AI governance for development teams works when it is short, concrete and enforced by tooling: a list of approved tools and model providers, clear rules on which data may enter prompts, mandatory human review for agent-written code, restricted agent permissions and logging. Policies that only say "use AI responsibly" change nothing; policies that name tools, data classes and owners do.
On this page
Development teams adopted AI coding tools faster than most companies could write rules for them. The result in many organisations is a mix of personal subscriptions, unclear data flows and inconsistent review practices. At the same time, the same teams are asked to build AI features into products, which raises a second set of questions about models, data and accountability. Good governance addresses both without slowing developers down. This article describes the policy elements that make a difference and how to enforce them.
Two governance scopes
The first scope is AI as a development tool: coding agents, AI editors and chat assistants used to write, review and test code. The risks here are leaking source code or secrets, introducing insecure or licence-incompatible code and agents taking unintended actions in development environments. The second scope is AI as part of the product: features that call models at runtime. Here the risks concern customer data, output quality, transparency obligations and regulatory classification. One policy document can cover both, but the sections should be separate because owners and controls differ.
The policy elements that matter
An effective policy fits on a few pages and answers concrete questions:
- Approved tools and providers: which coding agents, editors and model APIs may be used, under which plans and contractual terms, for example business plans that exclude training on your inputs.
- Data classes: which categories of code and data may be shared with which tools. Public open-source code, internal code, customer data and secrets usually need different rules.
- Review duty: every agent-written change goes through the same pull request review and CI as human code, with a named human accountable for the merge.
- Agent permissions: agents run in sandboxes or containers, without production credentials, with network and file access restricted to what the task needs.
- Logging and accountability: which agent actions are logged, who owns each AI feature in production and how incidents are reported.
Everything else, such as preferred prompting techniques, belongs in team guidelines rather than in the policy.
Enforce with tooling, not memos
Policies that rely on individual discipline erode quickly. Make the approved path the easiest path. Provide company accounts for approved tools so that nobody needs a private subscription. Configure tools centrally where possible, including allowed models and permission modes. Add secret scanning and dependency checks to CI so that leaked keys and unknown packages are caught regardless of who or what wrote the code. Protect main branches so that no change bypasses review. For AI features, route all model calls through a central gateway or platform that handles keys, logs requests and enforces data rules. An enterprise platform such as KalaidoCreate provides this kind of central governance, compliance features and custom APIs, so individual teams do not each build their own controls.
Governance for AI features in products
For runtime AI, the policy should require a short assessment before development starts: the purpose of the feature, the data it processes, the model provider and hosting region, the risk classification under the EU AI Act and the planned human oversight. It should also require an evaluation set and a defined quality threshold before release, plus monitoring afterwards. Our primer on the EU AI Act for software teams explains the classification step. Keep the assessment lightweight for low-risk features so that teams do not route around it.
Roles and ownership
Governance fails when nobody owns it. Assign three roles. A policy owner, often in the CTO's office or the security team, maintains the policy and the approved tool list. Feature owners are accountable for each AI feature in production, including its quality and incidents. Champions in each development team help colleagues apply the rules and feed problems back. Review the approved tool list quarterly; the market changes too quickly for annual cycles.
Training is part of governance
Rules without understanding lead to workarounds. Developers need to know why secrets must stay out of prompts, how prompt injection affects agents that read external content and how to review generated code critically. Short, practical training is more effective than long compliance modules. neooutput runs agentic coding workshops that cover secure use of Claude Code, Codex and Cursor alongside productivity, and our article on secure agentic coding practices lists the guardrails every team should learn. For organisation-wide rollout, see rolling out agentic coding in enterprise teams safely.
Getting started
Start by inventorying which AI tools are already in use, including private accounts; the answer is usually more than expected. Choose a small set of approved tools, provide company accounts and write the five policy elements above in plain language. Add the CI controls, run a training session and schedule the first quarterly review. A first version can be in place within weeks. It will not be perfect, but a short, enforced policy beats a comprehensive one that nobody reads. If you need support beyond policy, OnlineAppCreator helps organisations implement governed AI development as part of its custom development services.
Frequently asked questions
Should developers be allowed to use personal AI subscriptions?
Generally no. Provide company accounts for approved tools with contractual data protections, which makes the compliant path also the easiest one.
Does agent-written code need special review?
It needs the same review and CI as human code, with extra attention to security-relevant files, new dependencies and changes to tests.
How often should an AI tool policy be updated?
Review the approved tool list at least quarterly, because tools, plans and model providers change quickly.