Enterprise AI stalls as nine in ten prototypes fail
- Corporate AI projects fail as nine out of ten prototypes never enter active production.
- Unwritten company workflows and missing operational context prevent models from running useful enterprise processes.
- Systems crash during long execution runs unless external platforms provide sandbox guardrails and state memory.
Nine out of ten corporate artificial intelligence projects die before touching a production system.
The ease of generating code with artificial intelligence sparked massive corporate experimentation, but prototypes routinely break under real-world pressure. On September 15, 2026, Temporal CEO Samar Abbas detailed on This Week in Startups why ninety percent of enterprise prototypes fail to reach production. While developers easily assemble working demos on local laptops, those projects collapse under scale. When asynchronous agents run long-duration tasks, basic network blips crash mid-execution, leaving systems unable to recover state or track ongoing transactions.
Abbas argued that enterprise development must shift off individual laptops into harness-governed cloud platforms. When agents operate in code mode - generating and executing arbitrary scripts on the fly - letting models run unmonitored creates severe security and operational hazards. Resolving this requires independent execution harnesses acting as central control brains. These harnesses intercept model commands and execute them inside isolated sandboxes, providing the durable execution memory required before security teams will approve live enterprise deployments.
"The model predicts the next token, but the harness protects the business."
- Samar Abbas, This Week in Startups
This infrastructure bottleneck aligns with broader warnings about slow operational adoption. Earlier on September 12, 2026, venture capitalist Anish Acharya noted on The a16z Show that economic diffusion moves at a human pace because real-world operations face physical friction. While software engineering teams successfully build automated loops for bug reports and patches, full automation hits a hard ceiling. AI agents excel at local optimization, but they cannot pivot strategy or manage complex exceptions without human oversight.
The disconnect extends beyond execution harnesses into basic organizational knowledge. On September 18, 2026, Databricks CEO Ali Ghodsi argued on The a16z Show that enterprise artificial intelligence stalls because models lack institutional context. Most corporate operations rely on unwritten rules and informal networks that do not exist inside standard training data. Deploying basic chatbots for document search fails to deliver real productivity gains if systems cannot navigate internal workflows or organizational permissions.
"Enterprise AI stalls on missing organizational context rather than model intelligence limits."
- Ali Ghodsi, The a16z Show
Ghodsi emphasized that unmonitored agentic workflows also create massive financial waste. Databricks found that running identical model versions on different software harnesses generated a twofold cost difference. Without external harnesses setting strict budget limits, multi-agent workflows executing over days consume vast corporate token allocations without delivering actionable business outcomes.
These operational hurdles challenge Silicon Valley's narrative of rapid, self-improving superintelligence. Ghodsi pointed out that recursive self-improvement requires training costs and compute times to shrink superlinearly, whereas frontier model development operates in exact reverse. Building state-of-the-art models now costs upwards of $10 billion, with hardware limits acting as a hard governor. Acharya similarly observed that the market is bifurcating into expensive frontier models for unbounded problems and cheap open-weight models for bounded corporate tasks.
Until enterprise platforms bridge the gap between model prompts and organizational context, massive corporate capital expenditure will keep producing prototypes that never launch.
Source Intelligence
- Deep dive into what was said in the episodes
Databricks CEO on AI Pacing, Cyber Risk, and the Enterprise • Sep 18
- Ali Ghodsi states that enterprises fail to realize productivity gains because frontier models lack organizational context. Databricks resolves this by building a digital graph, or ontology, representing internal workflows and permissions.
- Ali Ghodsi warns that businesses must implement budget controls to prevent wasteful token consumption. Databricks discovered that running the exact same model version on different software harnesses can generate a twofold cost difference.
Also discussed on this episode: (7)
Safety (4)
- Ali Ghodsi asserts that the existential risk of AI is close to zero. Ali Ghodsi argues that tech leaders behave irresponsibly by fueling public panic over humanity-ending scenarios.
- Martin Casado criticizes the concept of pacing AI development as a weak public relations compromise. Martin Casado argues that slowing down is functionally impossible because market competition forces companies to race for commercial advantage.
- Ali Ghodsi outlines four simultaneous criteria required for dangerous recursive self-improvement, which are currently unfulfilled. Instead, frontier training runs are growing more expensive, brittle, and resource-heavy.
- Ali Ghodsi identifies automated cyber warfare as the primary immediate threat from AI. The time window between the publication of a software vulnerability and its active weaponization has collapsed from years to minutes.
Open Source (1)
- Sarah Wang reports that open source models represent over sixty percent of total token volume but command only five percent of market spend by dollar.
Agents (1)
- Ali Ghodsi notes that Neon and Lakebase dominate the database market for AI agents by focusing on lightweight branching and sub-second startup times. Over ninety percent of databases on Neon are now created by agents.
Models (1)
- Ali Ghodsi highlights successful real-world AI integrations, including Crisis Text Line for suicide prevention and Merck's Teddy model for cellular drug discovery.
Why Companies Are Becoming a Series of Loops | Anish Acharya on Lenny’s Podcast • Sep 12
- Anish Acharya argues that most business problems are limited by physical and operational constraints rather than pure intelligence. Consequently, superintelligent AI will not allow single players to exponentially dominate traditional industries like logistics or food delivery.
- Anish Acharya views modern company building as a cascading series of automated loops run by AI agents. Humans remain critical to provide out-of-distribution thinking and navigate the strategic plateaus that sit between local maxima.
- Anish Acharya notes that corporate AI integration is shifting from individual tool usage to complete organizational redesign. Historical precedents show major technological shifts take decades to fully restructure business environments.
- Anish Acharya predicts a split between high-cost frontier models for high-upside, unbounded problems like drug discovery, and cheap open-weight models for bounded, verifiable tasks like legal and accounting.
Also discussed on this episode: (8)
Society (1)
- Anish Acharya dismisses fears of a permanent AI underclass as a Silicon Valley dark fantasy. He notes that the current AI stack is highly decentralized with dozens of active players, preventing the winner-take-all centralization of the mobile era.
Models (2)
- Anish Acharya rejects the idea that AI models are commodities, arguing they have distinct behavioral shapes. He uses Qwen 3-8 Max for creative video generation and GLM-53 for precise, analytical product work.
- Anish Acharya uses AI music models alongside his thirty-year DJ practice. He predicts generative models will democratize music creation, shifting the medium from broadcast back to active composition and expanding the industry's total scale.
Startups (3)
- Anish Acharya argues that consumer AI opportunities lie in entertainment and human connection rather than productivity. The industry must design new interfaces because chat only suits high-agency users, leaving room for a format between chat and TikTok.
- Anish Acharya claims that defensible moats are typically discovered through execution rather than designed in initial business plans. Early execution, high user engagement, and capturing user reasoning traces eventually yield compounding advantages.
- Anish Acharya recommends several foundational texts on organizational strategy, scale economics, and cultural development to help founders navigate market dynamics and team psychology.
Coding (1)
- Anish Acharya advises product builders to ship small, low-stakes projects weekly to build genuine intuition for model capabilities. This practice treats building as an exploratory activity rather than a pressured commercial outcome.
Health (1)
- Anish Acharya highlights healthcare and education as major opportunities for AI-driven deflation. Automating administrative tasks in healthcare can significantly lower costs, while AI tutors can unbundle learning from traditional status credentials.
90% of AI prototypes never reach production (w/ Temporal's Samar Abbas) | AI Basics • Sep 15
- Samar Abbas states that 90% of AI prototype ideas die at the proof-of-concept stage and never reach production due to instability and brittleness. This gap separates simple local experiments from reliable, production-ready enterprise software.
- Samar Abbas notes that as AI agents transition to production, they execute over longer durations and become highly asynchronous. This shift exposes traditional distributed system challenges like reliability and scalability, making laptop-based execution unsustainable.
- Samar Abbas argues that the industry is moving from an MS-DOS era of local agents to a cloud environment requiring external harnesses. These harnesses serve as a centralized brain, coordinating multi-agent loops and tool runs lasting days.
- Samar Abbas asserts that enterprises cannot safely adopt AI agents until they move execution off individual laptops into distributed team environments. Security, sandbox memory, and organizational coordination require shifting from isolated tasks to multi-user corporate processes.
Also discussed on this episode: (3)
AI Infrastructure (1)
- Samar Abbas warns that system crashes during multi-step AI tasks, like issuing a refund before notifying the customer, can lead to costly double-processing. Temporal addresses this by automatically tracking state without requiring custom developer code.
Agents (2)
- Samar Abbas highlights that Temporal's transactional engine provides complete visibility into agentic actions. This allows developers to monitor LLM queries, tool calls, and execution progress in real time rather than keeping the operations completely abstracted.
- Samar Abbas cautions that code mode agents, which generate and execute code at runtime, create severe security risks for enterprises. Temporal mitigates this danger by intercepting commands through workflows and activities to apply protective guardrails.