Mikkel Freltoft Krogsholm
Menu

A mind, not a brand.

AI · mennesker · fremtidKøbenhavn · Danmark
Essay
10 min read

An Agent-Ready Denmark

Denmark digitised the meeting between citizen and state. The next step is open, secure access that a citizen’s own AI agent can use under a limited mandate.

Mikkel Krogsholm holder en lille messingnøgle, som åbner præcis én luge i en monumental offentlig arkivvæg.

When Allerød Municipality invited every citizen to a meeting about emergency preparedness through e-Boks, the municipality expected 150 people to attend. After a week, 1,200 had signed up, Politiken reported.

It sounds like a democratic success. It also sounds a little like spam.

The municipality sent a message to everyone, even though it expected the vast majority not to be interested. In the end, far more people than expected did want to take part. So the story is not a simple example of poor public communication. The invitation clearly struck a chord.

But it did so through a channel we do not treat like an ordinary inbox.

I do not open e-Boks because the sender has written a good subject line. I open it because the message may contain a decision, a deadline or a bill I cannot afford to miss. e-Boks has been able to function without a spam filter because of an unspoken agreement: when the public sector writes here, it matters.

If public authorities begin to use that access in the same way they use Facebook, websites and newsletters, they do more than change the contents of the mailbox. They spend the attention the channel possesses only because we do not dare ignore it.

This creates an obvious task for AI. A personal agent can read my post, explain official language, find the deadline and tell me what I need to do. It can also separate the invitation from the formal decision. That would be useful, but the need itself would show that something had been lost: before, I may have needed help understanding the important letter. Now I may need help finding the important letter among the unimportant ones.

A Country Built for Humans at a Screen

Denmark has already completed a digital transformation that many countries are still working towards. MitID, Digital Post, borger.dk and the many public self-service systems have made it possible to manage much of one’s relationship with the state from home. The European Commission describes Denmark as a country with robust digital infrastructure and high-quality public services. In a 2025 survey, 81 per cent of Danes said that the digitalisation of public and private services made their lives easier.

That is worth remembering before we criticise the next step. The alternative to digital self-service was not a warm civil servant who knew us personally and had plenty of time. It was also queues, paper forms, opening hours and letters travelling across the country.

But almost all of digital Denmark is built on one assumption: a human is sitting in front of the screen.

The human logs in. The human finds the right page. The human reads the letter, understands the fields and moves information from one system to another. Even when everything is digital, the citizen remains the manual connection between public authorities.

In When the Interface Became Optional, I wrote about how AI can change that relationship. When an agent can understand my intention and work in the systems, I no longer need to know every button, menu and workflow. I move from operating the system to directing the one that operates it.

If that movement is real, the next step in Denmark’s digitalisation will not merely be a better public website. It will be a public Denmark that a citizen’s agent can work with.

The Small Layer on Top of Jobnet

I have built a small example of what that might look like.

At Jobagenten, a jobseeker can give their own AI access to current job listings from Jobnet, Jobindex, Jobdanmark and Akademikernes Jobbank, among others. The user can remain in ChatGPT, Claude or another tool and describe the work they are looking for in ordinary language. The agent can search, read and compare the listings and lead the user back to the original sources.

Jobagenten cannot log in, change a profile or submit an application. It has a limited set of read-only tools that the site provides through WebMCP and MCP.

I have previously built MCP servers for Danish public data from Statistics Denmark and the Central Business Register. Jobagenten makes the same basic idea more human. The user does not need to know which search field Jobnet offers, or how four portals describe geography and job types differently. She describes her need. Her agent handles the translation.

But there is something backwards about the construction. It is my layer on top of Jobnet. Jobnet has not chosen to provide an official agent interface. My adapter uses the portal’s public interface and must follow it if it changes.

That is acceptable in an open job search where the agent can only read. It is not a model for a citizen’s tax case, health record or Digital Post.

Here, the public sector itself should provide a stable, documented and secure point of access. If it is built specifically for OpenAI or makes Claude the official Danish citizen assistant, we have merely replaced the portal lock-in with a new one. The interface should be based on open standards that different agents can use on equal terms.

The technology matters less than the principle. MCP gives an agent access to specific data and tools. A2A can be used when the citizen’s agent needs to conduct a longer process with an agent at the public authority. What matters is that the state builds for the protocol and lets the citizen choose their representative.

The State’s Assistant or Mine

The joint public-sector digitalisation strategy for 2026–2029 has already recognised part of the opportunity. It proposes investigating an AI-based digital assistant so citizens can receive answers and communicate with the entire public sector without navigating complex systems across different authorities themselves. The strategy also mentions shared digital architecture, principles and standards.

That is the right challenge. But the wording conceals a political choice.

Should the public sector build an assistant that citizens must use? Or should it build the infrastructure that citizens’ own assistants can use?

A shared public assistant could be better than yet another collection of self-service systems. It could know the authoritative sources, show the appeals process and be built to Danish requirements. But it could also become a new front door where the state owns the system, the conversation and the interpretation of the citizen’s needs.

The personal agent reverses the relationship. It works for the citizen. It can know her situation across work, finances, calendar and family without all that information having to be collected in one public profile. It can help her deal with several public authorities, just as a skilled adviser or family member does today.

In When Stories Grew Hands, I wrote about agents receiving identity, communication channels and a mandate to act on our behalf. Here, that idea becomes concrete. If the agent is to represent me in dealings with the public sector, the state does not need to know its soul or favour its model. The state must be able to verify whom it represents, what it is allowed to do and what it actually did.

It looks less like a new person and more like a new kind of power of attorney.

MitID Is Not the Mandate

Several building blocks already exist. NemLog-in allows public systems to offer login and signatures with MitID, web-service access, Digital Power of Attorney and access management. In the guidance for Digital Power of Attorney, an authority must define what the authorised person may see or do. Access can be limited by case number, amount and service, among other things. Denmark’s Shared Medication Record even distinguishes between authority to view and authority to act.

This is not a finished solution for AI agents. A human representative and a software program are not the same, legally or practically. But the principles are already recognisable: identity, limited access, a distinction between viewing and acting, and the ability to revoke permission.

MCP’s current authorisation model is based on OAuth. This makes it technically possible to send the citizen through a secure login and then give the agent a short-lived access token that applies only to the selected authority and selected actions. The division can be stated simply:

MitID shows who I am. OAuth describes what my agent may do. MCP gives it the specific tools.

The agent must never receive my MitID codes or be able to approve MitID on my behalf. I log in myself and grant a precise permission. The agent may be allowed to read new letters, identify deadlines and prepare a reply. It is not automatically allowed to send the reply, move money or sign a declaration.

When an action is binding or sensitive, the system can stop and require me to return. Not necessarily at every small step, but at the point where the mandate ends. This is the technical version of the point I previously expressed by saying that an AI agent needs a job description: a text can describe the mandate. The system must enforce it.

The Right to Choose Your Intermediary

There is an obvious objection. Why place yet another private technology company between the citizen and the state? Are we not already dependent enough on American platforms?

Yes. If agent readiness means every public system is custom-built for one provider, we have made the problem worse. A citizen’s access to the state must not depend on whether she pays for the right chatbot, and an authority must not send sensitive information into a private system without a clear legal basis.

That is precisely why open standards matter. They make it possible to switch agents without rebuilding the entire public connection. A local Danish solution, a specialised disability assistant and a large international model could, in principle, use the same public tool under different mandates. The state can require security, logging and identification without deciding which intelligence the citizen should trust.

Nor does this mean that the human route should be closed. Some people will not use an agent. Others will use a bad one. An agent can misunderstand a decision with the same confidence with which it can misunderstand anything else. The authoritative text, the understandable explanation, the telephone number and the route of appeal must therefore remain.

The agent should be a new way in, not an excuse to make the old ways impassable.

Public authorities also have a responsibility not merely to make more information machine-readable, but to improve the interaction itself. When citizens get AI, municipalities may get more work, because citizens can suddenly produce more and longer enquiries. An agent-ready public sector should not only make it easier to send text. It should make it easier to ask the right question, attach the right information and move the case through a structured process.

Otherwise, we merely automate the queue from both ends.

Back to the Mailbox

A personal agent in e-Boks could be a great help. It could translate a letter from official language into everyday language. It could put the deadline in my calendar, find the relevant guidance and prepare the information I am missing. For many people, that could be the difference between formal digital access and a real ability to act. It could also separate the invitation from the formal decision.

But the ability to filter must not become permission for the public sector to send everything to everyone. If e-Boks ends up as just another noisy inbox, we have not modernised the relationship between citizen and state. We have destroyed a valuable shared expectation and then asked AI to repair it for each of us individually.

An agent-ready Denmark therefore requires two movements at once.

The public sector must preserve discipline in the channels that citizens cannot opt out of. At the same time, it must open its systems to the agents that citizens choose for themselves.

Denmark has already given its citizens a digital identity and built a public sector they can meet through a screen. The next step is not to remove the human from the meeting. It is to give the human the option of sending help ahead.

An agent with a key that opens only the doors I have chosen.

Sources and Further Reading