Pillar - AI Governance
Governance your platform team can actually deploy
Most AI governance advice stops at a policy document. This is the other half: the platform controls that make the policy true, and the evidence that proves it to an auditor.
The gap most enterprises are standing in
Two things happened at once. Teams shipped generative features quickly, often calling model endpoints directly with keys in config files, and the regulatory picture moved underneath them. The result is a platform nobody can inventory and obligations nobody has mapped.
The thing most teams have wrong
The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) deferred the high-risk obligations - it did not defer the transparency duties. Article 50 took effect on 2 August 2026 and applies by what a system does, not by which risk tier someone assigned it. A chatbot, a summariser or a generated-image feature that nobody ever classified as high-risk is in scope today.
Where the deadlines actually sit
| Obligation | Applies from | Status |
|---|---|---|
| Prohibited practices (Art. 5) | 2 Feb 2025 | In force |
| GPAI provider obligations (Art. 51–56) | 2 Aug 2025 | In force, unaffected by the Omnibus |
| Transparency duties (Art. 50) | 2 Aug 2026 | In force |
| Art. 50(2) machine-readable marking, generative systems already on the market before 2 Aug 2026 | 2 Dec 2026 | Grace period |
| New Art. 5 prohibition: non-consensual intimate imagery and CSAM generation | 2 Dec 2026 | Added by the Omnibus |
| High-risk - Annex III standalone | 2 Dec 2027 | Deferred from Aug 2026 |
| High-risk - Annex I embedded | 2 Aug 2028 | Deferred |
What I build
An AI landing zone is, concretely, an API gateway sitting between your teams and your model endpoints, plus the plumbing that turns it into evidence:
- One front door. API Management in front of Azure OpenAI and any other provider, so there is exactly one place where policy is enforced and one place to look when someone asks what was called.
- Model allowlists per team. Which models a product group may call, at which versions - a technical control, not a wiki page.
- Token quotas and chargeback. Per-team budgets with usage ingested into a pipeline finance can read, so spend stops being a month-end surprise.
- Prompt and response logging with the retention and redaction your DPO agreed to - the substrate for both incident response and Article 50 evidence.
- Tenant and network isolation. Private endpoints, managed identity end to end, no standing keys.
- Control mappings to ISO/IEC 42001, DORA and NIS2, generated from the same Terraform that deploys the thing.
Why this is one person's job
The mapping and the implementation drift apart the moment they belong to different people. When the
architect who writes the control narrative is the one running terraform apply, the document
describes what is actually deployed - which is the only version an auditor cares about.
Writing on this
The long version
-
Building an AI Landing Zone on Azure — Part 5: Turning tokens into euros
Raw usage records are not a bill. This is the rate card as data, cost computed at query time, the cached/cache-write/reasoning token nuances, and the report that gave finance a monthly AI bill per cost centre.
-
Building an AI Landing Zone on Azure — Part 4: Metering every call
The usage pipeline behind the AI gateway: APIM to Event Hub to a Function App to Cosmos DB, managed identity end to end.
-
Building an AI Landing Zone on Azure — Part 3: One front door for many models
How a single, VNet-injected API Management instance becomes the only way to reach Azure OpenAI and Mistral OCR, with policy fragments as composable building blocks and an identity per hop.
-
Building an AI Landing Zone on Azure — Part 2: The platform underneath
Before the gateway and the metering there is a private AKS cluster, a hub-and-spoke network and a GitOps setup where no application ever holds a secret. This is how that foundation is put together.
-
Building an AI Landing Zone on Azure - Part 1: Why every enterprise needs an AI gateway
Teams are calling LLM endpoints directly with API keys in config files and nobody knows who is spending what. Here is the landing zone I built to fix that, and why.
Start with an assessment
Two to three weeks: an inventory of what your teams have shipped, a classification against the Act, and a mapped remediation plan your platform team can start on the following Monday.