Replacing Microsoft 365 with a Sovereign AI Platform

The Challenge

We are an IT consultancy that builds infrastructure for regulated industries — healthcare, legal, government. Our clients need to know their data never leaves their control. But our own operations were running on the same SaaS platforms we were advising clients to move away from.

Microsoft 365 handled our email, documents, and collaboration. GitHub hosted our code. Slack ran our communications. Every tool was a US-headquartered company subject to the CLOUD Act, storing data on infrastructure we did not control.

Beyond the credibility gap, there were practical problems. Per-seat licensing costs scale linearly with team growth. AI features in M365 Copilot require sending data to Microsoft's models. And every SaaS tool added another subprocessor to our GDPR data processing register — another vendor to audit, another DPA to maintain, another point of failure in our compliance chain.

We needed to eat our own cooking. If we were going to sell sovereign infrastructure to healthcare providers and law firms, we needed to run it ourselves first.

What We Built

We designed a four-layer platform that replaces every Microsoft 365 service with self-hosted, open-source alternatives — then extends beyond what M365 offers with AI agents, intelligence gathering, and operations monitoring.

Layer 1: Infrastructure

Bare-metal servers on EU-based providers, provisioned with Terraform and configured with Ansible. Kubernetes (RKE2) runs workloads. Ceph provides distributed storage. WireGuard encrypts all inter-node traffic. Every component is infrastructure-as-code — the entire platform can be rebuilt from a single command.

Layer 2: Management Hub

Identity and access management via Zitadel (replacing Azure AD). Secrets management via OpenBao (replacing Azure Key Vault). GitLab for source control and CI/CD (replacing GitHub). Prometheus, Grafana, and Loki for monitoring and logging. Velero for backups with Ceph snapshots.

Layer 3: Workloads

This is where the M365 replacement lives, alongside services that go far beyond what Microsoft offers:

Layer 4: AI Platform

Open-source language models run on our own GPU hardware. LiteLLM provides a unified API gateway with model routing, cost tracking, and fallback chains. AI agents handle tasks across the platform — from code review to compliance monitoring to competitive intelligence — without any data leaving our infrastructure.

This is the critical differentiator. Microsoft Copilot sends your data to Microsoft's models. Our AI agents run on models we host, on servers we control, in data centres we chose. The inference happens on the same infrastructure as the data. Nothing leaves.

Multi-Tenant Architecture

The platform is designed from the ground up for multi-tenancy. Each client gets hard isolation: separate Kubernetes namespaces, tenant-scoped identity, and network policies that keep workloads apart. A central hub provides shared infrastructure (identity, monitoring, secrets), while dedicated spokes run client workloads.

This is not a theoretical design. It is the architecture running our own operations today, with the encryption and key-custody model described honestly above. When we deploy for a client, we clone the same platform with their isolation parameters. The internal deployment is the production reference and the quality gate — if it works for us, it works for them.

Results

Cost

• £0 per-seat licensing. No Microsoft 365 subscriptions, no per-user identity fees, no per-mailbox charges. Infrastructure cost is fixed regardless of team size.
• GPU ownership beats API billing. Running our own inference hardware costs a fraction of per-token API pricing at our usage levels, and the cost does not scale with query volume.
• Predictable infrastructure spend. Bare-metal servers on committed pricing with no surprise egress charges or tiered storage fees.

Compliance

• Built for GDPR from the start, not retrofitted. Data residency is EU (Hetzner, Finland and Germany). No US CLOUD Act exposure. No subprocessor chain to manage for core operations.
• DSPT-ready infrastructure. Full audit trail control. Asset register generated from infrastructure-as-code. All 10 NDG data security standards addressable with first-party evidence.
• Operator-managed encryption via OpenBao. Tenant data is encrypted in transit and at rest with tenant-scoped keys. Genitco retains key custody today; customer-held-key (BYOK) encryption is roadmapped, not yet built.

Operational

• Full infrastructure-as-code. The entire platform deploys from a single command. New environments — client tenants, acceptance clusters, disaster recovery — are provisioned in hours, not weeks.
• AI agents integrated at every layer. Code review, security monitoring, competitive intelligence, and operations triage all run as AI-powered workflows on our own infrastructure.
• No vendor lock-in. Every component is open-source with standard APIs. If a better tool emerges, we swap it. No migration fees, no data export negotiations, no contract negotiations.

What This Means for Clients

When a healthcare provider, law firm, or government contractor asks us to build their infrastructure, we are not showing them a PowerPoint. We are showing them the production system we use every day. The monitoring dashboards are real. The backup recovery procedures are tested. The compliance controls are battle-hardened.

Client deployments start from this reference architecture and are customised to their specific regulatory requirements. A healthcare client would get DSPT-aligned controls. A legal client would get infrastructure designed to support SRA obligations around document handling. A government client would get infrastructure designed to support Cyber Essentials Plus accreditation. The platform adapts — the security fundamentals do not change.

The result is infrastructure that the client owns, on servers they choose, with data that never leaves their control. No per-seat fees that punish growth. No AI that sends their data to a third party. No vendor that can be compelled by a foreign government to hand over their information.

Technology Stack