All insights

Article

Where Your Code Goes When an AI Developer Runs in Your Own Azure

The first question a head of engineering asks about an AI developer is not about speed. It is about location: where does the code sit, who can reach it, and what leaves the building. It is the right question, because a tool that writes software sees more of your codebase than most of your staff do.

Our answer is structural rather than reassuring. We install the Gyzer AI Developer Platform in your own Microsoft Azure and maintain it. Nothing runs on Gyzer's infrastructure. Your organisation holds every account, and you pay for the Azure and AI usage directly.

What "installed in your Azure" actually means

The platform runs as infrastructure inside the Azure subscription your organisation controls. The AI identities it uses, the work queues it reads from, the private previews it builds and the test databases behind those previews all live in that subscription.

That has three practical consequences, and they are the reason we build it this way.

Your security team can inspect the configuration with the tools they already use. They do not have to accept an architecture diagram; they can read the resource groups, the role assignments and the network rules.

Access can be revoked by you, not requested from us. If the engagement ends, the accounts are yours to close.

Audit evidence sits where your other audit evidence sits. The logs and records are in your tenant, subject to your retention rules, rather than in a supplier's dashboard that disappears at the end of a contract.

What stays with you

You hold the Azure account. You hold the AI service account. You hold the code repository, the pipelines, the test records and the logs. We hold the platform software and the responsibility for maintaining it. Gyzer Technologies holds ISO 27001 certification.

We are deliberately precise about what we claim. You hold every account, and nothing runs on Gyzer's infrastructure. If a security reviewer wants to know what is sent to a model and under what terms, the answer comes from your own configuration and your own agreement with the model provider, which is a stronger position than a supplier's reassurance.

The limits that apply in your tenant

Installing it in your environment does not change what the AI is allowed to do. The limits travel with the installation. Inside your subscription it still never touches production, never crosses between products and never reaches the open internet, and those three are enforced by configuration your own team can read. The remaining four are in the seven limits an AI developer must never break. Every action is logged against its story, with its cost.

Read those together with the tenancy point and you get the shape of the answer a risk committee wants. The AI acts inside your subscription, under your identity model, inside a network boundary you can inspect, on a work queue that is separated per product, with a cost ceiling per story and a log of everything it did.

The question your team will ask second

After location comes the one about the code itself: what does the model see? The platform works from the repository it has been given, on the product it has been assigned to. It does not browse, it does not fetch, and it does not carry anything from one product's work into another's, because each product has its own identity and its own queue.

That separation is worth testing rather than accepting. Ask to see two products configured side by side and check that the identities are distinct, the queues are distinct and the network rules are distinct. It takes ten minutes and it tells you more than an architecture slide. Three organisations run it today inside their own subscriptions, 678 stories delivered, two of them healthcare software teams of six and eight developers.

What your team needs before install

Four things, all of which most Azure estates already have: an Azure subscription the platform can be deployed into, the ability to create service identities and role assignments, network rules that permit an approved list of addresses, and a repository and pipeline arrangement the platform can integrate with.

Installation takes two weeks. The first story runs in week three.

What changes for your Product Owners and reviewers

Nothing about the roles. A Product Owner writes the user story with acceptance criteria. The platform writes the tests, writes the code and builds the private preview. A Code Reviewer approves every change, and a person releases to production.

What changes is where they do it. They work inside your environment, against your repositories, with your standards. The AI does the writing; your people keep the judgement, which is the part that has to stay in your organisation.

What to do next

What the platform may do inside those boundaries is the other half of the conversation, and the story log an auditor reads is what your security architect will ask for. Bring your security architect to the assessment. The useful conversation is not about whether AI writes code, it is about which subscription, which identities, which network rules and which logs. Those are answerable in thirty minutes.

Where Your Code Goes When an AI Developer Runs in Your Own Azure | Gyzer