Healthcare network moves AI from pilot to production with enforced guardrails
Healthcare · ~1,200 employees, highly regulated · Switzerland / EU · On-premises deployment in 5 weeks; 7 pilots in production within 2 quarters
Short answer
A ~1,200-employee healthcare provider network moved seven blocked AI pilots into production within two quarters by deploying bluebill on-premises with runtime PII redaction, prompt-injection defence and complete audit logging mapped to EU AI Act and ISO/IEC 42001 requirements.
7
blocked pilots moved into production
5 weeks
to on-premises deployment
0
requests leaving the network boundary
100%
of model traffic covered by logged guardrail decisions
The challenge
- Seven AI pilots had passed clinical usefulness review but failed data-protection review.
- Patient data could not leave the network, ruling out unmediated use of public model APIs.
- There was no runtime enforcement — policies existed as documents, not as controls.
- Audit preparation was a manual exercise nobody wanted to repeat quarterly.
What bluebill did
- Deployed the control plane on-premises so no request or log leaves the network boundary.
- Enforced guardrails at four layers: input validation, data redaction, output checks and tool/action limits.
- Redacted patient identifiers before any model call, with the redaction decision itself logged.
- Scoped agent permissions to explicit tool allow-lists with budget ceilings per use case.
- Mapped every logged control to EU AI Act and ISO/IEC 42001 evidence requirements so audit export is a query, not a project.
“The pilots were never the hard part. Proving what happens to the data was.”
— CISO, anonymised healthcare provider network
What changed afterwards
- Data-protection review moved from a blocker to a checklist, because the controls are demonstrable at runtime.
- New AI use cases are approved against an existing control set instead of starting a fresh review each time.
- Audit evidence is exported on demand rather than reconstructed quarterly.
Environment: On-premises / air-gapped deployment · Self-hosted open models · Existing IAM · Audit log export
Questions this engagement answers
Can AI governance work in an air-gapped environment?
Yes. This network runs the control plane on-premises with self-hosted models, so no request, prompt or log leaves the boundary while guardrails and audit logging still apply in full.
What evidence does an AI audit actually need?
A model inventory, retention-bounded prompt and response logs, guardrail decisions including what was redacted or blocked, and an approval history for model changes. Producing these as a by-product of normal traffic is what removed the quarterly manual effort here.
Client names are withheld by agreement. Engagement profiles are anonymised; figures are the measured results of the engagement described and are consistent with bluebill's aggregate programme results. Individual results vary with environment, scale and starting maturity.