Well that turned out to be a miserable experience.
Not at first though. At first, it felt great. I felt amazed. I was stunned by the power of generative AI. The development power I had access to now. And with the first results in the pocket, I started dreaming about what do do next. What to build next. How I could fit that new technology into my work. Into my private life.
And then the aha moment kicked in. AI had succeeded. I failed.
For some time now, I have been playing around with Claude. Code, Cowork. So when several colleagues asked me to prep a ‘state of AI’ presentation, it felt like a confirmation that I was on the right path. Enthusiasm ignited and that night, I started prepping the presentation. With AI. I made good progress and quickly started to work on the live demos I wanted to incorporate in the story.
For that, I fired up Visual Studio code and asked it to built me a CRM system. Given the audience would have a serious background in business applications, I thought it would be interesting to show them how easy it had become to build applications like that. Accounts, contacts, activities, leads and opportunities. Tabs with related information. Process flows. Modals. Reports. Dashboards. Claude Code took off and pretty quickly delivered a working CRM system to me. Several extra prompts from my end and I felt I had what I wanted to have.
Eureka!
I shared a quick note with the organizing committee that I made good progress with the presentation. And on the side shared that I had produced a CRM system. Let’s see how they’ll respond is what I thought.
Several hours later, I realized I had made a big mistake. I had asked Claude to build me a CRM system. But I had asked Claude to build me a CRM system as per my definition of a CRM system. Lists, Cards. The known entities. The flows I had worked with before. Everything in accordance with the specs of a CRM system as I knew it.
Never in the process did I challenge my own definitions of what a CRM was supposed to be. Sure, a CRM system, and any other system of record thrives on structured data. Running a CRM system based on a bunch of Word-documents stored in an equally large number of folders is difficult for a machine and likely impossible for a human being. Structured data you need. But what about the interface? Do we need lists, cards, process flows and all that? Or could we instead survive and thrive with nothing more but a chat based interface at the core, grounded in structured data with a reasoning engine on top of it? Not only to retrieve and process information but for that chat based interface to take effective action and help me do more with less.
An interface like that could and should be enriched ‘on the side’ by relevant related information. At the core though, all you need is a way to convey wat you want the system to “do”. And no, scrolling through lists and opening the right records, for more scrolling and clicking to happen, is not the most effective use of your time and brain capacity.
Which reminds me of a tagline I read the other day. Nothing new, nothing fancy, but super relevant in the year 2026 and the era of AI.
The best interface is no interface.
Anything else is a distraction. Emails, teams messages, web interfaces and the likes. Clutter. Detours that stand between intention and action.
I don’t know if the world is ready yet for neurally-infused interfaces where all you have to do for something to happen is to think about it. I don’t know if I am ready for that yet and certainly some proper guardrails would be required in that scenario. But more and more these days, I am skipping the endless filling of text boxes on web pages. Instead, I’m verbally brainstorming with Copilot and Claude to help me structure my thoughts.
Fortunately it took Claude just one more attempt and one API key. It now has a CRM copilot agent embedded that can act as a primary way of navigating the system. Whenever asked, it leverages Claude in the backend to process information, take action and update the structured dataset underneath.
Can’t wait what the future holds.
