In the first half of 2025, social media and the news were full of absurd ideas, such as “Human-written code is dead” or “Future requirements are written by machines.” Hype was running completely wild as part of the great AI gold rush. As we are descending from the peak of the hype curve in the second half of 2025, important substance is emerging from behind the noisy gold diggers.
This article is about the tools we use to create software. Artificial intelligence is currently showing potential to improve these tools by an order of magnitude that will affect society economically and socially. Startups like Bolt, Lovable, and Windsurf have set new growth records with their AI-assisted development tools. It tickles my curiosity: Is it purely gold diggers driving up the price of shovels, or is there actually gold out there?
We know from previous technological waves like the internet and mobile waves that adoption is an indicator of economic and social changes. So, let’s look at three different levels of adoption for AI-assisted coding and see how it could potentially affect our software product development practice in Denmark.
Level 1: Better IDEs for developers
Although it sounds relatively boring to the outside observer, even 20-30% productivity improvements for individual software developers will lead to major changes. Over the last 30 years, the shortage of software developers in Denmark has been a bottleneck in our software development practice. Development environments (IDEs), programming languages, compute platforms, and education have improved over those 30 years, but productivity improvements never kept up with demand. We have therefore been busy sending tasks to India, Vietnam, and other countries where capacity is higher and prices lower.
If we increase productivity or development speed by 20-30%, we eliminate many of the negative consequences of outsourcing abroad, but new challenges also emerge:
- Greater variation in developer competencies: Auto-generated code is good if it meets our sustainability requirements. That is, it must be maintainable and further developable. Auto-generated code is bad if it leads to bloatware and an architecture that drives us deeper and deeper into the sand. Everything from software education to software tools must be refactored to ensure productivity improvements push us in the right direction.
- Lack of good product management and design: Product management and UX design capacity can very easily become the new bottleneck. If we can’t specify good solutions, we can’t develop them either. Garbage In, Garbage Out. We can naturally hope for corresponding productivity improvements here, but they are harder to spot on the horizon.
If we succeed in solving these two tasks sensibly, we will be able to create sustainable digital transformations across our society. When standing in the middle of the noise from all the gold diggers, it’s good to pull out Amara’s Law: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
The startup Windsurf illustrates the approach in scenario 1. They focus precisely on the IDE and were acquired by OpenAI for $3 billion in early June 2025. But positioning themselves as an IDE may be the first step on a longer journey, which brings us to level 2.
Level 2: New ways of collaborating
Another possibility is that AI-assisted coding and vibe coding move tasks between professions or create entirely new software development roles. Companies like Lovable and Bolt are trying to open programming up to a broader group of people.
Tasks that previously sat as tedious standard tasks with developers can potentially be moved to other professions like designers or product managers or a new group of junior developers. Experienced developers can thus focus on the 40-50% of tasks that are either technically complicated, have security implications, or affect the performance of the overall solution. The result is a new division of labor and higher motivation across each profession.
Such a change in responsibilities will ultimately require a completely new way of integrating all the tools – from discovery (e.g. Dovetail and Figma) through development of user stories (e.g., Jira and Miro) to coding (Github, IDE, etc.) and feedback across teams (e.g. Slack and Mixpanel).
The result at this level depends on how much prompt-generated coding can help us make quality software at reasonable prices. In addition to the problem of hallucinations and the quality of auto-generated code, we also need to solve the cost profile of using large language models. It is still expensive in terms of server capacity usage and is not sustainable in the long run. It will be exciting to follow improvements in the underlying metrics over the next 12-24 months.
This brings us to the last level.
Level 3: The majority of all code is generated automatically
Here we’re out in gold rush hype curve territory. We are told that software developers become architects and supervisors, while AI handles the majority of actual coding. Likewise for product management and design: Machines do all the work.
This is the utopian scenario where all the technical and economic challenges with current technology are solved. The scenario is also the prerequisite for us to achieve Artificial General Intelligence, that is, the theoretical concept in AI research where computers can perform intellectual tasks at the same level as humans.
Although Elon Musk and a lot of viral posts on LinkedIn would like us to believe otherwise, we are not heading toward this scenario within the foreseeable future.
Experiment with artificial intelligence… thoughtfully!
The analysis of the three levels reveals my recommendations to software product companies and those of you working with software products. Personally, I start with level 1 and experiment with AI development tools. Challenge yourselves with AI-assisted development tools in your teams. Learn what works and have an open conversation across professions. Try a few different tools and understand their limitations. Especially the latter is important, so you don’t generate a lot of unsustainable code based on garbage requirements.
Level 2 is a bit further out, but we can already be curious about organizational and social change today. One of my recommendations to product managers is to experiment with vibe coding for prototype work but remember to reap the value by reviewing the outcome in your retrospectives with your colleagues. Vibe coding is cool but does it provide the value you expect?!
Ignore level 3 completely and shift focus to a sound product management practice now. When coding hopefully becomes easier, the ability to define the right product becomes even more important.
