Notifications
Clear all
Topic starter
19/08/2026 3:39 pm
TL;DR: AI agents fail in production 70% to 95% of the time, with performance falling further in repeated runs and multi-agent chains, according to Fiddler's analysis of benchmark and enterprise data. The governance challenge is no longer whether agents can demo well, but whether organisations can constrain scope, verify outputs, and audit actions before failures cascade into business damage.
NHIMG editorial — based on content published by Fiddler: AI Agent Failure Rate: Why 70-95% Fail in Production
By the numbers:
- AI agents fail between 70% and 95% of the time in production environments, depending on task complexity and how success is measured.
Questions worth separating out
A: Treat agents like governed runtime actors and block any side effect until the output passes validation.
Q: Why do AI agents that succeed in demos fail so often in production?
A: Demos hide the conditions that break real workflows, including expired tokens, changing APIs, concurrent activity, and incomplete context.
Q: What breaks when AI agents are given broad standing access?
A: Broad standing access breaks governance because the agent can move from one task to another without a fresh authorization check.
Practitioner guidance
- Implement human approval gates for high-stakes agent actions Route refunds, deletions, financial transfers, and other irreversible actions to a reviewer before the agent executes them.
- Enforce output validation before side effects Use schema validation for structured outputs, assertion tests for constrained tasks, and a second-pass judge only where open-ended review is necessary.
- Instrument span-level tracing for every agent action Capture each tool call, intermediate decision, retry, and final output so you can reconstruct failures and measure drift.
What's in the full article
Fiddler's full blog post covers the operational detail this post intentionally leaves for the source:
- Benchmark-by-benchmark comparisons across WebArena, AgentBench, Carnegie Mellon, Princeton, and MIT findings
- Latency and cost tradeoffs between schema validation, assertion tests, and LLM-as-judge verification patterns
- Observability guidance for OpenTelemetry integration, token telemetry, prompt versioning, and compliance monitoring
- Practical examples of co-pilot routing, retry logic, circuit breakers, and AI registry design
👉 Read Fiddler's analysis of why AI agents fail in production →
AI agent failure in production: what controls are missing?
Explore further
19/08/2026 4:18 pm
AI agents are now governed runtime identities, not just software features. Once an agent can call tools, read data, and act across systems, it inherits many of the same governance problems associated with non-human identities. The article shows that execution risk rises when those privileges are assumed to be safe simply because the agent passed a demo. Practitioners should treat agent approval, traceability, and scoped access as baseline governance requirements.
A question worth separating out:
Q: How should security teams govern agent workflows at runtime?
A: Security teams should govern agent workflows with controls that evaluate prompts, tool calls, and outputs during execution, not only after deployment. Runtime checks matter because risk can appear at each stage of the workflow. The goal is to stop unsafe behavior before it becomes an executed action or a leaked response.
👉 Read our full editorial: AI agent failure rates expose the governance gap in production
