Amazon Revealed High AI Agent Prototype Failure Rate

A senior Amazon executive disclosed that 90% of early AI agent prototypes failed to reach production in 2024.

Updated on Sept. 24, 2026 in Artificial Intelligence

Isometric editorial illustration of a geometric server rack component and a solid plinth, representing enterprise AI infrastructure challenges. — Amazon reported that 90 percent of its internal AI agent prototypes failed to reach production in 2024, citing challenges in scaling autonomous tools. AI Illustration. Upload story photo >

Live Poll

Do you trust that businesses are currently getting measurable value from their investments in AI?

Amazon vice president Swami Sivasubramanian reported that 90% of the company's early AI agent prototypes failed to make it to production during 2024. The disclosure highlighted significant internal challenges in scaling autonomous tools before recent infrastructure consolidations.

Why it matters

The high failure rate highlights the widespread difficulty enterprises face in moving experimental AI agents into production environments. Common obstacles for these projects include misaligned objectives, poor governance, and a lack of clear measurement tools.

While 90% of early prototypes failed, Amazon's Kiro coding tool now supports 100,000 engineers and achieved an 83% reduction in file-reading waste. Additionally, six developers re-engineered Bedrock in 76 days, which now serves 80% of Fortune 100 companies.

The players

Swami Sivasubramanian

He serves as a vice president at Amazon and provides strategic oversight for the company's artificial intelligence initiatives.

Amazon

This technology corporation provides extensive cloud computing services through AWS and develops internal AI coding infrastructure.

The details

Amazon addressed these development bottlenecks by creating a deterministic layer known as a box to govern agent tool calls and by consolidating hosting services. The company also rolled out Kiro Crew, which gained 39,000 internal users within 30 days of its launch.

Amazon teams built their early AI agent prototypes throughout 2024.
Swami Sivasubramanian disclosed the findings on September 23, 2026.

The Tech Race

Amazon's move to standardize its agent infrastructure onto Bedrock and AgentCore follows the industry-wide 7% ROI measurement rate for AI agents. This transition marks a departure from fragmented development patterns toward a unified, security-approved path for enterprise scaling.

Enterprise customers using AWS Bedrock may see more reliable and secure AI agent integrations following the company's infrastructure consolidation. These changes focus on streamlining production-level tools, potentially accelerating the development timelines for software built on Amazon's platforms.

The takeaway

The high failure rate of early AI prototypes suggests that governance and infrastructure are as critical to success as the underlying models themselves. Organizations should focus on creating deterministic layers to manage AI behavior before attempting to scale complex agent-based workflows.