Aokumo at AWS Bedrock LLM Day Japan
On July 28, Aokumo exhibited and presented at AWS Bedrock LLM Day Japan in Tokyo. Around 300 technology and business leaders attended, and we joined a partner lineup that included NRI, KDDI iret, and Fujisoft.
Our thanks to the AWS Japan team for the invitation and for running an excellent event.
Our five minutes
We had five minutes on stage and used them to make one point:
AI agents do not fail to reach production because the models are weak. They fail because approval, rollback, and audit controls are added afterward.
Enterprises already have monitoring, alerts, dashboards, and increasingly capable LLMs. What they often lack is a reliable mechanism for turning an agent’s proposal into a policy-governed, reversible, and audit-ready action.
Reimplementing guardrails for every agent or relying on instructions in a prompt does not survive contact with a regulated production environment.
Governance must be part of the execution architecture, not an optional feature added at the end.
We call this Governed Execution, built on four pillars:
Policy enforcement: Execution boundaries are defined structurally rather than left to the agent’s judgment.
Risk-based approval: High-risk actions require explicit human approval, while pre-authorized actions can execute autonomously within clearly defined policies.
Rollback by design: Every change includes recovery controls, with the proposed diff, blast radius, and dry-run results available before execution.
Tamper-evident audit trail: The proposal, policy decision, approval where required, execution, and verification are recorded automatically, helping teams prepare compliance evidence, including J-SOX documentation, in minutes.
What the room told us
The attendee survey shared at the event reflected what we heard repeatedly at our booth:
65% of organizations are actively pursuing generative AI, with more than 40% already working on concrete projects.
Yet 56% remain stuck at the proof-of-concept stage. Only around one-third have reached production.
The two most commonly cited barriers were engineering talent at 62% and security at 57%.
Nearly every booth conversation followed the same pattern:
The PoC works, but the production gate will not open.
The ambition is there. The technology is often ready. The bottleneck is the final mile between “the demo works” and “security and compliance approve production use.”
That final mile is exactly what Aokumo is built to address.
A next step: Free AWS security assessment
Putting AI agents into production starts with understanding the current control state of your environment.
The free AWS security assessment announced at the event is also available to readers of this blog.
Scope
The assessment covers four areas:
Identity and access management
KMS and encryption keys
Network security
Logging and threat detection
It uses the assessment engine we previously applied during a pre-release security review for a financial group.
Terms
Limited to one AWS account per company
Read-only access
No changes to production environments or applications
Applications close on August 31, 2026
Deliverables
Each participating company will receive:
An AWS security scorecard
Approximately 10 prioritized findings with attack-path explanations
A practical remediation roadmap ordered by risk and impact
Apply for the free AWS security assessment
To everyone who visited our booth and challenged us with difficult questions: thank you. Those conversations reinforced our conviction that enterprises do not need another AI agent that merely recommends actions.
