# AI governance framework: a practical template for European companies

> A practical AI governance framework template for European companies: a one-page frame, a register with one row per AI capability, six steps to fill it in, and how it maps to the AI Act.

Language: en · Category: governance · Author: Nicolas Corod (Founder and chief executive officer) · Published: 2026-11-10

Many companies now report an AI governance framework. Far fewer check what their AI actually produces. The swissVR Monitor, cited by Alexandre Buga of Deloitte Switzerland in [PME in October 2026](https://www.pme.ch/invest/ia-en-entreprise-deloitte-decrypte-le-role-du-cfo/4ynpp4t), makes the gap visible. Many companies have set up a governance framework for generative AI, yet fewer than one in five have a person validate all AI-generated content. A framework on paper and a framework in operation are two different documents. This article gives you the second one: a practical template for European companies, filled in one AI capability at a time.

## What is an AI governance framework?

An AI governance framework is the set of decisions that says which AI capabilities a company runs, who owns each one, what each one may touch, and what evidence shows that the rules were followed. It turns principles into owners, limits and records that someone can check.

Most published frameworks stop at principles such as transparency, fairness and accountability. The principles are correct and easy to agree on. The hard part is the sentence that should follow each one: who does what, for which use, and how anyone would know.

The broader five-pillar model (registry, owners, permissions, traceability, lifecycle) is covered in [the enterprise AI governance framework for leaders](/en/blog/enterprise-ai-governance/). This article goes one step further and gives you the template itself.

## Why do most AI governance frameworks stay on paper?

Most frameworks stay on paper because they govern documents instead of uses. A framework that lists principles but names no owner per use cannot be applied on a Tuesday afternoon, when a team asks to connect a new tool to its assistant.

Three symptoms come back in most companies:

- **No owner per use.** A committee "oversees AI", so nobody decides on a specific request. Without defined responsibilities for who deploys, reviews and approves, governance defaults to "no", because "no" is always the safe answer.
- **No link to how people work.** Employees keep using personal accounts, which is how [shadow AI](/en/blog/what-is-shadow-ai/) grows. In French-speaking Switzerland, a Qualinsight survey of 1,030 working people was [reported by PME in July 2026](https://www.pme.ch/strategie/ia-au-travail-72-des-romands-lutilisent-souvent-en-secret/rl801nw). Of the respondents, 49% say their company has not defined clear rules on AI. Among people who use AI at work, 27% do so without approval from their company or its IT department.
- **No evidence.** When an auditor or a client asks what an assistant did last quarter, the framework has nothing to show.

The innovation lead of a 40-person consulting firm described the trap in a meeting with skilder in August 2026. Requests to connect borderline tools are hard to approve when no written method sits behind them, he explained. The longer a company waits to bring scattered skills together, the harder the clean-up becomes.

## What should an AI governance framework template include?

A usable AI governance framework template has two parts: a one-page frame that sets the rules once, and a register with one row per AI capability. The frame changes rarely. The register changes every time a team starts, changes or stops using AI for a piece of work.

### Part 1: the one-page frame

| Block | The decision to write down | Who signs it |
| --- | --- | --- |
| Scope | Which assistants and agents are covered (Claude, ChatGPT, Copilot, any agent compatible with the Model Context Protocol, or MCP) and which uses are out of scope | Executive sponsor |
| Risk tiers | Three tiers, for example: internal drafting, work on client or personal data, actions in business systems | Risk or compliance lead |
| Owners | Every AI capability has one named business owner, not a committee | Executive sponsor |
| Limits | Per tier: which data may be used, which tools may be called, which actions need a human approval | IT and security lead |
| Evidence | What is logged for each run, how long it is kept, who may read it | Data protection lead |
| Review | How often each tier is reviewed, and what triggers an earlier review | Executive sponsor |

The rules for staff (what employees may paste into an assistant, which accounts they may use) belong in a separate document written for them.

### Part 2: the capability register

The register holds one row per capability: one piece of work an assistant does for a team. The sample row below uses invented data for a fictitious logistics company.

| Field | What to fill in | Sample row (illustrative data) |
| --- | --- | --- |
| Capability | The workflow, in one sentence | Compare three supplier quotes and draft a recommendation |
| Business owner | The person who knows the job and answers for the result | Head of purchasing |
| Assistants | Where it runs | Claude and Copilot Studio |
| Tools and data | Systems it reads or writes, and the data classes involved | Reads the ERP supplier table; no personal data |
| Risk tier | The tier from the frame | Tier 2: client and supplier data |
| Human approval | Which outputs a person must validate before use | The recommendation, before any order is placed |
| Evidence | What the log keeps | Who ran it, inputs, outputs, version |
| Version and review | Current version, last and next review date | Version 3, reviewed on 15 October 2026, next review in January 2027 |

The register answers the three questions an auditor, a client or a board member will ask: what runs, who answers for it, and what proof exists.

## How do you fill in the template in six steps?

You fill it in from the uses that already exist, not from a blank page. Start with what teams already do with AI, give each use an owner and a tier, then set limits and evidence. A first version can be in place within a few weeks.

1. **List the real uses.** Ask each team which tasks they already hand to an assistant. Treat hidden uses as a starting point, not as a fault.
2. **Group them into capabilities.** Merge duplicates. Five people with five prompts for the same task become one capability with one method.
3. **Name an owner per capability.** Choose the person who knows the work, not the person who knows the tool.
4. **Assign a risk tier.** Use the three tiers of the frame. Most uses land in tier 1, so spend your review time on tiers 2 and 3.
5. **Set limits and approvals.** For each capability, write which data and tools it may reach and which outputs a person validates. Limits apply to the capability, never to watching people.
6. **Decide the evidence and the review date.** Logging comes before the first incident, not after it. An [AI audit trail that compliance teams can use](/en/blog/ai-audit-trail-execution-logs/) records who ran what, through which capability, with which result.

Step 3 deserves the most care. The owner approves changes, decides on colleagues' requests and retires the capability when it no longer serves. That turns governance into a working method rather than a permanent veto.

## How does the template map to European and Swiss rules?

The template does not make a company compliant on its own. It gives each rule a place to land: the register shows what runs, the frame shows who decided, and the evidence block shows what happened.

| Framework or law | What it expects | Where it lands in the template |
| --- | --- | --- |
| [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj), Regulation (EU) 2024/1689 | Duties by risk level; prohibited practices banned since 2 February 2025; high-risk rules from 2 December 2027 after the 2026 omnibus regulation | Risk tiers, human approval, evidence |
| General Data Protection Regulation (GDPR) | Lawful processing of personal data, and the ability to demonstrate it | Tools and data, evidence |
| Swiss Federal Act on Data Protection (FADP), in force since 1 September 2023 | Applies to AI processing of personal data, including transparency and human review of automated decisions | Data classes, human approval |
| [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) of the US National Institute of Standards and Technology, January 2023 | Four voluntary functions: govern, map, measure, manage | Frame (govern), register (map), evidence and review (measure, manage) |
| ISO/IEC 42001:2023 | A certifiable management system for AI | The frame and the register become its working records |

A Swiss company falls within the scope of the AI Act when its AI systems, or their outputs, are used in the European Union.

## How do you keep the framework alive after day one?

You keep it alive by treating each capability as something that is never finished. Each one follows the AI Capabilities Lifecycle: you author it, test it, deploy it, control it, monitor it and improve it, version after version. The register is where each version is recorded.

Two habits make the difference. First, each new version ends with a short report to the owner's manager: what changed, how the capability was used, which requests are waiting. Second, the limits of a capability, its [permission boundary](/en/blog/least-privilege-ai-agents/), change only through a reviewed version, never through a quiet edit.

Ali Mahfouz, Area Service Lead for French-speaking Switzerland and member of the executive committee of Microsoft Switzerland, put it plainly at Vision Entreprise 2026 in October: "Guardrails are not there to slow down adoption. On the contrary, they are there to ensure deployment at scale."

## Where does skilder fit in this framework?

skilder gives the register a place to run. Teams keep their own assistant (Claude, ChatGPT or Copilot), and skilder serves each capability inside it, with its owner, its limits and its trace.

- **Capabilities grouped into roles.** A role bundles a team's skills, its scoped MCP tools and its access rights. Roles are configured from blueprints, starting configurations adapted to your own processes.
- **Versions.** Each skill is versioned, so the register matches what actually runs.
- **Limits.** Role-based access control and enterprise single sign-on decide who reaches what. Guardrails, rules that check a tool call before or after it runs, are in beta.
- **Evidence.** skilder keeps an execution trace attributed to each user, and the audit log is exportable.
- **Hosting.** Switzerland, with Infomaniak by default, and the European Union; on-premise deployment on request.

Automated re-testing of a capability after each change is coming soon; until then, testing is manual. To fill in the first rows of your register with someone who has done it before, [book a free 30-minute exchange with an expert](https://koalendar.com/e/meet-with-nicolas-9?utm_source=skilder.ai&utm_medium=blog_ai_governance_framework).

**Takeaway:** write the frame once, then govern one capability at a time. A framework that names an owner, a limit and a proof for each use is one that people can follow on a Tuesday afternoon.

To map your current AI uses onto this template, [book a free 30-minute exchange with an expert](https://koalendar.com/e/meet-with-nicolas-9?utm_source=skilder.ai&utm_medium=blog_ai_governance_framework).
