Tech CEOs are collectively hallucinating AI’s capabilities—overestimating automation’s readiness while underestimating the “last mile” of implementation. Box founder Aaron Levie’s diagnosis of “AI psychosis” cuts to the core: executives demo a proof-of-concept, see a contract drafted or a prototype coded, and leap to the conclusion that AI agents can replace entire workflows. The disconnect? They’ve never trained a model on enterprise-specific data, debugged hallucinated outputs, or reconciled API latency with real-world SLAs. This isn’t just hype—it’s a systemic misalignment between executive vision and engineering reality, with billion-dollar consequences.
The Last Mile Problem: Why CEOs See Agents Where There Are Only APIs
Levie’s critique hinges on a fundamental architectural mismatch. AI systems today are orchestration layers, not autonomous agents. Take contract review: a CEO might see an LLM flag a clause in a sample document and assume the system can handle all 50,000 contracts in a legal database. But the reality? Fine-tuning a model on proprietary legalese requires:
- Domain-specific embeddings (e.g., training on Lawgical’s contract corpus, not generic web data).
- Prompt engineering for edge cases (e.g., distinguishing between “force majeure” in US vs. EU contracts).
- Human-in-the-loop validation for hallucinated references (e.g., citing non-existent case law).
The result? What looks like a 90% automation rate in a demo becomes a 10% usable output in production. CEOs are solving for the wrong problem: they’re optimizing for the “wow” factor, not the “ship” factor.
The 30-Second Verdict
AI isn’t failing—it’s being misapplied. The gap between executive promises and engineering constraints is widening. For every public demo of an “AI-powered” feature, there’s a private Slack channel where engineers debate whether the underlying API even meets basic fetch latency requirements.
Ecosystem Fallout: How AI Psychosis Accelerates Platform Lock-In
When CEOs overpromise AI capabilities, they don’t just mislead investors—they lock in their stacks. Consider the rise of proprietary AI cores:
- NVIDIA’s TensorRT (optimized for CUDA) vs. ARM’s Ethos-U NPUs (designed for mobile edge AI). The choice isn’t just about hardware—it’s about ecosystem dependency. A CEO who bets on NVIDIA’s AI stack today may find themselves in a vendor lock-in trap tomorrow when fine-tuning models requires proprietary tools.
- Open-source alternatives (e.g., Hugging Face’s pipelines) are being sidelined in favor of “enterprise-grade” AI suites that require custom hardware. This isn’t innovation—it’s strategic moat-building.
“The real tragedy isn’t that AI isn’t working—it’s that CEOs are using it as a Trojan horse to consolidate power over their tech stacks. By the time engineers realize the limitations, the company’s already committed to a 10-year NVIDIA contract.”
Under the Hood: The API Latency Tax No One’s Talking About
Levie’s examples—contract generation, code review—rely on APIs that are notoriously unstable. Take OpenAI’s gpt-4o API:
| Metric | Public Demo | Enterprise Reality (2026 Q2) |
|---|---|---|
| Token throughput (req/sec) | ~120 (ideal conditions) | 30–50 (with retries, rate limits) |
| Latency (p99, ms) | 800ms | 2.5–4.2s (with queueing) |
| Cost per 1M tokens | $3.00 (list price) | $8.50–$12.00 (with usage spikes) |
These aren’t bugs—they’re features of the architecture. LLMs are stateless by design; every request is a cold start. For a CEO to claim “AI handles X% of our workflows,” they’re ignoring:
- Context window decay (e.g.,
gpt-4o’s 128K limit vs. Real-world document lengths). - Hallucination cascades (e.g., an AI-generated contract clause referencing a nonexistent law).
- API deprecation cycles (e.g., OpenAI’s 2025 endpoint sunsetting).
“I’ve seen CEOs greenlight AI projects based on a single API call that worked once. They don’t understand that LLMs are not like traditional software—they’re statistical black boxes with no guarantees. You can’t treat them like a function in your codebase.”
Regulatory Whiplash: How AI Psychosis Fuels Antitrust Scrutiny
The FTC isn’t just watching AI hype—it’s documenting it. In a recent hearing, Commissioner Slade Mitchell warned that “AI-driven consolidation is accelerating at a pace unseen since the dot-com era.” The parallels to cloud computing’s early days are eerie:
- 2006: AWS launches, promising “unlimited scalability.” Enterprises migrate en masse, only to discover unexpected egress fees.
- 2026: AI vendors promise “agentic workflows.” Enterprises adopt, only to find hidden training data costs and vendor lock-in.
The key difference? AI’s opacity. Unlike cloud pricing, which can be audited, LLM training data is often proprietary. A CEO who claims “our AI is unbiased” can’t prove it—because the model’s decision-making process is effectively unexplainable.
The 90-Day Reality Check
By late 2026, we’ll see a wave of AI project cancellations as CEOs confront the “last mile.” The companies that survive will be those that:
- Treat AI as a co-pilot, not a replacement.
- Benchmark against real-world SLAs, not demo conditions.
- Avoid proprietary lock-in by using open standards (e.g., ONNX for model portability).
The rest? They’ll be the ones explaining to investors why their “AI-powered” product is still 80% manual labor.
The Takeaway: How to Avoid AI Psychosis
Levie’s advice—”use AI a ton to see what it can’t do”—is the only rational path forward. But it requires:
- Engineering-led pilots (not marketing demos).
- Cost-benefit analysis beyond “cool factor.”
- Contingency plans for hallucinations (e.g., OpenAI’s evaluation frameworks).
The tech industry has survived many hype cycles. But AI psychosis isn’t just another bubble—it’s a structural risk. The companies that thrive will be those that treat AI as a tool, not a messiah.
