Enterprise internal AI copilot: 9,000 employees, 1 secure stack

The situation

• A German Fortune-500 manufacturer had piloted Microsoft Copilot, ChatGPT Enterprise, and two domestic alternatives.
• None passed their security review, all required sending employee data outside their VPC.
• The CIO wanted an internal copilot for 9,000 employees with retrieval over their internal docs (4M files), tool integrations into their ERP, and audit logging, all running inside their AWS Frankfurt VPC.

What we did

• vLLM self-hosted inference inside the customer's AWS VPC, zero data egress
• Vector + keyword retrieval over 4M internal documents on OpenSearch
• Permission-aware retrieval mirroring corporate AD groups
• Read-only SAP tool integrations with role-scoped queries
• Full audit log of every prompt, retrieval, tool call, response
• Copilot UI deployed behind Azure AD SSO with per-department controls

The result

• The pilot went company-wide within 8 weeks at the board's direction. Internal employee usage stabilized at 6,400 weekly active users.
• The CIO has since used the architecture as a reference for two of the company's joint ventures. Independent third-party audit reported zero data egress events across the first 90 days.