsky’s hard stop is that every company with 1,000-plus employees or more than 10 years of history contains “a massive [amount of] stuff that’s sitting there waiting to be integrated,” while “AI actually doesn’t help to integrate anything.” Levie argues this makes work by Accenture, Deloitte, and other systems integrators entirely logical: people must implement the agents that may later automate work.
The central architectural shift is to treat an agent as a worker with identity, onboarding, and bounded authority—not merely software embedded in another product. Casado argues that companies already spent 40 years designing interfaces and processes for messy, nondeterministic humans, so an enterprise can “hire the agent,” give it email and application access, and reuse those controls. The unresolved problem is context: agents can operate at enormous parallel scale but do not naturally know which Sally or Bob to ask when the documented system fails.
SaaS may gain machine seats, but the API-versus-browser path remains contested. Levie sees Salesforce’s “full headless” move as a bellwether and says machine use could reach 100x or 1,000x human activity; Sinofsky calls an agent “another seat” because sharing human credentials would be indefensible. Casado and Sinofsky favor an API/MCP/CLI-first approach with browser use when those interfaces fail, while Casado notes that agents may need ordinary Safari when headless browsers are blocked.
Agentic volume creates two separate risks: familiar infrastructure scaling and less-understood operational entropy. Sinofsky asks what happens when 10,000 employee agents each hit a SaaS system 500 times more often; Casado says caching and standard distributed-systems techniques can address that load. His deeper concern is that AI-generated code “gets worse over time,” potentially creating as many problems as solutions, and companies do not yet know how to govern long-running agents whose output continuously changes shared systems.
The speakers expect AI to expand software, infrastructure, and skilled employment before it eliminates them. Box saw AI produce roughly 80%–90% of one feature, yet security review still constrained release; Levie therefore estimates perhaps 2x–3x engineering productivity, not 5x–10x. More code creates more systems to secure, upgrade, and repair, while John Deere, Caterpillar, Eli Lilly, and thousands of other companies can employ engineers using Claude Code, Codex, and Cursor: “We’re just getting started with the jobs on this front.”
Deep dive
1. Enterprise adoption lags because coding is the unusually easy case
Levie describes his job as trying to “bring reality to the valley and then bring the valley to reality.” The gap is not simply executive ignorance: Silicon Valley engineering combines extraordinary technical aptitude, constant internet awareness, freedom to choose tools, and the ability to debug failures quickly.
Coding also supplies unusually favorable agent economics. Models are strong at code, outputs are comparatively verifiable, and engineers can alter their own environments. Enterprise knowledge workers are less technical, their workflows are different, and their data is “much more fragmented” inside older systems they cannot casually replace.
The result, in Levie’s view, is a genuine workflow and technology divide rather than two camps merely talking past one another. Bottling up what works in coding agents and computer-use agents for the rest of knowledge work will happen, but diffusion into mainstream enterprises should take “a number of years.”
Casado adds that secular technologies often begin with individuals while large companies make decisions centrally. Employees may already be using AI successfully even when the organization has not adjusted its processes around data, governance, operations, and compliance.
2. Central mandates and architectural churn manufacture AI failures
Casado’s reconstruction of the boardroom loop is blunt: the board tells the CEO, “We need more AI”; the CEO hires a consultant; a centralized project appears that nobody understands. Without operational alignment, those programs predictably fail even while individual AI use continues growing.
That distinction makes the reported MIT statistic that 95% of large-company AI efforts fail “clearly silly” to Casado. He nevertheless sees only tepid enterprise inroads over the past few months because first-wave CEO AI failures created bruising and skepticism that second-wave projects must now overcome.
Management can also optimize the wrong proxy. Levie says he and his coworkers assign agents useless tasks to increase measured token usage: “You get whatever you measure.” Sinofsky then recalls the early web, when companies commissioned thousands of team sites that became “dead web” after the next reorganization.
3. Fast-moving architectures make forty-year buyers hesitate
Levie assigns perhaps 5% of the adoption problem to architectural paralysis, but it is still consequential. Labs leapfrog one another while disagreeing about whether the harness belongs inside or outside the computer, whether agents run locally or in the cloud, and which tools or hosting models should govern them.
Enterprises burned by AI choices three or four years earlier now debate two or three paradigms before committing. Avoiding lock-in means engineering for multiple paths; choosing one creates deprecation risk. Either response slows deployment into the workflows where the technology would matter most.
Casado frames the progression as pure software, then an unsuccessful AI-software fusion, and now an agentic model: expose the product through a CLI or similar tool and let AI consume it as a user. Companies have effectively had to re-architect twice within a year while “speed-running” earlier cloud transitions.
Casado says a Silicon Valley career of five two-year startup stints does not teach long-lived enterprise buying; Sinofsky adds that such people lack the frame of reference for choosing an accounts-payable system expected to last 40 years. For a large company, mapping the long-term bet is a rational response to replacement cost, not mere conservatism.
4. Integration and permissions survive every agent abstraction
Sinofsky uses Silicon Valley’s antisocial archetype Gilfoyle to expose the enterprise limit: an agent can independently “do my thing” until it hits integration. Any business with at least 1,000 people or more than a decade of history contains enormous accumulated complexity, and “AI actually doesn’t help to integrate anything.”
Casado drills into the integration problem: as humans are passed between roles, each step carries different access controls; an agent that bypasses those steps creates a security risk, while one with only its user’s permissions repeatedly hits walls.
Legacy permissions are especially informal. A person knows to ask Sally for an inaccessible document or Bob for the number inside another system; an agent may simply stop, choose a non-authoritative source, or return the wrong number. Enterprises must modernize systems and provide the right data, documents, and context before automation becomes dependable.
That is why Levie finds announcements involving OpenAI, Codex, Accenture, Deloitte, and other integrators obvious rather than ironic. “No, that’s exactly how it works”: change management, implementation, and integration must precede automation, supporting businesses that could do this work “for decades.”
5. Agents may need employee-style onboarding, not another index
Casado’s counterproposal is to stop viewing AI as unusual software. LLMs are smart, nondeterministic, and able to handle long-tail complexity—the same messy traits enterprises spent 40 years accommodating in humans. A company could “hire the agent,” assign it an email address, let it access documents, and let it request what it lacks.
Levie accepts the analogy but preserves an important asymmetry. Agents can work in parallel at effectively immense scale, yet humans receive organizational context through relationships and know whom to tap on the shoulder. That knowledge is rarely documented in a form an agent can inherit automatically.
The conversation pushes the analogy toward “agent onboarding”: Sinofsky imagines orientation, a CEO culture discussion, and departmental explanations, while Levie compares it with humanoid robots. He saw a small robot need a separate button-pusher for an elevator because there was “no headless version of the elevator”; the world’s existing interfaces were designed for humans.
6. Machine seats could enlarge SaaS while changing its interface
Levie calls Salesforce’s move toward “full headless” access a bellwether for enterprise software. The business-model details remain open—an API tax, an agent seat, or another scheme—but SaaS platforms may serve agents as well as traditional users.
Levie says a headless user could operate at 100 or 1,000 times the scale of human users, creating new use cases no person could perform manually.
Sinofsky argues that every agent still needs an identity and access rights, making it “another seat” even if the price differs. It should never know more than the person directing it, and a SaaS vendor should not encourage credential sharing; that would establish bad security practice from the outset.
Sinofsky also rejects treating an LLM’s output like a clean SQL record to which access controls can simply be applied: it may disclose a protected number inside a sentence. For him, this makes the “SaaSpocalypse” thesis “much dumber” than it first appeared. He allows that clever pricing might make agents cheaper—for example, if they are read-only or tied to a person—but insists they are seats.
7. APIs should win on efficiency, but browsers remain the escape hatch
Casado disputes whether every product needs a headless future. OpenClaw uses a Mac mini partly for iMessage and partly because ordinary Safari can work where headless browsers are rejected by anti-scraping measures; an agent seeking a Zillow estimate may therefore need to behave like a human.
Casado says that when good APIs exist, agents should prefer them, then use a browser when they encounter an execution problem. Sinofsky likewise argues that an agent will likely check for an MCP or CLI before switching to a cloud, local, or parallel browser. Levie says API access is a faster way to search documents than clicking through an interface.
Levie expects increasingly rigorous training data for the MCPs and APIs of SaaS platforms, while Sinofsky’s first use of Codex computer use was for a task without an available API.
Sinofsky introduces the time dimension: internet search transformed both product documentation and how documentation was written. Agent APIs will likewise evolve from developer-oriented endpoints toward user intent—such as “I’m waiting to meet this customer; what must I know?”—while, as he warns, “layers never go away. They just get layered.”
8. Verification limits productivity, while complexity expands the labor pool
Box’s agent already searches differently from a human: it fans out across multiple queries, inspects hundreds of results, and re-ranks them instantly. Observing those patterns led Box to improve its scaffolding, and some improvements flowed back to human users—evidence that agent-oriented architecture can “lift all the boats.”
Sinofsky nevertheless asks what happens when 10,000 human users gain 10,000 agents, each producing 500 times the traffic. Casado calls load scaling standard computer science—cache read-only data and architect mutable shared state properly—but says the novel risk is entropy: AI code may “get worse over time,” creating problems alongside solutions.
Levie’s own evidence supports a bounded productivity claim. AI wrote roughly 80%–90% of a new Box feature, but release waited for a full security review against code injection. Reviews, security, and deployment pipelines still rate-limit output, yielding perhaps 2x–3x gains across engineering rather than 5x–10x.
The same boundary applies beyond code: AI can inspect unlimited accounting data for anomalies, creating net-new visibility, but humans still perform the fine-tooth-comb audit. Levie remains “unbelievably optimistic on jobs” because people must initiate, review, and incorporate the work even as their abstraction level rises.
Sinofsky recalls earlier predictions that computers would eliminate accountants and uses lawyers as another example of digitization expanding and broadening professional work. More valuable information also requires more people to consume and act on it, because “the essence of a company is acting on information.”
Levie’s closing logic is that more code means more complex systems—and therefore more upgrades, outages, security incidents, and engineering work. AI-native companies are already hiring aggressively, Casado says infrastructure businesses are benefiting from increased software volume, and future engineers may build intelligent farming at John Deere or therapeutics at Eli Lilly rather than another social-network algorithm.