Amazon Bedrock in Healthcare: Every Token on a Budget
In two weeks, a healthcare organization gave its teams 13 Anthropic and Mistral models on Amazon Bedrock in the EU, with SSO and token budgets per project.
Challenge
Employees at a healthcare organization wanted to use generative AI in their daily work. Without an approved platform, they turned to personal subscriptions and API keys. IT had no visibility into which data reached which model, and finance could not attribute AI spend to projects. For an organization that handles health data, this was not sustainable. The organization needed a platform that processes data in the EU, uses existing corporate identities, and gives every project owner control over costs.
Approach
The organization engaged StackVera, an AWS Partner, for an Architecture Scoping engagement: a proof of concept with a fixed end date two weeks out. StackVera made Amazon Bedrock the single entry point to 13 Anthropic and Mistral models, deployed in AWS Regions in the EU with EU cross-Region inference profiles. Models that can only route globally are blocked. Amazon Bedrock does not use prompts or responses to train models. On top of Amazon Bedrock, StackVera built a user management portal. Employees sign in with the existing single sign-on and select a project. Every request is booked to that project, token by token. Each project receives a credit budget. When the budget is spent, requests stop. The project owner reviews the reason and decides whether to unblock.
Outcome
The platform went live after two weeks. Every team now reaches 13 models with one login. Each token is attributed to a project and its cost center, so finance books AI spend where it occurs. No request leaves the EU, and no project exceeds its budget without its owner's approval. Personal AI subscriptions and private API keys are no longer needed, and the organization can add new models as Amazon Bedrock makes them available in the EU.