Enterprise AI governance: a practical framework for leaders

Your teams did not wait for an AI policy to start using AI. According to the 2024 Work Trend Index from Microsoft and LinkedIn, a survey of 31,000 people in 31 countries, 75% of knowledge workers were already using generative AI at work, and 78% of those users brought their own tools. It looks a lot like the early days of the cloud: teams adopted the tools long before IT knew which ones. The question is no longer whether AI enters the company, but who steers it, and who answers for it in front of an auditor or a regulator.

Enterprise AI governance is the ability to know which AI capabilities run in the organization, who owns them, who can use them, with which data, and what they produced. It rests on a registry, named owners, permissions, traces and a lifecycle.

This article lays out a practical five-pillar framework, recaps the Swiss and European regulatory context, and points to a ten-question checklist to see where your organization stands.

Why a usage policy is not enough

Most programs start with a usage policy: what is allowed, what is forbidden, which data must never be pasted into a chatbot. That is useful, but it falls short for three reasons.

• It cannot be enforced. A policy does not say who has access to which system, or with which rights.
• It cannot be verified. Without a log, there is no way to show what happened after an incident, a complaint or an audit.
• It captures nothing. The methods that work stay in a few people’s chat histories, and leave with them.

The most visible symptom is shadow AI: employees using personal accounts because no governed alternative exists. The most expensive symptom is harder to see: know-how that never gets passed on, a topic covered in the article on knowledge versus know-how.

The five-pillar framework

1. A registry of AI capabilities

You cannot govern what you cannot see. The first pillar is a central inventory of approved AI capabilities: instructions, skills, connectors to internal systems. The registry answers one simple question: what is allowed to run in the organization, and in which version?

Instruction files scattered across laptops create exactly the opposite problem. That is the argument of SKILL.md, the king of shadow AI: an open, portable format, but an unmanaged one as long as it lives outside any registry.

2. Named owners

Every critical capability has an owner. Not a committee: one person who answers for what the capability does, approves its updates and makes the call when it produces a wrong answer. Without an owner, an AI capability ages in silence.

3. Permissions by role, not by person

Granting access person by person does not scale and does not survive departures. The right level is the role: the work a position on the org chart actually does. A role bundles the instructions, connectors and context that position needs, and nothing more. When an agent connects to a system through an MCP server, the role decides whether it is allowed to.

4. Attributable traceability

Every run leaves a trace: who launched what, when, with which capability, with what result. That trace serves audits, the improvement of capabilities and the measurement of real usage. It must be exportable and, to respect individuals’ rights, deletable.

5. A versioned lifecycle

AI capabilities change: a process evolves, a model is replaced, a business rule is corrected. Every change should go through a review, create a new version and be reversible. Without versioning, nobody can explain why the agent behaves differently than it did yesterday.

The regulatory context

In Switzerland: the FADP already applies

Switzerland has no AI-specific law, but that does not leave a legal vacuum. The Federal Data Protection and Information Commissioner (FDPIC) points out that the Federal Act on Data Protection (FADP), in force since 1 September 2023, applies directly to AI-based processing. In practice:

• transparency about the purpose, functioning and data sources of AI systems;
• informing the data subject of an automated individual decision (Art. 21 FADP), with the option of having the decision reviewed by a human;
• a data protection impact assessment (Art. 22 FADP) when processing is likely to create a high risk.

On regulating AI itself, the Federal Council announced on 12 February 2025 that it intends to ratify the Council of Europe Convention on AI and to pursue a sector-by-sector approach, in areas such as health and transport.

In the EU: the AI Act, in stages

Regulation (EU) 2024/1689, known as the AI Act, applies in stages. Prohibited practices have applied since 2 February 2025, and the obligations for providers of general-purpose AI models since 2 August 2025. The omnibus Regulation (EU) 2026/1744, in force since 27 July 2026, postponed the rules on high-risk systems to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems. It also rewrote the AI literacy obligation (Article 4), giving the Commission and the Member States a more active role.

A Swiss company is not outside this text if its AI systems or their outputs are used in the EU. More importantly, both frameworks converge on the same demands: know what runs, be able to explain it, be able to prove it. Those are exactly the pillars above.

Frameworks to structure the work

To avoid starting from scratch, the NIST AI Risk Management Framework, published on 26 January 2023 and voluntary, organizes risk management around four functions: govern, map, measure, manage. It is free and independent of any vendor.

The checklist: ten questions to see where you stand

Rather than a weeks-long audit, start with a quick inventory. The 10-question AI check-up asks ten yes/no questions in three groups:

• Capture and ownership: where do your approved AI instructions live, and who owns them?
• Control and traceability: do you know who can access what, and who used which capability?
• Durability: would your methods survive the departure of the person who built them?

It takes two minutes. The score is computed in your browser, with no form and no email address.

How skilder supports governance

skilder is a registry and control plane for the organization’s AI capabilities, independent of any vendor. Capabilities are grouped into roles, configured from blueprints on your own processes, then served in Claude, ChatGPT or Copilot. Your teams have no extra tool to open.

• Registry: skills, roles and MCP servers are versioned, with an internal, curated registry of MCP servers.
• Permissions: the role is the access boundary, with SSO and role-based access control; its rights do not widen at run time.
• Traceability: every task is logged and attributed to the person who launched it; skilder keeps this execution trace, which the customer can export and delete.
• Lifecycle: organization-wide promotion and rollback of versions.
• Measurement: usage per role, with an estimate of time saved drawn from real usage.

Key takeaway

AI governance is not a brake on adoption. It is what turns invisible individual usage into shared, verifiable and durable capabilities.

To see where your organization stands, start with the AI check-up. To map your usage with a founder, book a free 30-minute call. skilder’s security posture is detailed on the Trust & security page.

Frequently asked questions

What is enterprise AI governance? +

It is the set of rules, responsibilities and tools that lets an organization know which AI capabilities are in use, by whom, with which data and with what result. It covers the inventory of uses, the assignment of ownership, access rights, the traceability of every run and the upkeep of capabilities over time.

Where should an AI governance program start? +

With an honest inventory: which AI tools your teams actually use, with which accounts, on which data, and which working methods already exist only in people's heads. The skilder AI check-up asks ten yes/no questions to place your organization in two minutes, with no form.

Is an AI usage policy enough? +

No. A policy states intent, but it does not say who can access what, and it leaves no trace of what was done. Without a registry of capabilities, permissions and an execution log, it can be neither enforced nor verified after an incident or during an audit.

Does Swiss data protection law apply to AI? +

Yes. The Federal Data Protection and Information Commissioner points out that the FADP, in force since 1 September 2023 and worded in a technology-neutral way, applies directly to AI-based data processing: transparency about purpose and functioning, the right to have an automated decision reviewed by a human, and a data protection impact assessment when the risk is high.

When does the EU AI Act apply to companies? +

In stages. Prohibited practices have applied since 2 February 2025 and the obligations for providers of general-purpose AI models since 2 August 2025. The omnibus regulation in force since 27 July 2026 postponed the rules on high-risk systems to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). A Swiss company is in scope as soon as its AI systems or their outputs are used in the EU.

Does governing AI slow down adoption? +

Usually the opposite. As long as everyone keeps their own prompts, good practice does not spread and shadow AI thrives. Governance built on shared roles gives teams validated capabilities inside the tool they already use, which makes the governed path simpler than the workaround.