John Radosta, Principal AI Engineer & Partner
February 23, 2026
22 min read
AI Strategy
The Ivory Tower Problem
Silicon Valley has an information problem it doesn't know it has.
The most prominent voices shaping AI market narratives—venture capitalists deploying billion-dollar funds, FAANG executives managing platform strategy, and conference keynote speakers recycling each other's frameworks—share a critical blind spot. They're theorizing about a battlefield they're not fighting on.
Robert Greene's Mastery makes the case that practical knowledge, earned through direct engagement with reality, consistently trumps abstract theorizing. The AI market in 2025–2026 is proving Greene right at scale. After leading more than 100 enterprise AI transformations—across manufacturing, financial services, healthcare, legal, and government—and speaking with thousands of business owners and operators about their transformation goals, a clear pattern has emerged: the consensus narrative from Silicon Valley's elite is dangerously wrong on at least ten fundamental questions about where this market is headed.
What follows is a point-by-point correction, drawn from direct field experience, cross-referenced with the data that the ivory tower keeps ignoring.
In This Article
- Google's Warning: The Wrapper and Aggregator Extinction Event
- Sequoia's Framework: Vertical Agents Are "Act Three" of the Future
- The "Application Layer Captures All Value" Myth
- The Cursor Case: Misreading the Growth Mechanism
- The "McKinsey of AI" Opportunity Is Bigger Than VCs Think
- Outcome-Based Pricing Is the Future—But Nobody Knows How to Do It
- HealthTech AI Gold Rush: Real Opportunity, Wrong Assumptions
- The "Generative Engine Optimization" Gold Rush Is SEO 2.0 Hype
- The AI Personal Assistant "$100B Land Grab" Ignores the Trust Problem
- "LLMs Have Quietly Changed Everything" — The Accurate Claim with the Wrong Implication
Google's Warning: The Wrapper and Aggregator Extinction Event
What the Ivory Tower Claims
Google Cloud VP Darren Mowry warned that LLM wrapper and aggregator startups face an extinction-level threat, arguing that as foundation models improve with every release, the value a wrapper adds evaporates. His advice to founders: "Stay out of the aggregator business". The implication is that anyone building on top of someone else's models is doomed.
What the Battlefield Shows
The reality is the inverse of Mowry's prediction. After engaging with thousands of business owners and employees on their transformation goals, a consistent pattern emerges: LLM model preferences are fading fast as capabilities converge toward parity. The top 10 foundation models—both open and proprietary—now cluster within five percentage points on common benchmarks like MMLU, GPQA, and HumanEval.
The open-to-closed performance lag has shrunk from 24+ months in early 2023 to roughly 12–16 months. Open-source models like LLaMA-3, Qwen-72B, and DeepSeek R1 now match or exceed closed models in specific domains.
LLM Performance Convergence
Benchmark performance gap: Closed vs. Open models (percentage points)
This convergence doesn't kill wrappers—it makes them more valuable, not less. When every model is roughly equivalent, the differentiator becomes the application layer: the workflow integration, the domain-specific prompt engineering, the data pipeline, the user experience. Perplexity is the clearest proof point—a model-agnostic "wrapper" that processed 780 million queries in May 2025, achieved $148M in annualized revenue, and attracted a valuation approaching $18 billion. By definition, Perplexity is an aggregator that routes queries across multiple frontier models including OpenAI, Anthropic, and xAI systems. It should be extinct by Mowry's logic. Instead, it's one of the fastest-growing AI companies on the planet.
LLMs are becoming commoditized with rapidly diminishing differentiation across providers
The "wrapper" isn't a weakness—it's the product. Google's warning is self-serving platform strategy dressed up as market analysis. The historical parallel is precise: investing heavily in any single model provider today is like betting on mainframe computing in the 1970s.
Sequoia's Framework: Vertical Agents Are "Act Three" of the Future
What the Ivory Tower Claims
Sequoia's AI Ascent 2025 keynote, delivered by partners Pat Grady, Sonya Huang, and Konstantine Buhler, presented a three-act framework: Act One was novelty apps, Act Two was reasoning models, and Act Three (2025+) would be vertical agents—AI systems trained end-to-end for specific workflows using reinforcement learning, synthetic data, and user feedback.
What the Battlefield Shows
Agent Reliability Problem: Compounding Error Rate
End-to-end success rate for 20-step workflow
Agents are—at their core—prompt templates with tool-calling capabilities. The framing of "Act Three" as some qualitative leap beyond what exists today dramatically oversells the technical maturity. The fundamental limitation is mathematical: even at 95% per-step accuracy, a 20-step agent workflow achieves only 36% end-to-end success. At the 99.9% reliability threshold that production enterprise systems demand, even 99% per-step accuracy yields only 82% success across 20 steps.
Reinforcement learning, the mechanism Sequoia invokes for agent improvement, faces a critical limitation in business contexts: policy optimization requires consistent reward signals from environments with bounded variability. The majority of high-value business workflows involve tasks with a high degree of variability—edge cases, exceptions, ambiguous inputs, shifting requirements—that will always require a human-in-the-loop (HITL) or human-on-the-loop (HOTL). These are precisely the workflows where the most value is created. RL cannot handle these tasks because whatever policy optimization is used will always encounter lurking variables with insufficient historical data for training.
Fully automated vertical agents are too risky for the majority of high-value business workflows
HITL/HOTL architectures are the only path forward for enterprise-grade reliability. The "Act Three" narrative conflates demo-worthy agent behavior with production-grade deployment. Even Forbes acknowledges that human-in-the-loop models remain essential, with 65% of organizations now routinely deploying generative AI with human oversight requirements.
The "Application Layer Captures All Value" Myth
What the Ivory Tower Claims
Sequoia, Benedict Evans, and nearly every VC fund have converged on the thesis that foundation models are commoditizing and all value will accrue to the application layer.
What the Battlefield Shows
This is directionally correct but misleadingly oversimplified. The claim that "all value accrues to the application layer" ignores the reality that most application-layer companies will also fail—because the application layer is where competition is fiercest and switching costs are lowest. 28.5% of AI SaaS products built in 2024 are already dead by early 2026. The application layer doesn't automatically create value; it creates opportunity for value that requires deep domain expertise, workflow integration, and data flywheels to capture.
The VC narrative treats "application layer" as if it were a single thing. In practice, the application layer spans a spectrum from commodity chatbots (doomed) to deeply embedded workflow engines (defensible). What creates defensibility isn't being at the application layer—it's having proprietary data loops that train and align the intelligence layer to a real domain. Model weights commoditize, UX is easy to copy, and infrastructure is easy to rent. The data loop is the only durable moat.
"Application layer wins" is a half-truth
The application layer is also where most companies die. The differentiator is depth of domain integration and data flywheel velocity—not simply "being at the app layer."
The Cursor Case: Misreading the Growth Mechanism
What the Ivory Tower Claims
Cursor—which went from $1M to $100M ARR in 12 months and reached $1B ARR with ~300 employees—is cited by nearly every AI fund as proof that AI-native vertical tools grow explosively and that developer tools represent the premier AI investment category.
What the Battlefield Shows
Cursor is real and impressive. But the lesson VCs draw from it is wrong. Cursor succeeded not because it's an "AI-native vertical tool" in some abstract category sense. It succeeded because it eliminated switching costs entirely by forking VS Code (developers' existing environment), achieved a 36% freemium conversion rate against an industry standard of 2–5%, and grew with zero sales team through pure product-led organic adoption.
What VCs See
- AI-native vertical tools grow exponentially
- Developer tools are the premier AI category
- Technology innovation drives growth
What Actually Happened
- Zero switching costs (forked VS Code)
- 36% freemium conversion (vs. 2-5% industry avg)
- Product-led distribution (zero sales team)
Cursor is spending essentially all its revenue on AI API costs, betting that its proprietary models like Composer will eventually bring margins to 30–40%. That's a high-wire act, not a repeatable playbook. The lesson Silicon Valley should be drawing isn't "build AI developer tools"—it's that eliminating switching costs, deeply integrating into existing workflows, and letting product quality drive distribution matters more than any technology narrative.
Cursor's success is about distribution mechanics and zero-friction adoption, not AI-native architecture
Most companies applying the "Cursor playbook" are copying the technology surface while missing the distribution engine.
Enjoying This Analysis?
Get our latest thought leadership articles on AI strategy, transformation, and what's actually working in the field—delivered directly to your inbox.
No spam. Unsubscribe anytime. Business emails only.
The "McKinsey of AI" Opportunity Is Bigger Than VCs Think—and They Still Don't Understand Why
What the Ivory Tower Claims
Multiple sources describe an emerging "$18 billion AI consulting opportunity" being attacked by AI-native boutique firms challenging McKinsey, BCG, and Deloitte. The thesis: consulting is becoming a go-to-market wedge that creates enterprise relationships and switching costs.
What the Battlefield Shows
The ivory tower is accidentally right about the size of the opportunity but fundamentally wrong about why it exists and who will capture it. The gap isn't a "market opportunity" in the VC sense—it's a structural failure of the entire AI vendor ecosystem.
These failures aren't addressable with better tooling, smarter agents, or AI-native consulting platforms. The failure is happening at the human layer: 38% of failure points trace to user proficiency—people not knowing how to use the tools. 65% are organizational failures—governance, roles, process, and culture. Only 22% are genuinely technical.
The VCs building "Xavier AI" bots that charge $99/month and claim to replace McKinsey are solving the wrong problem. Mid-market companies—where 98.5% of CEOs believe AI has value but only 7% have a strategy—don't need a chatbot. They need practitioners who have been in the trenches, who understand that AI transformation is 10% algorithms, 20% technology, and 70% people and processes.
Where AI Initiatives Actually Fail
Root cause analysis of AI project failures
The AI consulting opportunity is real, but it belongs to battle-tested practitioners, not to AI chatbots or VC-funded platforms
The problem is organizational, not technical—and no amount of AI-native tooling changes that.
Outcome-Based Pricing: The Future Nobody Can Execute
What the Ivory Tower Claims
The VC consensus is that per-seat SaaS pricing is dying and outcome-based pricing is the inevitable replacement. Seat-based pricing adoption dropped from 21% to 15% in 12 months, while hybrid models jumped from 27% to 41%. Gartner projected 30%+ of enterprise SaaS would incorporate outcome-based components by 2025.
What the Battlefield Shows
The direction is correct. The execution reality is a disaster. Pure outcome-based pricing breaks in practice because attribution is messy, outcome definitions vary wildly across customers, and vendor unit economics become unpredictable. Forbes called outcome-based pricing "the most expensive myth in enterprise AI" because defining, measuring, and attributing "outcomes" in complex enterprise environments creates more friction than the pricing model eliminates.
Pure Outcome-Based Pricing
- Attribution is messy and disputed
- Outcome definitions vary by customer
- Vendor unit economics unpredictable
- Creates more friction than it eliminates
Hybrid Pricing Models
- Base subscription for revenue predictability
- Outcome tier for elasticity & expansion
- Usage proxies that feel outcome-aligned
- Pragmatic over theoretically pure
What's actually working on the ground isn't outcome-based pricing—it's hybrid models that blend a base subscription with usage or outcome-based tiers. The base provides revenue predictability for both vendor and customer. The outcome layer provides elasticity for expansion as AI results improve. The companies getting this right are measuring outcomes always but pricing on them selectively—using usage-proxy models that feel outcome-aligned without the fragility of pure performance guarantees.
Hybrid pricing models work. Pure outcome-based pricing is theoretically elegant and operationally broken
The battlefield favors pragmatism over purity. VCs pushing pure outcome-based pricing on their portfolio companies are setting them up for unit economics disasters in the first enterprise contract negotiation.
HealthTech AI Gold Rush: Real Opportunity, Wrong Assumptions
What the Ivory Tower Claims
Healthcare is the hottest vertical AI investment category, with PitchBook reporting a 300% YoY increase in VC flowing to AI health solutions. Digital health startups raised $9.9 billion through Q3 2025. MergeLabs closed a $250M seed at $850M valuation.
What the Battlefield Shows
Healthcare AI is indeed a massive opportunity—but the VC lens misreads why. The ivory tower sees healthcare as a vertical where AI agents can automate clinical documentation, diagnostics, and drug discovery. The battlefield sees healthcare as the ultimate proof point for HITL/HOTL architectures and the impossibility of full automation in high-stakes environments.
Nearly 86% of healthcare mistakes are administrative errors caused by manual processes or outdated systems. AI can address these—but only with human oversight baked into every workflow. Gartner projects 30% of new legal and healthcare tech automation solutions will include mandatory HITL functionality. Regulatory requirements (HIPAA, FDA, state medical boards) make fully autonomous AI agents a non-starter for the foreseeable future.
Healthcare AI validates the HITL thesis, not the autonomous agent thesis
The biggest returns will go to companies that augment human expertise rather than replace it.
The "Generative Engine Optimization" Gold Rush Is SEO 2.0 Hype
What the Ivory Tower Claims
GEO (Generative Engine Optimization) is the new SEO—a fundamental shift in how businesses get discovered, with 63% of websites now seeing traffic from AI platforms. Businesses need to optimize for ChatGPT citations, enable OAI-SearchBot crawling, and build structured data profiles.
What the Battlefield Shows
GEO is real in the sense that AI-driven search is growing. But the playbook being sold—"optimize for ChatGPT the way you optimized for Google"—fundamentally misunderstands the mechanism. ChatGPT commands nearly 40% of user market mind share, while Perplexity hovers between 5–10%. The traffic patterns from AI platforms are fundamentally different from traditional search: users ask fewer, longer queries, get synthesized answers, and have less reason to click through.
More importantly, the GEO playbook ignores what actually drives AI citation: authority and factual density. The websites that get cited by AI search engines are the ones that have genuine domain expertise, original research, and authoritative content. The same businesses that built SEO empires on keyword stuffing and backlink farms will fail at GEO because AI models are better at distinguishing real authority from manufactured signals.
GEO is real but overhyped as a tactical playbook
The winning strategy is the same boring one it's always been—build genuine domain authority and publish original, high-quality content. The tactical optimization layer matters far less than the substance underneath.
The AI Personal Assistant "$100B Land Grab" Ignores the Trust Problem
What the Ivory Tower Claims
LLM-powered personal assistants represent a $100B opportunity: auto-draft replies, schedule meetings, understand priorities, and run your inbox while you sleep. The "invisible ops" thesis posits AI copilots living in Gmail, Outlook, and Slack.
What the Battlefield Shows
The Compounding Error Problem for AI Assistants
Wrong emails sent per month at different accuracy levels (50 emails/day)
From the practitioner perspective, the technical capability for inbox automation is largely here. The trust barrier is nowhere close to being solved. Delegating email composition—where a single hallucination or misinterpreted context can damage a client relationship, leak sensitive information, or create legal liability—requires a level of reliability that current LLMs simply don't provide at scale.
The compounding error problem applies directly: an AI assistant managing 50 emails per day with 97% accuracy sends 1.5 wrong emails daily. Over a month, that's 30+ errors reaching real recipients—with real consequences in professional contexts. The economic value is obvious; the risk calculus for knowledge workers, executives, and anyone in regulated industries doesn't close.
Enterprise data security adds another layer. Browser-level and file-upload DLP systems exist precisely because organizations cannot trust automated processes to handle sensitive communications without guardrails. Any personal assistant that touches email at scale immediately becomes a data exfiltration vector.
Browser-Level DLP for AI Model Protection
DataFence intercepts file uploads and sensitive data at the browser level before they reach AI models, cloud storage, or unauthorized SaaS. Learn more →
The AI personal assistant market is real but gated by a trust and reliability threshold that no current model reliably clears
HITL review of AI-drafted communications will be required for years, limiting the "autopilot" framing.
"LLMs Have Quietly Changed Everything" — The Accurate Claim with the Wrong Implication
What the Ivory Tower Claims
Every SaaS product will have a native LLM layer by 2026. Andrej Karpathy admitted he's "mostly programming in English now". LLMs are no longer a feature—they're an expectation. The defensibility is in the data loop, not the model or the UX.
What the Battlefield Shows
This is the one claim from the broader AI discourse that is largely correct—but the implication VCs draw from it is still wrong. Yes, LLMs have changed the software landscape fundamentally. Yes, every serious SaaS product will integrate LLM capabilities. But the VC conclusion—"therefore, invest in LLM-native startups"—ignores the structural advantage this shift gives to incumbents.
When LLMs become a commodity input, existing SaaS companies with distribution, customer relationships, and data gravity can integrate LLM layers faster than startups can build distribution. The 28.5% death rate of AI SaaS built in 2024 is evidence of this dynamic playing out. Startups are building LLM-native products; incumbents are adding LLM layers to products that already have customers. Late 2025 saw AI features shift from gated add-ons into core plans across major SaaS platforms.
Startup Advantage
- Speed to build LLM-native features
- No legacy technical debt
- AI-first product design
Incumbent Advantage
- Existing distribution & customer base
- Years of proprietary training data
- Can integrate LLM layers quickly
- Data gravity creates durable moat
The data loop thesis is correct—but data loops require data, and incumbents have years of it. The startup advantage is speed; the incumbent advantage is data gravity. In most categories, data gravity wins.
LLMs changing everything is accurate. The startup advantage from this shift is smaller than VCs think
The same commoditization that enables LLM-native startups also enables incumbents to integrate the same capabilities with better data and distribution.
The Meta-Lesson: Why Silicon Valley Keeps Getting It Wrong
The pattern across all ten corrections is the same: Silicon Valley optimizes for narrative coherence, not operational truth.
- VCs need investable theses, so they construct frameworks ("Act Three," "application layer captures value," "$100B land grabs") that make categories legible to limited partners.
- FAANG executives need platform lock-in narratives, so they declare categories dead ("wrappers are extinct") when those categories threaten their margin structures.
- Conference speakers need dramatic arc, so they project exponential curves onto linear realities.
None of these incentives align with accurately describing what's happening on the ground inside actual enterprises trying to make AI work. The practitioners—the people redesigning workflows, cleaning data, managing change, and building HITL architectures—see a different reality:
- • Models are commoditizing. The moat is not model access; it's domain knowledge and organizational execution.
- • Full automation is a trap for high-value workflows. HITL/HOTL is the only architecture that survives contact with enterprise reality.
- • The failure rate is organizational, not technical. 70% of the equation is people and process. 22% is genuinely technical.
- • Speed and pragmatism beat elegance. Mid-market companies achieve 90-day pilot-to-production timelines when they pick specific problems and work backwards.
- • Distribution beats technology. Always has. Always will.
The AI market isn't about to see a "massive shakeout" in the way Silicon Valley frames it—as a Darwinian event that rewards their portfolio companies and kills everyone else. The shakeout will reward practitioners who understand the messy reality of making AI work inside real organizations.
Everyone else—wrappers, agents, platforms, and consulting bots alike—is just selling shovels to people who've never dug a hole.