Data Governance in the AI Era

AI is no longer just a chatbot where employees type questions. The new generation of AI tools works with connectors, skills, agents, files, workflows, APIs, and shared context. As a result, the question is shifting from 'Which AI tool should we use?' to a much more important one.

“What data does AI get access to, on whose behalf, and under what conditions?”

For many organizations, this is not sufficiently clear today. AI often gets access to the same data as the user, without the organization always understanding what is being read, shared, stored, or reused. Data governance is therefore no longer an IT detail, but a fundamental prerequisite for working with AI safely, scalably, and reliably.

1. From document access to AI access

Previously, data governance was relatively straightforward: an employee had access to a folder, a CRM record, a report, or an ERP module. AI changes that model. An AI assistant can not only read information but also combine, summarize, interpret, restructure, and forward it to other contexts.

Traditional access control is no longer sufficient. The organization needs to know:

• who has access to what data;
• which AI tools can use that access;
• what actions those tools are allowed to perform;
• what data ends up in prompts, files, skills, or connectors;
• whether logging, monitoring, and revocability are in place.

2. Real-life case: shared AI skills and hidden access rights

A tangible risk arises when employees share AI skills, agents, or automations with colleagues. Such a skill might seem harmless—a convenient way to generate reports or summarize client files. But behind such a skill can lie instructions, scripts, file permissions, connectors, and system access.

A possible scenario

A consultant builds an AI skill to perform client analyses more quickly, linked to project documents, templates, and data sources. He shares the skill with a colleague who uses it for another client case. Without malicious intent, the skill could call up, combine, or process information from the original context into a new document.

The problem is that the organization has insufficient visibility into:

• which skills exist and who built them;
• who is allowed to use them;
• which sources are linked to them;
• what permissions they inherit from the user;
• whether sensitive data can travel to other contexts.

3. AI agents as digital insiders

AI agents are not ordinary software buttons. They can plan, make decisions, propose or execute actions, retrieve information from various systems, and repeat tasks without manual oversight of every step. Treat them as digital insiders.

Questions every organization should be asking today

• What AI agents exist in our organization?
• What systems can they access?
• Can they only read, or also write, modify, delete, or send?
• Do they operate with personal permissions or with shared service accounts?
• Are their actions logged? Is there an owner for each agent?
• Can an agent be disabled immediately in case of an incident?
• Is there an approval flow for new agents, skills, or connectors?

4. Why this is relevant for SMEs

Many SMEs think AI governance is a problem for large corporations. In reality, SMEs are often more vulnerable because their processes tend to grow faster than their structure.

• Shared folders without a clear owner.
• CRM data that is not systematically cleaned.
• Excel exports containing client, project, or financial data.
• Power BI reports without data classification.
• AI tools used with personal accounts.
• No central registry of AI tools, connectors, or automations.
• Access is rarely reviewed when roles change.

“AI accelerates what already exists. If governance is strong, AI accelerates the organization. If it's weak, AI accelerates the risks.”

5. The five governance questions

1. What data do we have?

Know the location of your critical data: ERP, CRM, document management, mailboxes, spreadsheets, reporting, planning, and external platforms.

2. Who has access?

Access based on roles, responsibilities, and necessity—not historical exceptions.

3. Which AI tools use that access?

Not just official tools, but also personal accounts, browser plugins, automations, agents, and SaaS tools with AI functionality.

4. What is AI allowed to do?

Distinguish between reading, summarizing, exporting, writing, modifying, sending, and deleting.

5. Who is responsible?

Every dataset, every report, every AI agent, and every integration needs an owner—otherwise, governance becomes a paper exercise.

6. A practical maturity model

Level 1 — Invisible

AI is used, but no one knows by whom, for what purpose, or with what data.

Level 2 — Permitted but uncontrolled

General agreements, little technical control.

Level 3 — Registered

AI tools, data sources, and use cases are registered. Basic policies on sensitive data, prompts, and access.

Level 4 — Controlled

AI access linked to roles, logging, approvals, and periodic reviews. New agents, skills, and connectors are assessed.

Level 5 — AI-ready governance

Data governance is integrated into operations. AI can scale safely.

7. Where do you start?

Not a large governance program, but a concrete AI Data Access Scan that answers three questions:

• What data can be accessed by AI tools today?
• Where are the biggest risks?
• What quick wins can make AI usage safer without blocking innovation?

Typical quick wins

• Create a registry of used AI tools.
• Classify sensitive data.
• Review shared folders and reports.
• Limit exports from ERP and CRM.
• Establish rules for using AI with customer data.
• Set up an approval process for new connectors, skills, and agents.
• Provide for logging and ownership.
• Review access rights per role.

“If you want to deploy AI, you must first understand your data access governance.”