Digitization 2.0: Building AI-Ready Organizations

We need to talk about that AI use case you’re trying so hard to get production-ready. Here’s the truth: that particular use case is all but irrelevant. Yet the future of your organization depends on wether or not it succeeds. To understand why, let’s look at where we stand today, and where we’re headed over the next few years.

Phase 1: Custom AI for Custom Tasks

For decades, we’ve been seemingly stuck when it comes to AI. Every use case required a highly customized model, trained on internal data for highly specific tasks. I’ve personally been involved in a few, from training neural nets to identify key points in livestock based from a camera feed, to pose estimation via the TrueDepth stream from an iPhone.

We’ll call this “Phase 1” of AI: custom models for custom tasks. It is only thanks to the invention of large language models (LLMs), and more specifically the transformer architecture, that we’ve finally broken into “Phase 2”.

Phase 2: Generalized AI for Custom Tasks

Phase 2 is unique in that we’re able to use generalized AI on custom tasks. It means companies are able to apply AI without needing to custom train it to their data. What is more, the amount of money and brain-power invested into AI research yields a constant stream of improvements that companies can just tap into, at no additional expense.

The challenge in Phase 2 is that we still need custom implementations of generalized AI to complete custom tasks. Virtually any mildly complex tasks requires infrastructure setup, custom agentic loops, prompt engineering, context engineering, evaluations, logging and tracing, etc.. While the cost of achieving intelligence has gone down massively, the cost of implementing intelligence is still fairly high.

Phase 3: Generalized Implementation for Custom Tasks

What we are already beginning to see on the horizon is the onset of Phase 3: generalized implementation for custom tasks. OpenAI has recently moved into the workflow automation space with AgentKit, spinning up a tool similar to n8n, which itself is closely related to Zapier. These tools enable users to create workflows via a UI, requiring very little to no coding. The way OpenAI wants to compete in the space is that you are now supposed to describe a process, rather than create it via the UI, and the workflow will be generated for you, along with the necessary integrations.

In Phase 3, the cost of implementing AI will diminish. However, it is still your organization’s responsibility to identify and automate processes, albeit with massively reduced overhead. This is where we turn on the crystal ball and look a little into the future.

Phase 4: Autonomous Process Discovery and Automation

Once “text to automation” is sufficiently robust, the natural next step will be autonomous process discovery and automation, which we’ll call “Phase 4”.

Back in 2017, I worked with banks to identify process suitable for robotic process automation. As a newly minted technical consultant, I would shadow and interview employees to discover repetitive tasks that followed strict rules. My biggest learning at the time was that non-technical people struggle immensely to analyze and describe a process in a way that is required for automation.

In the previous phase of “text to automation”, that ability is still a bottleneck for agentic adoption. In Phase 4 however, the AI will self-identify processes and propose automations on its own. What can only be described as “process mining on steroids”, human ability is finally removed as the bottle neck for AI implementation.

Phase 5: Self-Improving Systems

At this stage, Phase 5 is a mere formality: with AI already self-automating processes and virtually every standard process being managed by AI agents, we can expect AI to self-improve workflows and processes, as they are no longer constrained by boundaries of human knowledge, expertise, or communication.

Why This Matters Now

So why should we care and why is this relevant now, when we are just getting settled into Phase 2? Well, remember how I said everything depends on the success of your current AI use cases? That is because the true benefit to any organization is not about the gains achieved by improved efficiency at all, that is simply “nice to have”. Instead, it is all about the learnings, the tools, and the organizational transformation that are implemented along the way while building out your first AI use cases.

It’s effectively “Digitization 2.0”. Similar to how we had to move documents, data, and communication into the virtual world for companies to benefit from software and the internet, we now have a similar task in front of us to become “AI ready”. But now it is not about data, it is about the processes and systems themselves. Without it, no organization will be able to transition to Phase 3 or 4, which is where the true unlock lies. They’ll be outcompeted, outmaneuvered, and eventually fail entirely.

Let me be very clear: in Germany, “Digitalisierung” is still something many companies are working on (and the government is hopelessly failing at) in 2025. The potential unlock of AI is several orders of magnitude greater than that of the internet (see “Software is Eating Labor” by A16Z). This time, international competition is fiercer that it has ever been.

We don’t have another 25 years for our organizations to become “AI ready”. There is only one path forward: Commit today, commit fully, and be hell bent on making sure that those AI projects are successful.