Key points
- Airservices Australia is exploring a “broker platform” within its AWS tenancy to act as a single governed entry point for all AI interactions.
- The AI broker would initially be trialled by around 100 users before expanding to more than 3000 users, per a request for information.
- The platform must run in AWS ap-southeast-2 using Amazon Bedrock as its primary model gateway, with audit trails retained for at least seven years.
Airservices Australia is exploring options for a “broker platform” within its AWS tenancy that can act as a “single governed entry point for all AI interactions” across the organisation.
The government-owned organisation responsible for air traffic control in Australia indicated it wants to stop the “proliferation of unmanaged point solutions” as AI adoption gains pace internally.
The AI broker would initially be trialled by around 100 users, before being expanded to 3000-plus users, Airservices indicated in a request for information.
“As part of its digital transformation strategy, Airservices seeks to consolidate and govern the use of artificial intelligence across its workforce behind a single enterprise-grade platform,” it said overnight.
“[The broker] acts as the organisation’s ‘AI front door’, replacing fragmented, unmanaged AI tool usage with a secure, auditable, and scalable capability.”
The platform ideally would present a “unified conversational AI interface for staff productivity, knowledge retrieval, and task automation”, according to the organisation.
It would also “support multi-model orchestration (routing requests to the most suitable foundation model based on task characteristics), integrate with corporate systems via governed connectors, and provide administrators with full visibility of usage, cost, behaviour, and compliance,” the request for information states.
Technically, Airsevices wants the broker to be deployed within AWS ap-southeast-2 and “Airservices-controlled accounts”.
Amazon Bedrock is to act as the “primary model gateway” for accessing AI foundation models.
The requirements list comes with some specific performance goals. Notably, these include a latency service-level objective on time-to-first token, marking the time between an AI prompt and the first part of the response being returned.
There is also a strong auditability component, with Airservices wanting “all AI interactions, tool invocations, and administrative actions must be logged with immutable audit trails retained for a minimum of seven years.”
