Samar Abbas warns 90 percent of AI prototypes fail
- Nine out of ten enterprise AI prototypes fail before reaching actual production.
- Distributed cloud harnesses are replacing developer laptops to stop autonomous agents from crashing.
- Companies are ditching heavy text generators for high-speed judgment models to cut costs.
Enterprise AI has hit a technical brick wall.
On September 15, 2026, Temporal CEO Samar Abbas revealed on This Week in Startups that 90 percent of enterprise AI prototypes die at the proof-of-concept stage. While autonomous coding agents make assembling an initial demo effortless, those applications collapse under real-world scale. Long-running, asynchronous multi-agent workflows break when basic infrastructure failures wipe out state mid-task, forcing developers to write endless custom recovery code.
"The model predicts the next token, but the harness protects the business."
- Samar Abbas, This Week in Startups
To survive, enterprise teams are moving off developer laptops into sandboxed cloud environments guarded by external harnesses. Abbas argued that local runtimes become operational liabilities when agents generate and execute arbitrary code on the fly. External harnesses act as control brains, intercepting commands generated by language models before execution to prevent erratic agent loops from derailing business operations.
The high failure rate is also driving a shift away from bloated text generators toward fast, specialized alternatives. On The AI Daily Brief the following day, host Nathaniel Whittemore detailed how Typesafe launched Jev, a probabilistic judgment model developed by ChatGPT co-inventor Diego Almeida. Instead of generating long prose responses that crash deterministic software pipelines, Jev outputs numeric scores for classification tasks up to 200 times faster and 400 times cheaper than standard large language models.
Data fragmentation presents an equally daunting barrier to production deployment. Speaking on The a16z Show that same day, Lightfield CEO Keith Peiris explained why he walked away from Tome, an AI presentation tool with 25 million users, to build an AI-native system of record. Peiris discovered that enterprise models fail when context is trapped across disconnected call logs, emails, and rigid relational database schemas. Lightfield replaces traditional database tables with semi-structured chronological activity logs analyzed directly by models.
"Relational database schemas constrain modern customer records."
- Keith Peiris, The a16z Show
Software business models are fracturing alongside the technical stack. Peiris noted that traditional per-seat pricing fails when power users consume thousands of times more compute, while pure consumption pricing creates cost anxiety that stifles user activity. Enterprise software vendors are turning to hybrid models with flat platform fees paired with consumption credits for automated workflows to keep deployments viable.
Building an AI prototype takes hours, but building production infrastructure remains a battle of endurance.
Source Intelligence
- Deep dive into what was said in the episodes
Nathaniel Whittemore
Why a New Class of AI “Judgment Models” Could Have Big Business Implications • Sep 16
- Diego Almeida claims Jev operates up to 200 times faster and 400 times cheaper than standard large language models with free output tokens. This cost-efficiency allows developers to run continuous checks across massive enterprise datasets that were previously cost-prohibitive.
- In a test of Jev's capabilities, Mike Taylor analyzed 37 documents against 21 questions to detect AI-generated writing patterns. Jev processed all 777 judgments in under 0.7 seconds at an estimated cost of a quarter of a cent.
- Matt Stockton argues Jev functions as an accessible interface for classical machine learning classification techniques. While businesses historically solved classification problems manually, Jev bypasses the complex data-labeling and hosting steps required by legacy methods.
- Nathaniel Whittemore argues that judgment models will excel in cross-team workflows by assessing organizational handoffs. These models evaluate if messages imply delivery commitments or violate established product roadmaps, helping manage complex internal responsibilities.
Also discussed on this episode: (5)
Safety (1)
- Mark Zuckerberg argued that AI safety pacing is the responsibility of individual labs driven by market liability and alignment incentives, rather than government-enforced slowdowns. To support this, Meta delayed releasing its Muse model by several months to prioritize safety.
Regulation (1)
- Bernie Sanders and Steve Bannon formed an unlikely populist alliance at the Future of Life Institute's Washington assembly to demand human-centric AI regulation. Both figures framed the technology's rapid deployment as a battle for agency against unchecked corporate oligarchs.
AI Infrastructure (1)
- Journalist Jasmine Sun reported that local opposition to AI data centers is primarily driven by a perceived loss of community agency over the future. This matches the broader populist shift toward framing artificial intelligence as a battle over civic control.
Enterprise (1)
- Salesforce launched KOA, a specialized vertical CRM model fine-tuned from NVIDIA's Neotron, alongside AI Force connectors that allow third-party agents to access enterprise data. The moves signal a transition toward headless, platform-agnostic software architectures.
Models (1)
- Diego Almeida co-created a new AI training method called Reinforcement Learning for Calibrated Decisions to power Jev, a new model architecture. Unlike traditional language models, Jev outputs precise probabilities rather than text to automate specific workflow decisions.
The AI-Native CRM • Sep 16
- Keith Parris pivoted away from Tome despite the presentation tool reaching 25 million total users and two million monthly active users. Parris realized LLMs lacked the contextual relationship data required to generate indispensable, professional-grade presentations.
- While running 12 pilots with enterprise sales teams, Keith Parris discovered that the primary obstacle to AI agent automation was fragmented and conflicting data sitting across legacy CRMs, call recorders, and databases.
- Leveraging their experience at Facebook, Lightfield founders modeled their CRM architecture after the chronological Facebook timeline. The system builds a canonical activity log of interactions to infer business changes and trigger automated updates.
- Alex Rampell argues that modern systems of record must prioritize intelligence over schema. Lightfield operates as a semi-structured database, allowing users to ingest unstructured emails and database records first and define structural fields afterward.
- Healthcare marketplace Power utilizes Lightfield's arbitrary schema to scrape clinical trials and match Alzheimer's patients with experimental treatments. Keith Parris notes the system completed a successful matching process within days of deployment.
- Keith Parris rejected pure seat pricing because power users consumed too much compute, and pure consumption pricing because cost anxiety halted user activity. Lightfield now uses a hybrid seat fee for core features and consumption pricing for automations.
Also discussed on this episode: (3)
Startups (2)
- To seed his early-stage CRM, Keith Parris offered startups free office space in exchange for using the unfinished software. This negative-pricing hack secured 10 active startup customers who provided constant product feedback via Slack.
- To prevent the organizational inertia that slowed Tome, Keith Parris structured Lightfield's 40-person team without strict functional departments. All employees attend a single daily standup and pull tasks dynamically from a continuously edited stack rank.
Enterprise (1)
- Keith Parris counters incoming VPs of Sales who demand Salesforce by distributing Lightfield to engineering, finance, and customer success teams for free. This strategy builds cross-departmental network effects that prevent the CRM from being easily replaced.
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.