Working a case at Klarna meant clicking through three screens of fields. Some mandatory, some not, and you couldn’t reach the next one until you’d picked something in the one before it. Occasionally a value that existed for no reason other than to get to what came after. Forget one and the case filed wrong. A few thousand people did this several dozen times a day.
I was chosen for a newly formed project team of six to move the company off Salesforce, one person out of each of the six processes in what was then called the Service Domain. Leadership had taken a significantly cheaper offer from Freshworks, and the six of us were supposed to make it happen. None of us had implemented a CRM before.
Who we built it with
Every team lead appointed one person from their team as our counterpart, to test what we built and tell us where it broke. I sat with them and learned processes I had never worked in myself. Eleven or twelve markets at the start. Seventeen by the end.
Every time a workflow refused to map cleanly, we asked those people whether the step had to be that way. Was there a regulatory reason. Was there a business reason. Often there wasn’t. That’s just how Salesforce does it. Some of them volunteered that they had thought the step was stupid for years, and had done it anyway.
The cascading dropdowns went. I have no number for you, but clicks per case came down noticeably, and everyone who tested it could feel it. The correctness of case handling didn’t go down when the clicking did.
What actually made it possible
Not that Freshworks was a better product than Salesforce. What mattered was that six people with no developer among them could create a field, map it, and change their minds if needed. The tool was never the constraint.
Somebody had built the old sequence inside what Salesforce allowed them, and after enough years nobody could separate the two any more. Asking why a step existed got us the honest answer: because that’s how Salesforce works.
When a system can’t be changed by the people using it, the constraints of the system quietly become the requirements of the business.
It took me another few years, and a second CRM, to understand what it meant.
The same instinct, blocked
Years later I’m at E.ON, working between the Swedish market and a solution management organisation in Essen. Dynamics 365, rolled out centrally, with its own IT organisation behind it.
Sweden wanted to launch new eMobility products. That needed extra fields. In Klarna terms this is a quick fix before lunch.
Here it was a change request. A written description. A justification. Bundled into a work package. Later, nothing would be accepted without a business case attached, which in practice meant inventing a revenue figure for a database field. Then came the actual work: explaining to people in Germany why Swedish customers identify themselves with BankID, and why a Swede who does everything else in ninety seconds on his phone will simply leave if you hand him a printable form to fill out. When you live here, none of that is strange. If you don’t, it’s inconceivable.
Some of it got built. But the distance between the person who knows what’s needed and the person allowed to build it was enormous.
Three hundred lines, three times
The Swedish e-mobility product catalogue was about three hundred lines in Excel. Dynamics had a bulk upload feature. Back then it wasn’t enabled, or it wasn’t working. Today I don’t care any more. But what it had was a preview grid where you could paste in line by line, and save. Crude, error prone, but it worked.
I did that twice. The better part of a week each time.
Once because supplier prices moved due to a new currency exchange rate, and once because the organisation decided that the product names didn’t match what people were used to seeing on invoices. And you couldn’t edit names after the fact, if you can believe it.
The third time, I said no. Not dramatically. I said this costs me almost a week, it’s not the best use of my time, and Germany should fix the upload. The Swedish head of sales who requested it, said fine. Immediately. I moved the request to the IT team, where it should have gone in the first place.
What it actually was
It would be easy to tell that as a story about a manager who didn’t understand what she was asking for. It wasn’t.
She had held my role a year earlier, before the system reached Sweden. What she passed on wasn’t a workaround. It was an expectation: whoever holds the role keeps the prices current. Perfectly reasonable. It was formed when that meant editing one table in excel.
Then the system changed, and the cost of that task went up by an order of magnitude. Nobody recalculated it. They just kept assuming.
A system change resets what everything costs. Ideally it would change for the better, but in reality that is often not the case.
People carry on with tasks given under an old system, at the new price. Morale drops.
Everyone notices and someone has to say it. The third time, that someone was me.
Why the question isn’t academic for me
For the ten years I ran my own companies, and for most of the employed years too, the binding constraint was never the idea, and it was rarely the market. It was finding a technical person who understood what I wanted to build, and who brought that little extra step. That extra thinking to make it practical for the user, and then the challenge for me as the founder was keeping that person. Everything I could envision had to travel through somebody else’s hands, their availability, their patience, and their interpretation of what I meant.
All that is gone today. I build the things myself now, with agentic coding, and they are considerably more complex than the CRM work I’ve described here. AI assisted coding and all tools around this space freed people to start building whatever they could think of, and to start businesses on the things they built. That was the promise I followed. I’d still say I haven’t achieved it. -yet.
I want to be clear about what working with AI actually means, because the enthusiasm around it gets it wrong constantly. AI didn’t remove the hard parts. Understanding a user well enough to know what should exist is still hard. So is launching, and iterating once your product meets people who don’t care how it was built. Those parts still rest with humans and human judgment.
AI can do a lot, and it can make things quicker. It can generate and create. But the steering, the iteration and the actual decisions, how something reaches people and how they will perceive it, should still be taken by humans.
What I do with this now
I build AI tools and systems, and I always start with the most important question: who will work on this, and how does it fit into what they’re already doing?
If the answer is a vague one about the team, it’s pretty clear nobody will be there to maintain it. If one poor person gets selected, the entire responsibility rests on them. So ideally every tool is fully operational, production ready, low touch and low maintenance. Anything that isn’t simple to use, people will stop using within a year. Or they will start building workarounds all over again, the very ones the tool was supposed to replace.
The question I always end on is the one in the title: who can change it next quarter? Depending on the answer, I can judge whether we need to go back to the drawing board.


