Quick answer: AI governance is the set of rules, roles and controls through which an organisation decides how AI is used, by whom, with what data and at what cost. Effective governance does not start with a twenty-page charter: it starts with measuring real usage, fits on one page and is proven by indicators.
What is AI governance?
Quick answer: AI governance is the way a company organises decision-making about its uses of AI: who can use what, with what data, under what rules, and how it checks that those rules are followed. It is not limited to legal compliance or to IT security; it links the two to the company's strategy.
AI governance: a decision and control framework that sets the authorised uses of AI, everyone's responsibilities, the technical guardrails and the monitoring indicators, and that adjusts based on what real usage shows.
An AI governance framework answers four questions: what is allowed, who answers for it, how it is applied, how compliance is known.
Public frameworks point in the same direction. The AI Principles of the Organisation for Economic Co-operation and Development (OECD), adopted in 2019 and updated in May 2024, place transparency, robustness and accountability among the five values-based principles they recommend.1 The AI Risk Management Framework of the National Institute of Standards and Technology (NIST AI RMF), published in January 2023, makes the “govern” function a cross-cutting function that runs through the other three: map, measure, manage.2
AI governance, data governance, ethics: where is the boundary?
Data governance organises the quality, access and protection of the company's data, with or without AI. AI governance builds on it, but deals with one more question: what a model does with that data, and what the company does with its answers.
AI ethics, or responsible AI, sets values: fairness, respect for people, no manipulation. Governance makes them verifiable: a value without a rule, without an owner and without an indicator remains an intention.
Why does AI governance reach the executive committee?
Quick answer: governing AI is an executive committee matter because decisions about AI commit the whole company, and management currently takes them without data. The uses already exist; the role of governance is to give the committee the figures it needs to decide.
The scenario is a familiar one: a salesperson pastes a proposal into a consumer tool, a finance team buys a subscription on a company card, and six months later nobody knows what data has left or what the whole thing costs. The phenomenon has a name, Shadow AI, and the guide devoted to it explains how to detect and reduce it without slowing adoption.
The usual reflex is to write a charter, then to ban. A ban without an alternative moves usage elsewhere, it does not remove it. A charter that nobody can check reassures the executive committee without changing anything about what actually leaves the company. The problem is therefore not the lack of a rule: it is the lack of data to decide.
The regulatory framework adds pressure, without being the main driver. The EU Artificial Intelligence Act (AI Act), Regulation (EU) 2024/1689, was amended by Regulation (EU) 2026/1744 of 8 July 2026, known as the Digital Omnibus on AI: its Article 4 now requires providers and deployers of AI systems to take measures to support the AI literacy of their staff, without requiring them to guarantee a given level for each person, and the obligations applicable to high-risk systems in Annex III are postponed.3 These obligations, their timeline and the risks of unmanaged adoption are detailed in the guide to adopting AI without losing control.
What principles should an AI governance framework rest on?
Quick answer: an AI governance framework rests on five principles: transparency, human oversight, accountability, security and data protection, monitoring. They overlap with the public frameworks. The difference lies in the evidence: for each principle, the committee must be able to show a fact drawn from usage, not just a paragraph in a policy.
| Principle | The committee's question | The evidence expected |
|---|---|---|
| Transparency | Does the committee know which tools and which models are used, by whom and for what? | The list of authorised uses, kept up to date, and usage statistics by team and by model. |
| Human oversight | Who validates an output before it leaves the company or is used to make a decision? | A written validation rule for each type of use, known to the teams and shown in the tool. |
| Accountability | Who answers for a use, an incident, a choice of provider? | One name per decision, and a log of the committee's decisions. |
| Security and data protection | What data goes to a model, and under what contract? | Masking of sensitive data before sending, the providers' contracts and an exportable audit log. |
| Monitoring | Are the rules applied, and do they have the expected effect? | Five indicators reviewed every quarter, and the rule changed when the figures call for it. |
This table works as a test. If a row can only be filled in with a document, the principle is not yet governed. If it can be filled in with a statistic, a log or a name, it is. That is the point of this guide's angle: a rule is proven by usage, not by how it is worded.
The data protection principle does not require banning all sensitive data, which would strip AI of its business value, but knowing what goes to the model and reducing it to what is necessary. Automatic masking of data before sending is its most direct form; how it works is explained in the guide Data anonymisation and AI, and the link between hosting, applicable law and data in the guide on sovereign AI in business.
Who decides what when governing AI in a company?
Quick answer: executive management sets the ambition and the accepted level of risk; a small AI governance committee turns that ambition into rules and arbitrates sensitive cases; business units decide day to day within the set framework. The committee does not approve every use: it sets the framework and looks at the figures.
The French data protection authority, the Commission nationale de l'informatique et des libertés (CNIL), recommends that deploying generative AI should rely on governance involving, from the outset, the data protection officer (DPO), the chief information security officer (CISO), the head of information systems and the business managers.4 This group of four forms the core of the committee, under the authority of a member of management.
What should an AI governance committee be made up of?
In a small or medium-sized enterprise (SME) or a mid-sized company, five to seven people are enough: a sponsor from executive management, who has the final say; the chief information officer (CIO) or equivalent; the CISO; the DPO; a representative of the finance department; two or three business champions, chosen from the teams that use AI the most. Human resources (HR) join the committee when training or work organisation is on the agenda.
Rhythm matters more than size: one hour per quarter, with a fixed agenda (indicators, incidents, requests for new uses, rule update), is worth more than a monthly meeting without data.
How do you share out decisions without creating a bottleneck?
The classic trap is the committee that approves everything. Requests pile up, delays grow, and teams go back to their personal tools. The solution is to classify decisions by level of risk and to escalate only sensitive cases.
| Decision | Who decides | Who is consulted | Frequency |
|---|---|---|---|
| Ambition, overall budget, accepted level of risk | Executive management | Governance committee | Once a year |
| Usage rule and list of authorised uses | Governance committee | Business champions, DPO | Every quarter |
| Choice of tools and model providers | Committee, on the CIO's proposal | CISO, DPO, purchasing | For each new tool |
| New use within an already authorised scope | Business champion | Nobody, if it follows the rule | As it arises |
| Sensitive use (health data, HR, decision about a person) | Governance committee | DPO, legal counsel | Case by case |
| Incident (leak, serious error, prohibited use) | CISO and DPO | Executive management informed | Within 48 hours (internal deadline) |
Does the committee replace the DPO? No. The DPO keeps the tasks assigned by the General Data Protection Regulation (GDPR), in Articles 37 to 39; the committee gives the DPO a forum to weigh in on the choice of tools.
Which public frameworks can serve as a foundation?
Quick answer: AI governance does not have to start from a blank page. The OECD principles provide the vocabulary, the NIST framework provides the steering logic and its generative AI profile a list of risks, and the CNIL recommendations provide the framework applicable in France. An SME or a mid-sized company can draw on them without aiming for certification.
| Framework | Nature | What it brings | What an SME or a mid-sized company should keep |
|---|---|---|---|
| OECD AI Principles | Intergovernmental principles, adopted in 2019 and updated in 2024 | A shared vocabulary: transparency, robustness, accountability | The basis of your preamble, in five lines |
| NIST AI RMF 1.0 | US framework for voluntary use, 2023 | Four functions: govern, map, measure, manage | The logic of steering by measurement |
| NIST AI 600-1 | Profile of the framework for generative AI, 2024 | Twelve risks unique to or exacerbated by generative AI | A checklist for your use cases |
| CNIL recommendations | Recommendations from the French data protection authority, 2024 | Start from a need, govern uses, train, involve the DPO and the CISO | The framework most directly applicable in France |
The four functions of the NIST AI RMF give a governance framework a sound structure: govern (culture, responsibilities, processes), map (the context and risks of each use), measure (quantitative or qualitative tools to track those risks), manage (allocate resources to the risks identified).2 Its profile for generative AI, NIST AI 600-1, published in July 2024, lists twelve risks unique to or exacerbated by these models, including confabulation (a false answer delivered with confidence), data privacy, information security and intellectual property.5
For organisations that want to go as far as a certifiable management system, the ISO/IEC 42001:2023 standard, published in December 2023, specifies the requirements for establishing, implementing, maintaining and continually improving an AI management system.6 How to set it up, its steps and how it fits with governance are covered in the guide on the AI management system (AIMS).
On the security side, the ISO 27090 standard describes the threats specific to AI and the measures to address them: the CISO will find what is needed to complete the committee's risk analysis.
Where does your organisation stand: informal, ad hoc or formal?
Quick answer: the maturity of AI governance can be read on three levels. Informal: everyone picks their own tool. Ad hoc: a rule or a tool exists, with no follow-up. Formal: a committee, a one-page rule, a shared tool and indicators. The level shows in the facts observed, not in the documents that exist.
| Level | What you observe | The main risk | The next step |
|---|---|---|---|
| Informal | Everyone picks their own tool; no written rule; management discovers uses in passing during a meeting. | Customer or strategic data sent to personal accounts, with no contract. | Measure real usage and provide official access. |
| Ad hoc | An internal memo or a charter exists; a tool was bought for one team; nobody follows up on how it is applied. | A rule nobody can check, and scattered spending. | Set up the committee and plug in indicators. |
| Formal | Regular committee, one-page rule, shared tool, indicators reviewed every quarter. | Heaviness: a framework that no longer keeps up with new models. | Keep the loop short and revise the rule based on the figures. |
To see where you stand, ask your CIO three questions: how many employees used an AI tool last month? With which models? For what total amount? Without a figure, the organisation is at the informal level, even if a charter has been signed. With precise answers and a recently reviewed rule, it is at the formal level.
In practice, the hardest step is the first one: as long as the official tool is less convenient than the personal one, no rule holds.
How do you set up AI governance from real usage?
Quick answer: governing AI comes down to four steps that form a loop: measure usage, decide a one-page rule, enforce it through the tool, then review it every quarter based on the figures. The guiding principle fits in one sentence: start from usage data, not from the budget or a charter template.
Measure before you write
The committee's first decision is not a rule, it is a measurement. Open official access to several models for a pilot group, then observe for four to six weeks: which teams use it, for which tasks, with which models, what sensitive data appears in the requests, at what cost.
This snapshot replaces weeks of workshops: it shows the uses that create value, those that expose data, and actual spending. The register of uses is then built from what exists, not from what everyone declares.
A one-page rule, technical guardrails
With these figures, the committee writes a rule that fits on one page. It can later give rise to a more detailed AI charter, but the page remains the reference.
- Authorised uses, by broad family (writing, summarising, analysis, code), and prohibited uses.
- The data that must never leave in clear text, and the mechanism that protects it.
- The human validation rule for any content that leaves the company.
- The official tools, and the procedure for requesting a new one.
- The names of the sponsor and the champions, and the channel for reporting an incident.
Each line of the rule must correspond to a guardrail enforced by the tool, not by goodwill: rights by team and by model, automatic masking of sensitive data, spending quotas, audit log.
A quarterly review driven by the figures
Every quarter, the committee looks at the indicators, the incidents and the requests for new uses, then changes the rule: a new, better-performing model is opened to the teams concerned, a team that sends a lot of sensitive data has its training stepped up, spending that drifts is explained before it is cut.
Each decision thus rests on a figure, and the rule follows the evolution of the models instead of holding it back.
Which indicators should you track to steer AI governance?
Quick answer: five indicators are enough for the governance committee: adoption, governed uses, protected data, spending and incidents. They are read together every quarter and drawn from the usage and cost statistics of the official tool; rebuilt by hand, they turn governance back into paperwork.
The choice of tools that secure, steer and centralise usage is covered in the guide to adopting AI in business.
The five indicators to present to the committee
| Indicator | What it measures | The question it puts to the committee |
|---|---|---|
| Adoption | Share of employees active on the official tool, by team | Where does the official tool fail to meet the need? |
| Governed uses | Number of use cases authorised and actually used | Which uses should be added, which removed? |
| Protected data | Volume of sensitive data detected and masked before sending | Which teams handle the most sensitive data? |
| Spending | Cost by team and by model, compared with the planned envelope | Does spending follow the value created? |
| Incidents | Reports, serious errors, prohibited uses observed | Is the rule understood and applied? |
These indicators are read together: spending that falls while adoption drops often signals a return to personal tools.
What Crosstensor offers
Crosstensor turns AI adoption into a controlled process. Guard, its differentiator, detects sensitive data, replaces it with tags before sending it to the model and puts the real values back into the answer, with compliance reports. Three outcomes for management:
- Adoption: an interface as simple as a consumer tool, but provided by the company, with single sign-on (SSO), Member and Administrator roles and rights by team and by model.
- Financial control: a single subscription in the form of a pooled envelope of credits, quotas by team and a dashboard of usage and costs by team and by model. AI spending is controlled, no longer endured.
- Freedom to choose the model: models from Anthropic, OpenAI, Mistral and Google in a single interface, with an exportable audit log for the committee's review.
The platform runs as software as a service (SaaS), with nothing to install and no server to maintain. The details are on the Platform offer page, and the price grid on the Pricing page.
Conclusion
AI governance is not a binder of procedures: it is management's ability to decide on its uses of AI with full knowledge of the facts. It rests on five principles, a small committee, public frameworks and a short loop: measure, decide, equip, review. A one-page rule, enforced by the tool and checked with five indicators, protects better than a charter nobody can verify. Start from usage data, not from the budget: that is what lets you move fast without losing control.
