The March of Nines: Why AI Reliability is a Decade-Long Challenge

The pursuit of reliable artificial intelligence is often framed as a series of incremental improvements. However, Andrej Karpathy, co-founder of OpenAI and former Tesla AI Director, cautions that achieving true dependability in AI systems is far more complex than simply increasing accuracy. He describes this challenge as the “March of Nines,” where each successive increment in reliability – moving from 90% to 99%, then to 99.9%, and beyond – requires a comparable amount of engineering effort to all that came before.

The Compounding Math of Reliability

The “March of Nines” highlights a critical reality: a strong demo showcasing 90% reliability is merely the starting point. As Karpathy explains, “Every single nine is the same amount of perform.” This is particularly true for complex, agentic workflows, which involve multiple steps such as intent parsing, context retrieval, planning, tool calls, validation, formatting, and audit logging.

The probability of end-to-end success in these workflows compounds with each step. For example, a 10-step workflow with each step having a 90% success rate results in an overall success rate of only 34.87%, leading to a 65.13% failure rate.¹ Correlated outages, such as authentication issues or rate limits, can further exacerbate these failures unless underlying dependencies are hardened.

Here’s a breakdown of the diminishing returns:

• 90.00% per-step success: 34.87% overall success, ~6.5 interruptions/day (for 10 workflows/day)
• 99.00% per-step success: 90.44% overall success, ~1 interruption every 1.0 days
• 99.90% per-step success: 99.00% overall success, ~1 interruption every 10.0 days
• 99.99% per-step success: 99.90% overall success, ~1 interruption every 3.3 months

From Demo to Dependable Product

Karpathy’s observations stem from his experience at Tesla, where early self-driving demos appeared nearly flawless, yet the path to a production-ready system was significantly more challenging.² Similarly, today’s AI systems, including chatbots and coding agents, can perform impressively in controlled environments but struggle with the unpredictable complexities of real-world scenarios.

Nine Levers for Achieving Higher Reliability

To move beyond promising demos and build truly dependable AI systems, teams must focus on measurable objectives and invest in controls that reduce variance. Here are nine key strategies:

1. Constrain Autonomy: Utilize explicit workflow graphs with bounded states and deterministic handling for retries and timeouts.
2. Enforce Contracts: Implement strict validation of data formats (e.g., JSON Schema) at every interface.
3. Layer Validators: Go beyond syntax checks to include semantic and business rule validation.
4. Route by Risk: Utilize uncertainty signals to direct high-impact actions through more robust verification processes.
5. Engineer Tool Calls: Treat tool calls like distributed systems, incorporating timeouts, backoff mechanisms, and circuit breakers.
6. Make Retrieval Predictable: Treat data retrieval as a versioned data product with comprehensive coverage metrics.
7. Build a Production Evaluation Pipeline: Establish a system for quickly identifying and preventing regressions.
8. Invest in Observability: Implement robust monitoring and logging to facilitate rapid diagnosis and remediation of failures.
9. Ship an Autonomy Slider: Provide a mechanism to adjust the level of autonomy and offer deterministic fallbacks.

The Enterprise Imperative

The demand for higher reliability is driven by the business risks associated with AI failures. A McKinsey survey found that over half of organizations using AI have experienced negative consequences, with nearly a third attributing those consequences to inaccuracies.¹

Looking Ahead

Achieving the later “nines” of reliability requires a shift from simply building larger models to embracing disciplined systems engineering. By focusing on bounded workflows, strict interfaces, resilient dependencies, and rapid operational learning, organizations can move closer to realizing the full potential of AI.