Home

October 27, 2026

The €144 Bureaucracy Adventure

I recently recovered €144 from a subscription I thought I had cancelled.

The interesting part is not the amount. It is that an AI agent helped me get it back.

I have spent years building software, and lately I have spent a lot of time experimenting with agents that write code, inspect websites, and handle small tasks. Yet one of the most convincing uses of an agent so far was not technical at all.

It was an email conversation about a refund.

There is something beautifully disappointing about the future arriving in the form of customer support.

An expensive lesson in subscriptions

I had tried an online service and believed I had cancelled the subscription. Then I noticed a €144 charge.

I cancelled it about ten minutes after discovering the payment. But stopping the next renewal and recovering money already charged are two different problems. The first was a button. The second looked like paperwork.

I could have handled it myself, of course. Find the right support address, explain what happened, read the answer, look up the policy, decide whether to insist, write another message.

None of those steps is particularly difficult. Together, they form a small administrative obstacle course. And, like most obstacle courses, it is much more entertaining when somebody else has to run it.

Normally I would either spend time on it or decide that the time was worth more than the money. That is a strange choice to make: accept a charge you disagree with, or spend part of your life arguing about it.

This time I tried something different.

Hiring an agent for a very boring job

I asked Grok Bot, an AI agent I use for various experiments, to contact the service and request a refund.

That is an important distinction from asking a chatbot to write a complaint. A chatbot can give you a nicely formatted email, but you still have to copy it, send it, wait, read the answer, and start again.

Here I delegated the process, not just the prose.

The agent could work through the correspondence, keep track of the context, and suggest the next action. I still reviewed what it was doing and approved the actions. I did not want an autonomous diplomat inventing international incidents on behalf of my credit card.

The arrangement felt familiar. It is close to how I increasingly work with coding agents: delegate a bounded task, inspect the important decisions, and intervene when necessary.

Except that instead of changing a function, the agent was trying to change somebody's mind.

The first answer was no

The initial refund request was refused.

That is the point where an ordinary support exchange often ends. Not necessarily because the refusal is correct, but because the person asking has run out of patience. There are only so many polite emails one can write about money that has already left the account.

The agent did not get discouraged. It pushed back, referring to European consumer rights and making another case for the refund.

Eventually, the service agreed and refunded the €144.

That is the entire result. One request, one refusal, a follow-up, and the money returned. No spectacular technical trick. No complicated integration. Just persistence, context, and a tool willing to do a tedious job.

I cannot tell which part persuaded the company. Perhaps the reference to consumer rights mattered. Perhaps the second request reached someone who made a different decision. I did not run a controlled experiment with two identical subscriptions and a robot assigned to only one of them.

What I know is that the agent carried the exchange forward after the first refusal, and the refund arrived.

Bureaucracy is an interface

I keep thinking about why this felt so useful.

Most software interfaces are designed around a task. You press a button and something happens. Bureaucracy is often an interface made of documents, policies, forms, and people who can only perform a particular step.

The instructions are scattered. The state is difficult to see. An error may arrive two days later in an email beginning with a courteous apology.

As a software developer, I find the architecture fascinating. As a person trying to recover €144, slightly less so.

An agent is unusually well suited to this environment. It can read long documents, extract relevant details, maintain the context of a conversation, and produce a response without needing to be persuaded that writing another email is worth the effort.

It does not remove the process. It gives the individual a way to navigate it with less friction.

Companies already have systems for subscriptions, payments, renewals, and support. The customer, meanwhile, often arrives with a browser tab and whatever patience is left at the end of the day.

An agent can make that relationship a little less uneven.

Delegation still needs a human

There is an obvious catch: an AI can sound confident about a policy or a legal right without being correct.

In my case, the agent cited European consumer rights. That is what happened in the correspondence, not proof that every subscription refund is legally guaranteed. The circumstances and the rules matter.

I reviewed the steps and approved what was sent. That part is not ceremonial. If an agent is writing to another person or organization on my behalf, the responsibility is still mine.

I am happy to delegate the reading, drafting, and persistence. I am less interested in delegating judgment.

This is also why I find the comparison with coding agents useful. I do not want to inspect every keystroke, but I do want to understand the decisions that matter. A fast feedback loop is valuable whether the output is a pull request or a refund request.

A small glimpse of what comes next

When people talk about AI agents, the examples tend to be enormous: autonomous companies, entire applications built from a prompt, research accelerated beyond recognition.

I like those ambitions. But I suspect that a large part of the value will appear first in much smaller places.

Changing an appointment. Challenging an unexpected bill. Sorting out an insurance form. Following up on a request that has been sitting unanswered for weeks.

Things that are possible today, but consume attention out of all proportion to their importance.

I can imagine a future in which every person has a patient assistant that understands the process, knows when to ask permission, and handles the repetitive parts. Perhaps the organizations on the other side will have agents too, and two machines will exchange extremely polite messages while we go for a run.

That sounds slightly absurd.

It also sounds wonderful.

I got €144 back. More importantly, I got a glimpse of a kind of software that gives people some of their time and patience back.

Perhaps the most useful agents will not do extraordinary things.

They will make ordinary things less exhausting.


Thank you for taking the time to read this post.