Insight
OPS-0003 · 2026-11-19Interview with Jens Hagel, Founder & Owner, kjello
As founder and owner of a Hamburg-based IT services firm, how would you describe the work your company does and the operational AI systems you run today?
hagel IT provides IT services to small and medium-sized businesses in northern Germany. We have 32 employees and over 150 customers.
AI agents handle many tasks on our websites, blog, social media and servers. I create these agents with AI coding agents. Our own software products are also created by AI coding agents. At night, on weekends, or when all lines are busy, an AI assistant answers our phones and books appointments in our technicians’ calendars. In case of an emergency, it calls our on-call technician. Between 8 a.m. and 6 p.m. on weekdays, you will always reach a human.
What path led you from computer networking and IT services into building and operating AI-driven automations for your own firm and clients?
I started the company by myself from my home in 2004. Since 2025, Philip Kraatz has been running the company as CEO. I am the founder and owner.
One customer had orders coming in as PDFs. For every order they received (30-40 per day), they had to enter it into their enterprise resource planning (ERP) system, which handles their orders and invoices. Each order took 15–20 minutes to process, and if someone made a typo in the part number, the wrong item was delivered. We built an AI system that can read the email and PDF and create a table and invoice. Now each order can be checked in under a minute.
When you first put agents into real workflows, what made you decide that prompt instructions alone were not a sufficient safeguard?
They will follow the rules most of the time. Most of the time isn’t enough when your agent is working with customers’ websites or sending out emails.
I moved those restrictions into code that checks what the agent wants to do before it does it. You want to give your agent restrictions that it can’t break — with code. A prompt is something it can forget.
How have you translated your non-negotiable boundaries — such as prohibiting deletions, account changes or unapproved external contact — into technical checks that other small firms could adopt?
Each task your agent wants to do has to be checked against something. If it’s something on your “Don’t Do” list, such as other people’s domains, deleting data, or changing accounts or subscriptions, your agent won’t be able to do it. Emails are only allowed to be sent through one channel, with a limit on the number of emails per day, and the agent always copies me.
You can implement this with little work if your company is small. Make a list of things that will hurt you. Make sure your agent can’t do them through code. Make sure all emails go through one channel that you can check. I tell my agents what I want them to do, give them context and restrictions, but not how to do it.
What does an effective kill switch look like in a small-team environment, and what did your experience with queued jobs teach you about what it must stop?
It should stop your agent immediately, without having to deploy anything, and your team should understand how it works. Our website internal linking tool will stop when one file is deleted.
It will also stop itself. If three articles drop by more than three spots on Google within 14 days, it will stop itself and email me. It is also limited to making only five changes a day across all our websites.
The queue is what I learned from. It will stop new jobs from being added to the queue, but it will not stop jobs that are already in the queue. So you would have to stop those as well.
For customer-facing tasks such as order or invoice processing, how do you decide where an agent can act independently and where a person must remain the final approver?
For our manufacturer customers, the AI will prepare everything, and a human will check it before it reaches the customer.
This check takes under a minute. They are able to send out their invoices immediately instead of waiting 24-48 hours. One of our customers’ CEOs initially doubted that it would work because their orders varied so much that they weren’t sure how AI could understand them.
What practical limits have helped you keep AI-agent spending predictable, particularly when workflows can run repeatedly or call several services?
Limit how long your agent can run. One job in September 2026 would have cost us around $7,500 if we hadn’t had this limit. Ninety-three percent of that cost would have been from sub-agents.
Now, our agents stop after around 200 steps and pass the work to a new agent. We do this with code as well. Our agents are told to wrap up at 180 steps and can’t use any tools after step 221.
After a model provider change left your social-media pipeline failing silently, what monitoring checks did you add to confirm that a workflow completed rather than merely started?
Our social media automation was using a free large language model (LLM) that our provider cut off in July 2026. The automation failed on the first AI step, but nothing looked wrong. It took us around 10 days to notice that we weren’t posting.
I now ask my checks whether the task ran today, instead of asking whether there has been any activity in the last x amount of time. I also added checks that stop our builds if they don’t create at least a certain number of pages.
What have you learned about designing alerts that a busy owner or operations team will actually notice, understand and act on?
Employ fewer of them. One of our website checks emailed us on 26 of 41 days. Only four were actual outages. One deployment would send us 19 emails every 10 minutes. Our server checks were sending us around 200 emails a week.
We now start an AI to fix our issues if they turn red. We send an email only if they are still red afterward, once per incident, and provide a weekly summary. If half your inbox is from your automation, you won't read the emails.
Thanks for sharing your knowledge and expertise. Is there anything else you'd like to add?
Tell your AI what day it is. One of our companies' automated LinkedIn posts was telling people that Windows 10 would end support soon. Support ended in October 2025. A business partner asked us if we were a year late. Our AI wasn't aware of what day it was, and we had a check in place for old information that was too strict.
We had to edit 182 passages in our 41 blog articles. We now tell our AI when something ends: “Since October 14, 2025” instead of “soon.”
This article was first published on Featured News.
