Amazon Revealed 90% of AI Agent Prototypes Failed
The company overhauled its AI infrastructure after struggling to move early experimental agent designs into production.
Updated on Sept. 24, 2026 in Artificial Intelligence
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 built in 2024 failed to reach production. The internal retrospective highlighted significant scaling challenges as only 17% of organizations currently succeed at deploying AI agents.
Why it matters
The high failure rate underscores why many enterprises struggle to measure ROI for AI agents, as poor governance and misaligned goals often stall development. Amazon responded by consolidating its agent infrastructure onto Bedrock and a new deterministic control layer.
Amazon's Kiro coding tool, used by 100,000 engineers, reduced file-reading waste by 83% through self-diagnosis. Separately, a team of six developers re-engineered the Bedrock platform in 76 days to improve security and operational stability.
The players
Amazon
A cloud computing and e-commerce giant that provides infrastructure, development tools, and AI services via AWS.
Swami Sivasubramanian
A vice president at Amazon who oversees technical strategy and AI development initiatives.
The details
To govern agent tool calls, Amazon implemented a deterministic layer called a box, which acts as a constrained boundary for system actions. By moving agent hosting services onto a single path approved by security, the company reduced the complexity that previously hindered deployment. The development team expanded from six engineers during the re-engineering phase to 35 as the system scaled.
Timeline
In 2024, Amazon teams built early AI agent prototypes.
On September 23, 2026, Swami Sivasubramanian spoke at HumanX in Amsterdam.
The Tech Race
Amazon is contending with an industry-wide 17% success rate for organizational AI agent deployment. Its consolidation strategy mirrors a broader enterprise race to standardize infrastructure and prove measurable ROI.
The transition to Bedrock and the Kiro coding tool affects how Amazon engineers build software internally, with 39,000 users adopting Kiro Crew within 30 days. For external organizations, the 80% adoption rate of Bedrock among Fortune 100 companies suggests this platform will likely become the primary baseline for enterprise AI workflows.
The takeaway
Amazon's struggle highlights that scaling AI requires more than just models; it demands rigid governance layers and consolidated infrastructure. Watch for future benchmarks related to the 80% adoption rate among Fortune 100 companies to confirm if this consolidation model improves industry-wide deployment.
Further reading
For broader trends in enterprise implementation, see the Artificial Intelligence section.
Live Poll
Do you trust that businesses are currently getting measurable value from their investments in AI?
