This is a personal copy of a column published in IEEE Software (May/Jun 2026). Republished with permission.

AI, anyone? A decade ago, the conversation was all about big data and machine learning. Today, the focus has shifted to generative models as a general-purpose technology. But what is actually required for successful adoption? This issue focuses on “The Impact of AI on Productivity and Code,” and this column argues that outcomes depend on readiness. Let me introduce three experts who have given this careful thought. Felix Dobslaw leads a new one-year master’s program on AI transformation. He is joined by Lucas Gren, who drives this change in practice at the med-tech company Getinge, and Erik Sterner, with whom he offers a popular free and open online course on practical generative AI. Together, we discuss AI transformation requirements across three dimensions: organizational, individual, and technological. – Markus Borg

Felix Dobslaw, Lucas Gren, Markus Borg, and Erik Sterner

Is Generative AI (GenAI) really transformational for the software sector? The signals point in opposite directions. On the one hand, we keep hearing that many AI initiatives fail – MIT Press has amplified that narrative [2]. On the other hand, some organizations report dramatic productivity gains alongside major workforce reductions. Klarna is a widely cited example, describing a shift from roughly 7,400 to about 3,000 employees in just two years while maintaining efficiency [3]. If both stories are true, AI readiness must be the real differentiator. In this column, we discuss the requirements organizations must meet to realize the benefits of AI.

AI Transformation Readiness

Professional software development is, at its core, the art of building and working with useful abstractions. Over the last two decades, we’ve steadily moved from imperative programming—spelling out how to do something step by step—toward more declarative styles that focus on what we want. Concurrency handling is a good illustration. GenAI extends this trajectory by turning intent itself (that is, human-level instructions) into an interface. As a consequence, a much broader group of people can now participate in building software.

Certain characteristics, capabilities, and markers seem to separate organizations that succeed from those that stall. In a race full of hype, teasing apart signals from noise is a shared challenge for researchers and practitioners alike.

The technology is moving so quickly that a cautious, evidence-first stance feels almost impossible. We have also changed our language, often talking about AI as if it were synonymous with GenAI, and forgetting that the term Vibe coding was coined only recently in February 2025. New Large Language Models (LLMs) arrive monthly. Agentic architectures evolve weekly. Interfaces change how people actually work—often faster than we can study them.

It feels a bit like the early COVID era, when medRxiv took off as labs rushed to share findings—except this time we’re not observing a primary and visible impact like an unfolding pandemic. Instead, we’re trying to understand a secondary one: what happens to organizations, work, and productivity when GenAI becomes broadly available.

And we’re not just changing the wheel while driving. We’re arguing about the destination. Are we trying to stay on the road and use AI to optimize existing processes, or are we rebuilding the entire vehicle – turning it into an airplane, and attempting to take off, in motion?

We believe the implications are immense, simply because of the new ways humans can now communicate intent to machines. As the interface shifts, the transformation spans every area where software exists today, begging the question: are organizations ready for it? And what qualifies as ready?

AI transformation readiness is inherently socio-technical. It succeeds only when the social system (people and their organization) co-evolves with the technology. But readiness for what? Which parts of development do you intend to support with AI: coding, documentation, architectural analysis, or test case generation? This question is seldom spelled out, yet the answer shapes what readiness means and differs sharply by context. We structure the discussion around three dimensions that cut across these choices: organizational, human/individual, and technological.

Organizational Readiness

Organizational readiness starts with a genuine openness to change—even radical change—and it must be visibly supported by executives and employees alike. It also requires clarity on what you want to achieve with AI assistance: which parts of development to support, and where the payoff justifies the cost. That analysis should guide your choices, regardless of which solution provider you pick. From there, it becomes practical: people need room to experiment, along with low-friction access to the tools, data, and support that let them turn ideas into prototypes.

In large companies and high-reliability or heavily regulated domains, the usual structures often aren’t built for this. If every proof of concept is forced through production-grade requirements processes, governance gates, or legal reviews, experimentation stalls before it can demonstrate value. In those settings, readiness often means creating parallel innovation tracks—bounded sandboxes and pilots where teams can explore safely without being crushed by premature rigor.

Further, establishing a dedicated support function for GenAI at a sufficiently high level in the structure can, if done well, remove friction and speed up ideation across the whole organization [4]. One key responsibility of such a function is to stay up to date with the already diverse and growing palette of tools, integrations, and models offered by GenAI providers.

It’s also easy to aim too low. Operational use cases are obvious (and an internal chatbot can be genuinely useful), but the transformation potential is often strategic: helping leaders synthesize signals, explore scenarios, and make decisions with more context and less delay. If GenAI is confined to the “efficiency” bucket, the organization will miss its greatest leverage. Value creation is not fixed; with GenAI, companies can produce more customer value, so reducing ambition to “same value, but cheaper” misses most of the opportunity.

If GenAI is confined to the “efficiency” bucket, the organization will miss its greatest leverage.

Finally, AI transformation readiness heavily depends on how well an organization understands itself—its roles, its processes, and how responsibilities and decision rights are delegated. The clearer the internal map is, the easier it becomes to see where AI can identify waste, augment work, and enable realistic automation. Just as importantly, it shows where AI would only create ambiguity or risk. That understanding also enables prioritizing AI investments and conducting a sober risk-benefit assessment before scaling pilots.

And while agility used to be a competitive advantage, GenAI turns it into table stakes. The pace of change is simply too fast for slow, linear adaptation. Staying competitive now requires the ability to learn, adjust, and redeploy capabilities continuously. As software becomes a commodity, flexibility must extend to software services and products themselves – for instance, in software-as-a-service dependencies, where solutions can be created in-house quickly and tailored to local needs. There is, therefore, a greater need than ever for effective agile leadership across the organization.

Individual Readiness

Humans are at the center of all software change. Competency is central to individual readiness, especially the capacity to keep learning as roles and tools evolve. Upskilling and reskilling depend on it. In fast-moving software contexts, self-regulated learning becomes critical: setting goals, seeking resources, and adapting without formal training.

How people feel about AI and how willing they are to engage with it also shape readiness. Transformation follows what people want and how they collaborate in practice. It is therefore essential to meet employees where they are to understand how they think and how they relate to new technology and its implications.

Last fall, we introduced a free, open online course on working and learning with GenAI [8]. To understand individual readiness requirements, we asked participants at the start of the course a simple question: “What are your thoughts and feelings about AI in professional practice?”

Figure 1 summarizes the results. Overall, 59% of respondents expressed uncertainty or negative feelings towards GenAI. Notably, this sample comes from a population that voluntarily signed up for a 25h GenAI course. If anything, we would guess this group to be more positive than the broader workforce – but that remains to be investigated.

If uncertainty and negative sentiment are this prevalent among motivated participants, individual AI readiness clearly depends on addressing these concerns. Training programs play an important role, but so does transparent and continuous communication about AI strategy within the organization. This includes how the organization interprets ongoing developments in AI, and openly states what is unknown.

Technological Readiness

AI is not a quick fix for something that is broken. On the contrary, solid engineering practices are what make an organization technologically ready for AI transformation. Existing quality assurance processes are your safety net and foundation. As the cost of adding code approaches zero, organizations can quickly spiral into the deepest technical debt they have ever seen. We bet that many reckless projects will confront existential risks.

When introducing AI, rapid feedback loops on new or modified code are what save you from steering off the road. Once an initiative is underway, continuous feedback and follow-up at the program level let teams adjust scope and ambition as they learn—as with any innovation effort. Basic building blocks such as configuration management, CI/CD, and high-quality test suites must be the backbone of the transformation. If you’re not there yet, return to step one. Experience reports increasingly show that AI assistance tends to amplify existing weaknesses rather than compensate for them.

Among these building blocks, testing deserves special mention. While the others are more binary in nature, testing quality lies on a spectrum. Broad AI adoption requires efficient and effective test suites. They must provide good coverage, and they’ll be run more often than ever before. Don’t let AI loose unless you can verify that the results align with your goals [5]. Specifying requirements as test cases and test-driven development get another boost now. Also, test that your tests are meaningful – poorly aligned AI can unfortunately remove test cases just to make the suite pass.

Another consideration relates to the quality of the pre-AI code itself. As with human developers, LLMs perform better on simple, easy-to-read code. Using refactoring as a proxy for development activities, our recent research shows that AI is far less likely to break functionality during refactoring when code health is high [1]. In other words, our new virtual coding buddies also get lost in complex code. When planning an AI rollout, we recommend starting with pilot projects in parts of the codebase that are “AI-friendly.”

Are You Ready For This?

How do you know whether you meet the requirements discussed so far? Here, we outline four building blocks that we consider critical for a successful transformation. In the separate textbox, we take a broader perspective.

1. Obtain a clear understanding of the needs and bottlenecks in the part of the organization to be transformed. The potential for value creation with GenAI is maximized when related organizational processes, roles, and responsibilities are well understood. Start by examining constraints on a case-by-case basis and focus on simpler ones with favorable effort-outcome ratios [7].
2. Nurture employees’ autodidactic abilities. Organizations need to take responsibility for creating environments that support continuous learning and exchange, and foster a culture of knowledge sharing and experimentation that encourages initiative.
3. Make space for internal R&D. Relying solely on external advice means lagging behind. Change is moving too fast for that. Organizations must invest internally in structures for knowledge exchange, systematic experimentation, and active participation in the broader AI debate.
4. Organizations need technical mechanisms to continuously assess risk, observe AI behavior, and log outcomes over time. As agent autonomy increases, this becomes critical. Deployment will never be a one-off event. Logging must be continuous and deliberate. Tooling and infrastructure are needed to make guardrails and quality gates scalable. For example, Model Context Protocol servers for the agents, code analysis in the continuous integration pipeline, and human oversight where it pays off.

Organizations around the world are figuring out how to take a pole position in this AI transformation race. Those who understand the potential of AI for their specific needs, and succeed in embracing it in inclusive and empathetic ways, will be better equipped to navigate competition and inevitable hurdles along the winding road ahead.

We’d be happy to continue the discussion. What are your thoughts? Where is the world heading? Please reach out and share your perspective!

References

• [1] Borg, Hagatulah, Tornhill, and Söderberg. Code for Machines, Not Just Humans: Quantifying AI-Friendliness with Code Health Metrics (2026), To appear in Proc. of the 3rd International Conference on AI Foundation Models and Software Engineering.
• [2] Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025). The GenAI divide: State of AI in business 2025. MIT Nanda, 1-75.
• [3] Prestone Fore. AI enabled Klarna to halve its workforce—now, the CEO is warning workers that other ‘tech bros’ are sugarcoating just how badly it’s about to impact jobs. Fortune. Available online at: https://fortune.com/2025/10/10/klarna-ceo-sebastian-siemiatkowski-halved-workforce-says-tech-ceos-sugarcoating-ai-impact-on-jobs-mass-unemployment-warning/, October 25, 2025.
• [4] Gren, L., and Feldt, R. (2025). Cross-functional AI Task Forces (X-FAITS) for AI Transformation of software organizations. In Proc. of the 29th International Conference on Evaluation and Assessment in Software Engineering, pp. 793-796.
• [5] Gren, L. and Dobslaw, F. (2026). The Expert Validation Framework (EVF): Enabling Domain Expert Control in AI Engineering. To appear in Proc. of the 5th International Conference on AI Engineering – Software Engineering for AI.
• [6] MacAskill, W., and Moorhouse, F., (2025) Preparing for the Intelligence Explosion, arXiv preprint arXiv:2506.14863.
• [7] Wang, S., Yu, Y., Feldt, R., and Parthasarathy, D. (2025). Automating a Complete Software Test Process Using LLMs: An Automotive Case Study. In Proc. of the 47th International Conference on Software Engineering, pp. 747-747.
• [8] Course page:
  https://www.chalmers.se/en/education/programmes-and-courses/continuing-professional-development/open-online-courses/working-and-learning-with-generative-ai/