Norway’s Privacy Regulator Warns of ‘Digital Ghosts’: AI Agents That Outlive the People Who Built Them

An employee builds a small automation. An AI agent checks a case management system every morning, pulls new records, drafts summaries and files them. It saves hours. Six months later the employee moves to another department, and then leaves the company. The agent keeps running. It still has access, it still reads personal data, and nobody remembers it exists.

Norway’s data protection authority, Datatilsynet, has a name for this: a “digital ghost.” In new guidance titled “Good routines prevent AI from becoming an invisible threat,” the regulator warns that AI has moved from passive tool to independent actor, and that organizations need to govern AI agents the way they govern people and systems with access to sensitive data. (Quotes below are translated from Norwegian.)

From chatbot to digital co-worker

Datatilsynet starts from a simple observation. Many organizations already use chatbots to summarize text or draft documents. AI agents go further: they can fetch information, update business systems and carry out routine tasks on their own. “In areas where AI has mainly been a tool, it can now also act more like a digital employee,” the authority writes.

When agents get direct access to email, file shares and case handling systems, the regulator says organizations must already have a deliberate approach to data quality, security and access control. It points to the Norwegian Digitalisation Agency’s “order in your own house” concept as the foundation: you cannot govern an agent’s access to data you have not organized yourself.

How digital ghosts appear

The guidance describes several ways AI automation can slip out of view:

• Orphaned workflows. An employee who built an agent-based automation changes role or leaves, and the automation keeps querying systems and processing personal data in the background.
• Shadow AI inside approved software. Because AI is increasingly built into ordinary applications, unapproved AI features can appear inside tools the organization already approved.
• Growing context. Over time, agents can learn from and collect large amounts of context, from customer information to internal meeting notes.
• Agent swarms. When several agents work together, one may collect information, another analyze it and a third make a decision, creating what Datatilsynet calls “a new ecosystem for data processing” inside the organization.

In each case, the regulator says, it becomes unclear who owns the process, what data is processed and how it is managed. That is a privacy problem under the GDPR, which Norway applies through the EEA Agreement, because a controller must be able to account for every processing operation it carries out.

The three questions Datatilsynet wants answered

The authority stresses that “the goal is not to ban the use of artificial intelligence, but to ensure it happens in a safe and well-considered way.” Staff awareness is the first step, but it must be backed by solid technical controls and routines. To stay ahead, Datatilsynet recommends that organizations build a systematic overview by mapping three things:

1. What exists. Which AI tools, assistants and automations are in use.
2. What the agents can do. Which systems each agent may access and which actions it is allowed to take.
3. Who is responsible. Who monitors each agent, owns the process and has authority to switch it off if needed.

The guidance sums it up in one line: “An AI actor needs a defined identity, clear limits on access and a clear line of accountability.” The full text and a recorded talk by Eirik Gulbrandsen and Sara Rossebø are on Datatilsynet’s website.

Turning the guidance into controls

The three questions are easy to agree with and hard to answer in a large organization. Here is how they translate into work a privacy, security or IT team can assign.

Give every agent an identity

An agent should run under its own service account, not under a human employee’s credentials. That way its access can be reviewed, limited and revoked without touching a person’s account, and its actions appear in logs under its own name. When an employee leaves, offboarding should include every agent that person created or sponsored.

Scope access to the task

Data minimization applies to machines too. If an agent summarizes support tickets, it does not need read access to the whole HR file share. Our guide to purpose limitation and data minimization in agentic AI walks through what the GDPR expects.

Name an owner and a kill switch

Every agent should have a named owner who can explain its purpose and a documented way to disable it. If nobody can switch it off, nobody is in control of it.

Watch the vendors

Agents often call outside models and services, which can add processors the organization never approved. We covered this risk in shadow AI and hidden subprocessors.

Review on a schedule

An inventory taken once goes stale. Datatilsynet’s language about continuous attention suggests periodic review, for example quarterly, of every agent’s access, owner and activity.

Part of a wider regulatory pattern

Datatilsynet is not alone. The UK Information Commissioner’s Office has published its own thinking on agentic AI, and governance frameworks are being rebuilt around the idea that AI can act, not just answer. Our analysis of AI’s agentic leap and the data risk management framework for the shadow AI era cover the same shift from different angles.

How Captain Compliance can help

The first question in the guidance, what exists, is an inventory question. Captain Compliance’s data governance and data mapping tools help teams record systems, data flows, vendors and processing activities, and connect those records to assessments and AI governance work. Adding AI agents to that map, with their owners and the data they touch, turns a hidden process into a documented one. A record of processing is not proof that an agent behaves as described, so the map should be paired with access reviews and logs. We do not claim our software discovers every agent automatically.

Frequently asked questions

Is this guidance legally binding?

No. It is guidance from Datatilsynet on how to meet existing obligations, mainly under the GDPR as applied in Norway. It signals what the regulator will expect to see.

What is a “digital ghost”?

An autonomous AI process that keeps running and handling data after the person or project responsible for it has gone, so that nobody clearly owns it.

Does Datatilsynet want companies to stop using AI agents?

No. The authority says the aim is safe and well-considered use, supported by awareness, technical controls and routines.

What should go in an AI agent inventory?

At minimum: the agent’s purpose, owner, the systems and data it can access, the actions it can take, outside services it calls, and how to switch it off.

Does this apply outside Norway?

The guidance is Norwegian, but the underlying GDPR principles of accountability, minimization and security apply across the EU and EEA.

The audit nobody has run yet

Most organizations can produce a list of employees with access to their customer database. Far fewer could produce a list of the AI agents with the same access. Datatilsynet’s guidance is a polite prompt to build the second list before a breach, a complaint or an inspection forces someone to build it in a hurry.