A quick recap, and an uncomfortable addendum
Over the past few months, we've sketched out a clear, logical path:
- In Episode 11.0, we saw that building your own intelligence infrastructure on open source is doable.
- In My Data, My AI, we argued that your unique company knowledge is your real moat.
- And in Intelligence Ownership Becomes a Business Case, we finally ran the numbers: your own fine-tuned models often beat Big Tech's generic, rented AI on quality and price.
A quick clarification up front, to avoid misunderstandings: when we talk about “intelligence” here, we are of course not grading your employees' IQ. We mean the technological capability of your AI system, i.e. the interplay of models, your company knowledge and the applications built on top of them.
So you could say: the target picture is set. Keep the knowledge, own the models. Let's write a classic three-year plan, define the architecture, pick a vendor and execute. Right?
Wrong.
When I look at how companies are currently trying to implement this strategy, they all run into the same invisible wall: the problem is no longer the question “rent or own?”. The problem is that everything this decision rests on moves much faster than the decision itself.
There's a reason this post is called “AI as a moving target”. It's not simply that technology keeps evolving. The exciting part, and the dangerous one, is that the basis for decisions shifts massively while a company is still planning, procuring or implementing.
This post doesn't offer a ready-made solution (that comes later); we'll turn to those topics in upcoming posts. It lays out the complex reality that any future strategic solution will have to measure up against.
The great illusion of stability
Classic IT strategy works like building a house. It rests on silent assumptions so obvious that nobody says them out loud anymore:
- The ground (the technology base) is stable.
- You choose a builder (vendor) once.
- You plan years ahead (architecture & procurement).
- The house stands, and the costs can be depreciated over years.
- The building permit (regulation) is obtained once and then filed away.
AI is not building a house. AI is navigating a raging river whose current changes daily. No wonder very few are making headway, as a new Gartner study shows.
That doesn't mean, however, that no stability at all can be created with AI and we simply have to accept that everything is constantly wobbling. The crucial point is: not everything can or should be stable. The strategic foundations must be rock-solid, while certain technological components deliberately remain swappable.
A procurement programme that easily takes two to three quarters in enterprise IT ends up deciding today on the basis of a model, a price point and a compliance interpretation that simply no longer exist by go-live.
This gives rise to six fundamental problems that boardrooms need to understand:
Problem 1: The half-life of AI models
“We're training our own model now!” Good approach. But anyone who painstakingly trains an AI model for their catalogue management today will find, six months from now, that a new open-source model hits the market that is twice as good out of the box and half as expensive to run.
Does that contradict our thesis from the last post? Why did I recommend “Intelligence Ownership” to you if the model is outdated in six months? The answer:
Ownership remains right and important, but owning a specific model is not, on its own, technological independence.
The point is to own the knowledge, the data and the decision-making capability yourself, without tying yourself permanently to a particular technical implementation.
Intelligence Ownership is not a static state. It's a process.
If you haven't built this process so that you can switch painlessly when the next foundation model arrives, two years from now you'll own a highly specialised intelligence running on a system that perhaps nobody can or wants to maintain anymore. That risk has to be assessed honestly.
Problem 2: The invisible dependency (Vendor Lock-in 2.0)
You might think that whoever owns their model's weights is free. Only partly. With AI, the dependency simply moves one floor down. So where exactly does it arise?
An AI model is ultimately just a “brain in a jar”. To be useful, it needs a memory (vector databases) and hands to operate tools (agent frameworks). The big cloud providers make it incredibly easy to use their proprietary memory and their tools. But then you're giving away your intelligence and everything that makes you you.
The problem: the model itself may be open source, but if your application is deeply entangled with AWS's managed retrieval service or Azure's agent framework, swapping out the “brain” won't help you one bit. The switching costs for the rest of the ecosystem are enormous. You have to learn to distinguish between using a vendor and being structurally dependent on a vendor.
Problem 3: The hardware misunderstanding
We all know AI needs massive computing power, above all GPUs (graphics processors). But you absolutely have to distinguish between training (teaching the model) and inference (querying it day to day).
- Training is like a five-year university degree. It demands massive, focused concentration. For that, you build expensive, highly specialised GPU clusters.
- Inference is like working a busy front desk. You need quick reflexes, multitasking and pragmatism.
Companies are 95 percent inference operations. Yet they often plan their infrastructure like a research lab. And once AI starts acting as an agent that plans on its own, reads databases and triggers workflows, the load shifts back to classic IT and CPUs anyway. The question is no longer “Do we need GPUs?” but “Who's planning the right mix for our day-to-day?”
Problem 4: Regulation on a treadmill
The EU AI Act is not an inspection sticker you get once before going live. It applies in stages, is role-specific and is continually being refined by new standards.
That doesn't mean every technology switch has to trigger a months-long, completely new compliance review. The idea is rather this: your compliance has to be set up so nimbly that it keeps pace with technical change. When you switch model or vendor, documentation and risk classification must be adaptable in a modular way. Anyone who treats regulation as a one-off project phase at the end of the pipeline has misunderstood the problem.
Problem 5: The new bottleneck isn't called “compute”
AI-assisted development keeps making standard code cheaper and faster to write. That's good. So the bricklayers are getting cheaper. But we're short on city planners and plumbers.
What's becoming extremely scarce are people who can keep an AI platform running in production.
- We need plumbers (data, security & platform engineers) who lay the pipes (APIs, data pipelines, vector stores) securely and with high availability.
- We need city planners (AI architects, governance & FinOps experts) who design the big picture so that the shiny new AI district is connected securely and in a privacy-compliant way to the legacy ERP system from 2009.
This isn't just a recruiting problem. It requires a complete overhaul of the skill profiles in your core IT.
Problem 6: The meta-problem: what is still stable at all?
When we have five moving targets at once (models, vendors, hardware, regulation, skills), one thing becomes crystal clear, and it's the central message of this post:
The stable object of your IT strategy can no longer be a particular model, a particular vendor or a particular architecture. What has to be stable instead is the organisation's ability to swap technological components in a controlled way and adapt to change.
That's an uncomfortable insight for anyone who just got a slick, set-in-stone five-year architecture picture signed off by the board. The question “Who owns your intelligence?” is necessary, but no longer sufficient. The new follow-up question is:
How fast and how cheaply can we switch technologically when the ground shifts beneath us?
The litmus test: 6 questions for your strategy
Before you commit to any concrete organisational or technical solution, your AI strategy must be able to answer these six questions with brutal honesty:
- Model switching: How long does it take us to swap the AI model behind a production use case without rebuilding the entire process? Days, months, or “nobody knows”?
- Depth of dependency: Which of our AI applications could we not detach from their current cloud provider within a quarter? (And do we know that for sure, or are we just guessing?)
- Infrastructure profile: Are we planning our server landscape for training or for inference? And who decides the mix: the business side, our IT or our cloud provider?
- Compliance agility: If we swap a model tomorrow, how much of our AI Act documentation remains valid?
- Skills gap: Who in our organisation can run an inference platform and also knows how our historically grown access management (IAM) works? (If the answer to this question consists of exactly one name, you have a massive risk.)
- Sensing: How do we find out about relevant technological or regulatory changes? From our current vendor's marketing newsletter, or do we have our own systematic process for that?
Whoever has answered these six questions for themselves understands the real playing field. Answers to “How do we solve this?” are material for the next posts.
Until then, please take this one guiding principle with you: The answer to the dynamics of AI is not to give up on strategy, but to understand strategy differently. You no longer lock in a particular technology for five years; you build an organisation and an architecture that can go along with technological change in a controlled way.
And that brings us to an inescapable organisational reality: for companies already living genuine agile management, the motto now is: buckle up, you need to get even more agile; we're talking about adaptation cycles measured in weeks. And for the organisations that have so far dismissed agility as a buzzword and remain stuck in classic waterfall projects, rigid annual budgets and multi-year rollout plans? For them, this dynamic is no longer just uncomfortable. It's becoming business-critical.
Live long and prosper 😉🖖
Music reference (1 -> we've got 100Miles this year 😉 and 2 -> AI strategy is like a Miles Davis piece: the direction is set, but the band has to react to changes and improvise in real time. True agility)
