What is Software For?
Why vibe-coding replicas of commercial software is a mistake
Less than a year ago I wrote a piece describing how I vibe-coded a customized version of Qualtrics in about a week. That was my own front-line experience of the SaaSpocalypse and I concluded that we were entering an age of personalized, hand-crafted AI applications. A world where the build vs. buy decision was fundamentally transformed by the ease with which AI-coding could build almost anything. Recently, I read a story about a guy who was re-creating Photoshop using vibe-coding (if I remember, he figured it would take a week or so) and my basic reaction was — that’s stupid.
💥 Master AI & Tech Skills in 3 Months!
💰 One Subscription. Multiple Courses. More Savings.
🔥 Get Up to 50% OFF | ⏰ Limited-Time Offer
👉 Enroll Now & Start Learning
Stupid not because it makes sense to go on paying Adobe (though for some reason I do), but stupid because re-creating software that mimics the old way of working with applications now struck me as deeply mistaken. Why would I want a product like Photoshop with all its clunky mechanisms, its inscrutable left navigation toolbar, its fantastical array of little popup controls, and its layers of hidden complexity? Nobody new to Photoshop would use any of this — they’d just use its kind of clunky AI bolt-on. All anyone really wants is a way to bring photo-editing into the AI workflow.
That doesn’t necessarily contradict my thesis in the Qualtrics piece — after all, I built a version customized to my needs and preferences — I wasn’t trying to mimic Qualtrics even if I was trying to replace it. But in the last six months, my thinking about what most software is for and how it should be built has changed radically.
Go into the wayback machine to spring 2026, and AI coding was forcing a complete re-appraisal of how development works. If you’re a programmer or a vibe-coder, it’s been like one of those interplanetary journeys that spin faster and faster around a great mass until one is shot out at unimaginable speed toward some incomprehensibly far landing spot. But while coders have been on that journey, AI has been rapidly swallowing more of everything else. Somewhere in the recent releases of frontier models, AGI-like capability has emerged — systems capable of doing almost everything in the digital realm pretty well without any additional training.
That’s changed software…again.
Because there’s no point trying to build a custom application for something your off-the-shelf AI chat, co-work or agent can do perfectly well. Which makes me wonder what custom software is for right now? And the only general answer I could find is as a way to bring a problem into the AI-sphere.
Let me put this in concrete terms. Ten years ago, I started a company to build a crowd intelligence product. That product ingests data from a variety of sensors (lidar, camera, Wi-Fi, etc.) and helps users understand what people are doing in physical spaces. When I started that company, our vision was to extend the realm of digital analytics to the physical world — to re-create Google Analytics in a new realm. Our initial focus was retail — optimizing stores — but when my company got acquired, that focus changed to using the technology to optimize fan experience and safety at sports venues and live events (because that’s what the new company does).
The product is a robust reporting and dashboarding tool that takes most of what Google Analytics did and adds some powerful tools optimized for geo-located (as opposed to digital) data. In the past couple of years, we’ve added a bunch of AI-enabled features to it. Some of those are embedded in the product and provide better data highlighting and exploration. Others live outside the product and do more advanced things like simulation or tackle specific agentic questions in the live event world.
I’ve spent the last month re-thinking our product roadmap. It’s a humbling experience because almost everywhere I look, I see that current-gen AI can out-perform a user with our platform. There’s just no point using AI to produce dashboards and reports in a world where the best user of that data isn’t a person but the AI itself. Sure, I can build better dashboards, and I can build them a lot faster than I used to. But isn’t using an AI to build dashboards like asking a master mechanic to build a better toolset for someone deeply unmechanical to build a car with? Shouldn’t the mechanic just build the car? And drive it?
On the other hand, it’s not easy right now to use AI for real decision-making in a stadium or live environment. There’s a bunch of different and complex data sources, there are real-time data streams, any decisions need to be vetted and approved, and the communications flow is complex. Those are all big challenges in a world where the window for making decisions that matter is often measured in minutes. Off the shelf frontier models could do the core work and decision-making, but it’s almost impossible to leverage them well in the necessary timeframe. Good agents with the requisite integrations don’t exist, and there’s no way to control agentic recommendations and weave them into the business process. In a world where most decisions need to happen in real-time (the live event world), the cycle time from data to recommendation to decision to action is the most critical element in a system.
Get Gary Angel’s stories in your inbox
Join Medium for free to get updates from this writer.
Which suggests that the best thing our product can do (aside from sourcing crowd-intelligence data which is still too expensive to build using frontier models) is to bring all of that data to a frontier model, help people use that model well, and then facilitate the process of moving from AI-recommendations to implemented decisions in real-time. Our product needs to become a tool for enabling AI-usage, not an AI-enabled analytics tool. The difference is profound and leads to a very different vision of the product — a vision as a kind of control-center or stadium operating system.
I don’t think that shift is even a little bit unusual.
I’m not saying that EVERY piece of software you might build right now should be a tool for facilitating AI-usage. But in the enterprise world, the exceptions will do little more than prove the rule. Instead of trying to figure out how to bring AI into our product, I need to be thinking about how to bring our product into AI. Until recently, I had the problem backwards and I think a lot of people still do.
In my latest product mocks, there are no dashboards, no reports, and while you can still ask AI questions, that’s not fundamental to the product. It isn’t meant to be a verticalized chatbot or a way to explore the data. The product is almost entirely focused on decision-making and execution. AI creates the plan, makes recommendations, monitors the live event, suggests changes to configuration and staffing, surfaces options to the right people in the organization, and feeds its own learnings back into the next plan. The platform provides the data substrate to the AI, surrounds it with the organizational knowledgebase, adds the communications infrastructure to and from the AI and across the org’s decision-flow, and adds the decision-tracking substrate that lets the enterprise MAKE and EXECUTE decisions: before, during, and after a live event.
The whole design is based on the assumption that AI is the primary consumer and user of the data, and the product needs to facilitate that both upstream and downstream. Upstream, it needs to give the AI the data and context it needs. Downstream, it needs to facilitate the flow of information and recommendations from the AI through the rest of the organization.
For most readers, the important point isn’t specific to our product. Our product is just an example of a broader lesson that holds true for almost every class of enterprise tool. If you’re thinking about how to use AI in a product to make it better at doing what it did, you’re probably making a mistake. With current-gen AI, we don’t need software to do most of what it used to do.
It’s the same with that Photoshop application. All those clunky low-level photo-editing tools are worse that useless. We don’t need to be tinkering with that stuff because the AI is better than most of us at doing it.
Of course, this begs a huge question. Whatever we know right now about AI won’t be true in a year. AI is swallowing everything in technology faster than a Sumo wrestler in an all-you-can buffet. So yes, it’s quite likely that general purpose enterprise AI-agents will eventually swallow my highly-verticalized product vision of a decision-engine. In the past, the specifics of our vertical and its relatively niche status would have provided a formidable moat. But AGI essentially de-moats every such friction point. All you can really hope to do is get there first and make hay while the sun shines. And on the bright side, with all our AI-coding, we can harvest a helluva lot of hay in a short summer.
Next fall, I’ll likely be writing how a vision like this no longer makes sense. All I can really hope for is that the vision holds that long. That’s our life these days. Fortunately or not, 12 months is the far horizon beyond which a product strategy need not peer.
Thank you for being a part of the community
Before you go:
👉 Be sure to clap and follow the writer ️👏️️
👉 Follow us: Medium
Disclosure: This post includes affiliate and partnership links.
