The pursuit of reliable artificial intelligence isn’t about hitting arbitrary milestones like “90% accuracy.” It’s about understanding that each additional percentage point of dependability requires comparable effort, and often, a fundamental shift in how systems are built. As Andrej Karpathy succinctly puts it: “When you get a demo and something works 90% of the time, that’s just the first nine.” This isn’t a bug; it’s a core property of complex systems.
The Compounding Cost of Each “Nine”
The “March of Nines” describes how quickly diminishing returns set in. A basic AI demonstration might achieve 90% reliability with relative ease. But reaching 99%, and then 99.9%, demands orders of magnitude more engineering work. This is particularly critical in enterprise environments where even minor failures can trigger substantial business risk.
Why this matters: AI systems often operate in high-stakes scenarios where even small error rates can lead to financial losses, regulatory violations, or reputational damage. A 1% failure rate on a billion-dollar transaction is still a significant problem.
From “Usually Works” to “Dependable Software”
Many teams mistakenly focus on model accuracy while neglecting the broader system reliability. A typical enterprise workflow involves multiple steps – intent parsing, data retrieval, planning, tool execution, validation, and logging. If any step fails, the entire workflow collapses.
For example, a 10-step workflow with each step succeeding at 95% probability has an overall success rate of only 59.9%.
Defining Reliability with Measurable SLOs
The key to achieving higher reliability isn’t just better models, it’s turning dependability into measurable objectives. Teams must define Service Level Indicators (SLIs) and then invest in controls that reduce variance. Essential SLIs include:
- Workflow completion rate: The percentage of workflows that succeed or escalate gracefully.
- Tool-call success rate: The percentage of tool executions that complete within specified timeouts, with strict schema validation.
- Schema-valid output rate: The percentage of structured responses that conform to predefined schemas.
- Policy compliance rate: The percentage of outputs that adhere to security, privacy, and regulatory constraints.
- Latency and cost per workflow: Ensuring performance within acceptable bounds.
- Fallback rate: The effectiveness of fallback mechanisms when primary systems fail.
Once defined, these SLIs should be used to set concrete Service Level Objectives (SLOs) and manage an error budget to ensure experiments don’t derail reliability.
Nine Proven Levers for Enhanced Reliability
- Constrain Autonomy with Explicit Workflows: Limit AI’s freedom by defining strict state machines or directed acyclic graphs (DAGs) where each step has defined inputs, outputs, and failure handling. Use idempotent keys for safe retries.
- Enforce Contracts at Every Boundary: Use schemas (JSON Schema, Protobuf) to validate all structured data. Normalize units (ISO-8601, SI) and enforce strict data types to prevent interface drift.
- Layer Validation Checks: Beyond syntax, implement semantic and business-rule checks to prevent logically plausible but system-breaking outputs.
- Route by Risk Using Uncertainty Signals: Use confidence scores or secondary verification models to direct high-impact actions through stronger assurance paths.
- Engineer Tool Calls Like Distributed Systems: Apply timeouts, backoff strategies, circuit breakers, and concurrency limits to external dependencies. Version tool schemas to prevent silent failures.
- Make Retrieval Predictable and Observable: Treat data retrieval as a versioned product with metrics for coverage, freshness, and hit rates. Use canaries for index changes.
- Build a Production Evaluation Pipeline: Maintain a golden set of production traffic to run against every change. Use shadow mode and A/B canaries with automatic rollback on regressions.
- Invest in Observability and Operational Response: Emit detailed traces, store redacted prompts, and classify failures into a clear taxonomy. Use runbooks and safe-mode toggles for rapid mitigation.
- Ship an Autonomy Slider with Deterministic Fallbacks: Design AI systems with adjustable autonomy levels and deterministic fallbacks (retrieval-only answers, cached responses, human review) to ensure safe operation.
The Enterprise Reality: Reliability Drives Adoption
A recent McKinsey report shows over half of organizations using AI have experienced negative consequences due to inaccuracy. These risks force enterprises to prioritize stronger measurement, guardrails, and operational controls. The later “nines” aren’t just technical goals, they’re business imperatives.
Ultimately, achieving true AI reliability requires disciplined engineering: bounded workflows, strict interfaces, resilient dependencies, and rapid learning from failures. It’s not about avoiding mistakes; it’s about minimizing their impact and responding effectively when they occur.
