AI Radar is Jasmin Guthmann’s monthly column on Agentic AI, business transformation, and the future of digital value creation. With many years of experience in technology, transformation, and AI, she is one of the established voices shaping the conversation around the future of digital business. In April 2025, her book If There Is a Will, There Is a Way was published — a no-compromise guide to unlocking potential by focusing on what truly matters. Her message: take action, embrace responsibility, and stand for what you believe in.
In the diva-e Conclusion column AI Radar, she shares insights from her practical experience, provides context on current developments, challenges trends, and explores what technological change means for businesses, leadership, and organizations.
Accept all
I have clicked "Accept all" on roughly ten thousand cookie banners. Read: zero. Last Tuesday I accepted the terms of a hotel wifi in Munich in under a second, between two emails, and was back in my inbox before the page had finished loading. Legally, I consented. Factually, I wanted the wifi.
That click is the most honest thing in enterprise AI right now. Because it is exactly what "human in the loop" looks like once the slides are over.
The loop is empty
Ask any steering committee how they govern their agents. Same answer every time: a human approves. Sounds responsible. Sounds like control. Now walk down two floors and look at the human.
She gets 300 agent outputs a day. Supplier matches, ticket classifications, draft replies, price adjustments. The agent is right about 95 percent of the time. Nobody told her which five percent. So she does what you would do. She approves. Faster every week. By week six she is not a reviewer. She is a latency.
This is not a character flaw. It is the most replicated finding in human factors research since the nineties: the better the automation, the less the human checks. Aviation found out the hard way. In 2013 the FAA sent airlines a safety alert telling them to make pilots hand-fly more, because autopilot had quietly eroded their manual skills. Pilots. Years of training. Simulator every six months. Still happened. Your procurement clerk got a Teams message and a twenty-minute tutorial.
Compliance by checkbox
The EU AI Act requires human oversight for high-risk systems. It does not require the human to understand anything. Nowhere does it say the human must have time, context, or the authority to say no without a meeting. So companies do what companies do with a requirement: they satisfy the letter. A name in a field. A button labelled "Approve." A log that proves someone was there. Was she, though?
Here is the uncomfortable version: a lot of human-in-the-loop setups are not there to catch errors. They are there to have someone to blame. The agent makes the decision. The human signs it. When it goes wrong, the org chart knows where to look. That is not governance. That is a liability transfer with a login.
"So you want no humans. Full send."
No. I want fewer humans, and I want them where they actually add something. Three tests:
Does the human add information, or a signature? If every approval is a yes, the step is theatre. Remove it and put the saved time into a weekly sample review where someone reads twenty cases properly, with coffee.
Does the agent earn autonomy, or get it by default? An agent should start with narrow limits and widen them with a track record you can show. A new hire doesn't get procurement authority in week one either.
When the human says no, does anything happen? If a rejection goes nowhere and nobody asks why the agent got it wrong, you have not built oversight. You have built a complaints box with no bottom.
The pattern is older than AI. You don't put a quality inspector at the end of the line and call it a quality system. You design the line so the inspector mostly has nothing to do, and then you check whether that's true.
Who is clicking for you?
Most companies I talk to cannot answer three questions:
Which agent decisions can a human actually reverse?
How many did a human reverse last month?
And what happened to the agent afterwards?
If the answers are "not sure," "none," and "nothing," then the human is in the loop the way I am in the cookie banner. Present. Consenting. Absent.
Somewhere in your company, right now, someone is clicking Accept all on your agents. Go find out what they would say if you asked them why. Want to make AI agents more accountable and human oversight more effective?
