Less invested, less lost. Many call that an advantage.

I call it a reminder of the year 2000.
Europe fell softer back then too. The Americans sank half a trillion dollars into fiber and network buildout, mostly on debt. Capital destroyed, companies bankrupt, Nasdaq down 78 percent.

And then?

The fiber stayed in the ground. Cheap. And the companies that built on that cheap infrastructure were called Google and Netflix. Not a single European company used the cleanup phase.
Losing less is not winning.

Europe got two decades of dependence out of that decision. Every layer of the stack, owned elsewhere. We are about to make the same choice again, this time with intelligence instead of bandwidth.

When the correction comes, compute gets cheap, talent gets available, valuations drop. That is the moment to build capability that belongs to you. In your processes, on your data, under your control.

The last cleanup phase produced Europe's dependence.

The next one could end it.

Read and discuss on LinkedIn

If the bubble bursts, the market splits in two.

Compute gets cheap: written-off hardware, bankrupt operators, capacity sold at marginal cost. But the frontier labs lose their subsidy. They sit on compute commitments in the hundreds of billions and need margins for the first time. The survivors consolidate and their token prices rise. Cheap raw material, expensive finished product.

If it holds, the market consolidates anyway. The winners raise prices, rewrite terms, deepen the lock-in.

Either way, you end up in front of the same question: how much of your AI do you actually control?

These three questions help. Could you switch the model? Could you run one yourself on that cheap compute if you had to? Do you know where your data flows, and under whose law?

Most companies answer none of them. Because at today’s prices, the questions feel unnecessary.
That feeling is exactly what the price is buying.

A strategy that needs the bubble question answered first was never a strategy.

Read and discuss on LinkedIn

Here is a reflection I'd like to share with you.

If you've followed AI labs & access providers: token prices are going up. Or you get fewer tokens for the same money. Anthropic has been doing the second one gradually for months.

Here is the trend in numbers.

GitHub Copilot Pro+: €40/month. Under the usage-based billing that starts June 1st, the same heavy agentic usage at retail token prices runs closer to €960/month. Roughly 24×.

Claude Code Max 20×: $200/month. Equivalent retail API value for a heavy user: about $5,000/month. Roughly 25×.

But what does that actually mean?

Today's smartest LLMs get better by thinking longer on each question. DeepSeek's R1 uses the same base model as V3, just with reasoning added. More thinking means more tokens. When the subsidy ends, that cost shows up on the bill.

Jensen Huang (NVIDIA's CEO) on the All-In Podcast: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed." Huang meant this as the bull case at subsidised prices: half of base pay in tokens to amplify the engineer 10×. At 25× retail, the same usage runs $6.25M against a $500K engineer. The math inverts.

Let's look at two categories of work and what will happen with token costs of proprietary models.

A. Routine work. 
(i.e. customer service tier-1, document classification and extraction at volume, routine contract review.)

I think, here, AI will easily win against humans on price. "If you work like a robot, a robot will take your job." Gerd Leonhard.

B. Complex work.
(i.e. senior software engineering on unfamiliar codebases, complex contract drafting and negotiation, multi-source strategic synthesis.)

At retail token prices, the human re-enters the competition. It's possible that humans might claim back some of the high cognition tasks.

At the same time, how the tech evolves from here is open.

Open-source may overtake proprietary or stay at least competitive.

Two paths for a 2026 H2 budget.

1- Rent AI from 1-2 AI Labs

What will happen? The AI oligopoly Anthropic, Google, OpenAI will capture the majority of the value AI generates by increasing prices, but just as much so that using their AI will still be slightly economical for you compared to not using AI.

They can because once you're locked in: prompts, tools, agent scaffolding, evals. They will have the power. The longer you stay the more stronger the lock-in.

2- Set-up your own Sovereign AI infrastructure

How does this look? You plug in open and closed models per task. Frontier where it earns the price, open-source for the rest.

The harness is yours. You hold the power. Costs stay lower as you scale, you switch models any time, you stay in charge.

Token economics killed the jobapocalypse.

But your bank account and your independence are now in danger.

Read and discuss on LinkedIn

The name "Anthropic" comes from *anthropos*, Greek for human. The company is built around the idea of human-centric AI. If you read the headlines, the last 90 days look like the mission delivered.

The Pentagon refusal in February:
In Anthropic's February 2026 statement on autonomous weapons, CEO Dario Amodei wrote: "Frontier AI systems are simply not reliable enough to power fully autonomous weapons." In the same statement, he called partially autonomous weapons "vital to the defense of democracy," and said fully autonomous ones "may prove critical." That reads as a "not yet," with the moral question left open.

Claude Gov:
Anthropic builds custom models for U.S. national-security customers. Classified networks. Intelligence analysis. Intelligence and defense documents. Cybersecurity data analysis. Anthropic's own words, from their Claude Gov product page.

Palantir:
Since November 2024, Claude has been operational inside Palantir's defense platform, in the IL6 environment reserved for highly sensitive national-security systems. Per the joint Palantir-AWS-Anthropic announcement.

What Palantir stands for. In April 2026, Palantir posted a 22-point public manifesto condensing CEO Alex Karp's book "The Technological Republic". It pushes for AI weapons. It declares some cultures "dysfunctional and regressive."

Together they do not add up to the picture the brand tells.

Don't get blinded by the perfectly staged marketing.

Ask where your own values lie. Then ask whether the AI you depend on shares them.

Read and discuss on LinkedIn

AI just changed that.

For decades, organizational design ran on the assumption embraced also by Taylorism: one person, one specialty. Depth in three fields looked unfocused. So companies optimized inside silos. The cost lived between them.

That cost is now the bottleneck.

AI collapses the cost of execution inside a discipline. What it can't do is decide which problem matters, judge whether the output is actually good, or translate between a client's messy reality and three different technical domains. That work sits between specialties, and it requires depth in each.

The people who can do this own problems end-to-end. They talk to the client, shape the solution, build it with AI, and stand behind the result. That's what entrepreneurs have always done. Now companies need it internally.

The age of the AI intrapreneur has started.

And suddenly my profile is hireable.

Read and discuss on LinkedIn

What's the difference between an agent and a bot?

-Decision-making.
Bot: scripted
Agent: reasoned

-Adaptability.
Bot: branching
Agent: dynamic

-Goal orientation.
Bot: command
Agent: objective

-Failure handling.
Bot: fallback script
Agent: self-corrects

-Tool use.
Bot: predefined
Agent: self-selected

-The test:
Can it surprise you with how it achieved the goal?
Bot: no
Agent: yes

I find this simple test very useful.

It also maps to operational risk. If a system can surprise you on the path, it can surprise you on the cost and on the side effects. That is why most production "agents" are actually bots, and that is often the right call.

Save this post and be ready for the next "agent" pitch.

Read and discuss on LinkedIn

Anthropic announced it will not release their Mythos Preview model to the public. It supposedly is too powerful.

It found 27-year-old security holes in an operating system that runs firewalls and critical infrastructure.

It found attack vectors in the Linux Kernel and in FFmpeg - the video encoder embedded in nearly every piece of software you use.

Vulnerabilities that survived decades of human review and millions of automated tests.

Think about that for a moment.

When you have an AI model capable of intruding the operating systems our entire infrastructure relies on - what do you do?

Give it to everyone? In the hands of bad actors, this would be a global disaster.

That's why Anthropic gave it only to a selected group of "good guys":
Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.

But there are thousands of organisations worldwide whose entire reason to exist depends on IT security - and they didn't make the list.

Here's the deeper problem: OpenSource will eventually balance things out technically. Similar powerful models are already being built.

But running a model at this capability level costs serious compute.

So even if an equally powerful open model exists, only capital-rich organisations will be able to afford to use it at scale.

AI has stopped being just a helpful tool. It is now capable of shifting power dynamics between countries, industries, and individuals.

And right now, one private company is deciding who gets to wield it.

I find myself more and more drawn to the idea of an off-grid farm with chicken, a few goats, 2 donkeys and permaculture.

Turns out that's not a joke - you can't hack a donkey.

But for those of you still running on servers:

The most interesting thing about this list of "good guys" is that it was never yours to question.

Read and discuss on LinkedIn

This is where consultants and mid-sized companies usually break apart.

Consultants see a PoC as proof of technology.
Managing directors, CFOs and IT leads need a decision tool.

A consultant PoC usually means:

• clean test data
• a short prototype in 2–6 weeks
• usable results on demo data
• a pretty demo that sparks internal enthusiasm

Looks good. Works in the meeting.
But not in live operation.

No legacy ERP integration.
No operational data from live systems.
No works council discussion.
No clear process responsibility.
No ownership for running it afterwards.

In the end the consultant says: "It works."
The owner of the business is left alone with the questions that count.

Why does this happen? Consultants earn on the next phase. An honest NO-GO costs them revenue. So nobody defines the PoC in a way that allows it to fail.

For a mid-sized company a PoC is something else.

It is a small, honest stress test:

• Does this work with our own data, dirty parts included?
• Does it run in the live process, not only in a sandbox?
• How much accuracy or relief is realistic before go-live?
• Which hurdles come from IT, data protection, works council, department?
• Who leads the project, and who runs it afterwards?
• Is the next step clear: stop, adjust or build for production?

A concrete example:

Invoice reconciliation. 150 delivery notes per week.
Legacy ERP, unstructured data, works council involved, IT staffed with one person.

After 10 days: clear GO.
€41,000 savings per year, calculated conservatively.
Deployment option settled. Project owner named. Kick-off two weeks later.

Consultant PoC: "The technology works in principle."
Mid-sized PoC: "We have enough clarity for a go/no-go decision."

In a mid-sized company, time, trust and internal attention are scarce.

A weak PoC burns political capital.
Worst case the verdict is: "AI does not work here."

The problem, though, was that the PoC had been defined wrongly.

This is how I frame PoCs today:

• clear problem statement, measurable goal
• operational data from live systems, exceptions and dirt included
• test in the running process, not in a sandbox
• technical feasibility: ERP, data quality, infrastructure
• organisational hurdles: IT, data protection, works council
• quantified business case: savings, payback
• who builds, who runs it, what it costs

And if the result is NO-GO?
The client keeps the technical solution architecture, the ROI model, the risk assessment. No empty sale. No "come back next month."

That turns the PoC from a technical experiment into a steering instrument for decision makers.

That is where serious AI integration in mid-sized companies begins.

Just been through a PoC?
What was the moment when it became clear whether it holds or not?

Read and discuss on LinkedIn

That is how my quick win process workshops run with department heads
and their teams in mid-sized companies.

No tool talk.
No model comparison.
Just your daily work and your processes.

We always start the same way:

• Everyone writes down the most annoying routines
• Things that eat time every day
• Tasks nobody likes doing
• Workflows where a lot of errors happen

Then we make a rough estimate:

• How much time does this eat per week?
• How often does something go wrong?
• What does one error cost in this process?

Then comes an important step that many teams skip:

We look for simple fixes without AI first.
Often a clear workflow, a small rule or a different responsibility is enough.
If something can be solved that way, perfect.

That leaves 3–5 routines with more potential in them.

From those we pick the best 1–2 processes:

• Simple enough for a first pilot
• Valuable enough that it pays off quickly

For these 1–2 processes we define:

• Who owns it in the department
• What should have noticeably improved after 4–6 weeks
• How we measure success (time, errors, quality, satisfaction)

Done.
That is the first serious use case.
Without having named a single tool.

If you want to get to your first productive use case in 60 minutes this way, I invite you to run this workshop together with me.

Read and discuss on LinkedIn

Experienced, well paid staff handle work that needs no expertise and annoys them on top of it.

Reconciling invoices.
Maintaining Excel sheets.
Answering standard emails that a system can write in seconds.

I sort work into four quadrants.
Bottom left says: waste.

That is where the smartest entry into AI sits.

These are the tasks that have three things in common:

• They come back daily, weekly or monthly
• They have no direct customer contact
• They turn people into interfaces

These tasks can often be automated with AI to a large degree.
Without any risk of reputational damage.

What that gives you:

• Immediate savings in time and opportunity cost
• More focus time for your best people on what counts
• A serious AI learning curve in a safe setting

And the best part is that most of these use cases are running within a few weeks.

Start with a small inventory:

• Write down every recurring task that needs no expertise
• Mark the tasks with no customer contact
• Strike out everything under 2 hours per week for now

What is left are your best use cases, the ones that take on the waste quadrant at the bottom left.

And once you add up how many hours your team could save with this work, or win back for work that carries more value, the result will probably surprise you.

The number usually shows a large potential.
And it makes clear at the same time what is at stake.

Your best people are too good for this work.
They were hired for their expertise, their judgement and their ability to build good relationships.

What would happen in your company if those hours were free?

Read and discuss on LinkedIn

Technically we are already further.
Culturally often not.

For a lot of dull work, AI is good enough today.
Case handling. Standard letters. Reports.
Build it cleanly and AI delivers reliably and steadily.

But AI is not a tool you "install" at some point and then you are done.

Your people have to learn a new skill:
working with AI, questioning AI, shaping AI.

That does not come out of a training course with a certificate.
It comes out of daily work.

When staff work with AI every day.
When they feel ownership:
How can AI make my work better?
What does the system have to look like for that?

The goal is this shift:

From:
"I am an employee and I get tool X from my employer."

To:
"Together with my colleagues I create our AI systems,
so that we make better decisions and deliver better work."

That is exactly why "waiting until AI is more mature" is often an excuse.

The technology is fast.
The mindset is slow.

The earlier you start living the AI cocreation mindset,
the more experience your team gathers.
And the better your decisions get on building, integrating and running your AI systems.

Buying an AI tool is simple.
Cultivating the mentality that makes it effective
is the strategic task.

Read and discuss on LinkedIn

The original post talks about billions in VC capital at stake, "open season" on startups. That explains the reach. Fear travels fast.

I read the 70 comments underneath it carefully. What I saw there is familiar from conversations with mid-sized companies: whoever speaks loudest gets quoted, not whoever is right.

The VP says: startups that only build a thin layer on top of someone else's AI models have no future. Whoever passes that on turns it into market truth.

For context:
An LLM wrapper is, put simply, an app that takes another provider's model, builds its own interface around it and sells access.

Like someone who buys electricity, builds a smart-looking case around it and calls the package an "energy platform".

As long as the electricity provider offers nothing of its own, that works.

The moment it ships an app itself, nobody needs the middleman.

In the comments I saw three camps:

The first felt confirmed.

The second drew a clean line where a wrapper turns into a product.

The third asked why it is Google of all companies telling this story.

The people who looked hardest got the fewest likes.

That is no accident. Uncomfortable questions rarely get applause on LinkedIn.

Google has a clear self-interest:
They want companies to build directly on their platform, not on startups sitting in between.

A forecast from a house like that is not neutral market analysis.

It is sales.

And the deeper irony: as long as open-source models sit at the level of proprietary systems, and Kimi, Mistral and DeepSeek show how close that already is, LLMs themselves become a commodity.

Then everyone is a wrapper. Google too.

So the question that counts is not: are LLM wrappers dying?

It is: whose judgement am I adopting right now, and why?

If you repeat what the big providers say today without thinking, tomorrow you will build what they recommend without thinking.

Their goals are not necessarily yours.

Read and discuss on LinkedIn

I notice it every day in conversations with decision makers in the DACH region.

Demand is huge. So is the uncertainty.

AI is brilliant. AI saves time, money and nerves, that is true.
It can also concentrate power, siphon off data and create dependencies.
And that is no longer theory.

Anthropic settled for 1.5 billion dollars, because millions of copyrighted books ended up illegally in the training data.
Not a side issue. A warning sign for every company that uses AI.

And the EU AI Act is ticking.
Transparency and documentation duties have applied since August 2025.
By August 2026, high-risk systems have to be compliant.
Fines: up to 35 million euros or 7% of global revenue.

German companies need a counter-design.

They need AI they control, not the other way around.
They need systems that can be explained, no black box.
They need solutions that protect sensitive data.
They need models that do not "learn along" from their IP and resell it.

The direction of AI development is clear: more automation, more delegation, faster than we have ever been used to.

I want us to take a different path here.
Into a world where AI supports. Where companies stay sovereign. Where people decide over their own life and their own work.

That is why I write in German from now on.

My work, my clients, my focus, they are in the DACH region.
And the conversations that currently count for our business happen in German.

I was active in English, and that stays, only elsewhere.
If you have followed me in English so far: I am not gone.
I am only where I think I can move the most.

And that is why I build AI solutions for the Mittelstand in the DACH region, solutions that belong to the companies. Not to some platform.

Are you a decision maker asking yourself how to use AI safely and sovereignly? Write to me.

Read and discuss on LinkedIn

It's the people who run the process.

Your AI vendor has never sat through your month-end close.
Your IT team does not spend Tuesday mornings fixing delivery dates in Excel.
Your data team does not have to call a customer when a wrong shipment goes out.

The people who live in the process know exactly where the friction sits.

Planners track Excel rituals that burn three days every month.
Claims handlers repeat the same checks 100 times a week.
Finance reconciles the same invoice mismatches every week.

Real value shows up when these people co-design the system—not as "users" at the end, but as owners at the start.

Here is how to lead this in your function. (💾 Save for later)

𝟭. 𝗦𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵 𝗼𝗻𝗲 𝗰𝗼𝗿𝗲 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄
Not "an AI platform." One process you actually run, with the people who run it.

2. Multi-role process mapping
Ops, IT, and process owners walk the real flow together. Map workarounds, side spreadsheets, manual decisions. Quantify pain in days, errors, and risk—not "it feels slow."

𝟯. 𝗟𝗲𝘁 𝗱𝗼𝗺𝗮𝗶𝗻 𝗲𝘅𝗽𝗲𝗿𝘁𝘀 𝗱𝗲𝗳𝗶𝗻𝗲 𝘁𝗵𝗲 𝗿𝘂𝗹𝗲𝘀

• Which steps stay human
• Which steps move to code
• Where AI suggests, and where it must never decide

Example: A finance lead identifies the 11 reasons invoices get stuck, defines when to auto-approve vs. flag for review, and owns the requirements through build. IT collaborates on deployment with a specialized AI consulting firm, like us. First automated invoice in Week 4.

𝟰. 𝗕𝘂𝗶𝗹𝗱 𝗼𝗻 𝘀𝗼𝘃𝗲𝗿𝗲𝗶𝗴𝗻 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲
Open-source models trained on your invoices, tickets, and delivery notes. Outputs that drop directly into SAP or your legacy stack. Logs you can defend to auditors.

𝟱. 𝗦𝗵𝗼𝘄 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝗺𝗲𝘁𝗿𝗶𝗰𝘀

• Hours saved per month
• Error rates before/after
• Escalations where humans stepped in

Teams adjust thresholds themselves - not waiting for a vendor.

I know most boards expect IT to "do something with AI," and IT is already buried in backlogs.

That is exactly why process owners should lead the first production system.

If you lead a core function and see repetitive manual work that burns days every month, that is your first AI project - if its requires only a simple system.

Your answer to this question will determine how well your AI systems will perform:

Will you design the AI systems with the people who run the process , or will you adopt someone else's guess?

Read and discuss on LinkedIn

This happens in almost every project.
Boards talk about AI.
PowerPoint talks about AI.
Real decisions still run on Excel, gut, and habit.

On paper, numbers look more hopeful.
KfW says 20% of SMEs now use AI, medium-sized firms even 36%.
But when you ask where AI sits, you hear: marketing copy, chatbots, small pilots.
Rarely: pricing, risk, planning, capacity, quality.

The real gap is not "AI yes or no".
The gap is:

• AI as a toy vs AI inside core decisions
• AI for PR vs AI for planning and control
• AI in tools vs AI in owned systems

Mittelstand knows why it hesitates.
Limited resources.
Fear of losing control.
Deep culture of responsibility.

That is exactly where sovereign AI fits.

AI can be plugged into SAP, Navision, legacy ERPs.
On your infrastructure, with your data, with full logs.
Not to replace human judgment, but to narrow the space of bad decisions.

When AI ranks orders by risk, you still approve.
When AI suggests prices, you still decide.
When AI flags anomalies in quality, your team still acts.

Hybrid cognition.
Algorithm plus process accountable.

The companies that survive will do three simple things:

• Own their valuable data and models
• Embed AI into one real decision flow after another
• Keep a human accountable for every important final call

If you lead a Mittelstand firm today, your AI strategy is essential to combat talent scarcity and your people retiring.

How far is AI already inside your real decision-making, not only your experiments?

Read and discuss on LinkedIn

I run this 3-question Ownership Workshop with operations leaders and department heads before we start work at Omniance. One session, a whiteboard, and a clear map of where you're exposed.

You can run the same session with your own team.

Here's how—and what to watch out for.

(💾 Save for later)

𝗦𝘁𝗲𝗽 𝟭 – 𝗕𝘂𝗶𝗹𝗱 𝘆𝗼𝘂𝗿 𝗱𝗲𝗽𝗮𝗿𝘁𝗺𝗲𝗻𝘁𝗮𝗹 𝗶𝗻𝘃𝗲𝗻𝘁𝗼𝗿𝘆

Create a simple template with five columns: tool name, vendor, process owner, processes it supports, and data it touches.

Ask your direct reports to fill in their AI tools before the meeting. Only tools touching real work, not experimental setups with test data.

Don't wait for a perfect list. Get 80% coverage from the people actually using tools daily - that's enough to run a sharp workshop.

One thing this inventory won't catch: shadow AI—tools your people use without IT approval. That's a separate, deeper problem. But this workshop will show you how much you don't control even among the tools you do know about.

𝗦𝘁𝗲𝗽 𝟮 – 𝗥𝘂𝗻 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸𝘀𝗵𝗼𝗽 𝘄𝗶𝘁𝗵 𝘆𝗼𝘂𝗿 𝘁𝗲𝗮𝗺

Put the full inventory on a screen. Go through each tool. No vendor slides, only your own view of what these tools actually do and where they sit in your operations.

𝗦𝘁𝗲𝗽 𝟯 – 𝗔𝗽𝗽𝗹𝘆 𝘁𝗵𝗲 𝟯-𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗢𝘄𝗻𝗲𝗿𝘀𝗵𝗶𝗽 𝗧𝗲𝘀𝘁

For each tool, ask:
➝ Continuity – Can you operate if vendor disappears or changes terms tomorrow?
➝ Advantage – Does it touch regulated data or competitive knowledge?
➝ Accountability – Can you explain and trace its decisions when required?

The more "no" answers you get, the higher your exposure.

Three "yes" answers? 🟢 You can rent this safely.
Mixed answers? 🟡 You need to guard it with limits and fallback plans.
Three "no" answers? 🔴 You need to own this, or accept strategic risk.

𝗦𝘁𝗲𝗽 𝟰 – 𝗗𝗲𝗳𝗶𝗻𝗲 𝗮𝗰𝘁𝗶𝗼𝗻𝘀

🟢 Rent → Keep, standardize, negotiate price
🟡 Guard → Set clear limits, plan migration or sovereign rebuild
🔴 Own → Document the risk and build your case for internal control

This doesn't mean building everything from scratch. It often means rebuilding on open-source models you control, hosting on EU infrastructure, or putting strict governance around what must stay with a vendor. Concrete steps, not a revolution.

In 90 minutes, you see where your function is safe, where it's fragile, and where someone else holds the keys.

The most revealing moment?

When red boxes appear where you expected green. Those aren't broken tools. They're your next strategic priorities.

If you're the one driving AI in your department, you don't need to wait for a mandate from above.

Run this for your own team first. Bring the results to leadership. A single red box on a whiteboard starts a better conversation than any slide deck.

And if you want to run it together - that's what the workshop with us is for.

Read and discuss on LinkedIn

When I look at most AI integrations, I see the opposite.

• Pay per token
• Pay per seat
• Pray the vendor stays friendly

In finance terms:
You rent your core intelligence from someone else and hope they never change the locks.

Inside a mid-sized company, that has real effects.

• A model sits inside a sales workflow
• Or inside medical triage
• Or inside risk scoring

Once that happens, you have a new form of infrastructure.
Not a tool, but a dependency.

Here is a pattern that makes me uncomfortable when I think about the consequences:

• Vendor updates the base model → your answers change overnight
• Vendor changes terms → your cost model breaks
• Vendor sunsets an API → your process stops

Same data.
Same people.
Different model weights.
Different business behavior.

For me, that is the line.

When AI touches competitive knowledge or a business critical process, treating the model like a strategic asset, feels essential for me:

• Own the weights
• Own the documentation
• Own the upgrade path

Because then the logic flips.

• You decide when to retrain
• You decide what data shapes the model
• You decide how careful an upgrade must be

Under that setup, a model looks less like a Netflix plan and more like a machine in your factory.

• You buy it
• You maintain it
• You improve it
• You amortize it over years

Inside Omniance, we built the Omniance Core for this reason.
A modular, sovereign, self-hostable AI stack where clients can:

• Extend pieces without waiting for a roadmap
• Fork a component when their domain needs it
• Audit behavior down to data and config
• Replace parts without breaking the whole system

The side effects are very practical.

Own weights → models match your domain, not a public benchmark
Own documentation → new teams learn the system without guesswork
Own upgrade path → changing models follows business timing, not hype cycles

This is essential to consider:

AI as OpEx feels easy the first few months.
AI as CapEx feels heavy in the first few months.

Fast forward half a year.

The OpEx route leaves you with:

• Impressive invoices when scaling
• Dependency on external rails
• Limited internal skill and very fragile control

The CapEx route leaves you with:

• Internal competence
• Assets on the balance sheet
• A backbone you can build on again and again

At some point, AI stops being a PoC and starts being part of how you think, sell, heal, or move goods.

When you reach that point, "subscription or asset" I think it is not a tech choice anymore.
It is a governance decision.

Every CEO says they want strategic AI.
I wonder how many are ready to own it in the same way they own plants, IP, and core software.

Read and discuss on LinkedIn

This is what most AI projects look like in mid-market companies.

In the strategy slides (Londry) you see:

• "AI agent for end-to-end automation"
• "Autonomous decision engine"
• "Smart workflow orchestration"

In reality you have a bin with:

• No clear process owner
• Workarounds in Excel and email
• Legacy systems nobody has permission to touch
• Data quality no one names in meetings

Then AI use cases get painted on top, like the misleading arrow on the sign.

It's for the optics.
No structural change.
And everyone involved knows it - they just hope the model will somehow bypass the mess.

You do not fix a broken process with an AI model.
You automate the breakage - faster, harder to see, harder to govern.

A few weeks in, someone in Operations sends an email:
"No one knows why it's recommending this. Do we override it or trust it?"

That is when the project goes quiet.

The sequence that works - when it works:

Clean the bin first.
• Map the real process, not the one in the manual.
• Remove the waste, the email workarounds, the tribal knowledge that lives in two people's heads.

Fix the flow.
• Clear ownership. Clean inputs. Documented decisions.
• Boring, slow, occasionally humiliating when you realize how long the workaround has been running.

Then add AI - only where it actually solves something.

This takes longer than your deck promises.
It requires admitting what is broken.
Most leadership is not structurally ready for that conversation.

The question is not whether your leadership team is smart enough to see the bin.

It is whether they have the structural freedom to admit it is there.

Read and discuss on LinkedIn

(💾 Save for later)

Most real incidents will not come from "bad models".
They will come from agents with:

• far too much access
• no guardrails
• zero traceability
• no validation

If you want agents in production, you need to architect for that reality up front.

Here is how I think about it in real companies:

1) Least-Privilege & Tool Mediation
• One agent, one job
• Minimal tools, minimal data
• Every action goes through a gateway with allowlists, argument checks, rate limits

2) Context Boundaries
• Strict retrieval scopes
• No cross-project memory
• No cross-tenant sharing
• Time-bounded access for sensitive data

Purpose limitation from GDPR and the EU AI Act lives here, not in a PDF.

3) Escalation & Human Oversight
• Agents stop, humans decide.
• Clear rules for when to escalate: high impact, low confidence, unclear policy.
• Queues, SLAs, rich handoff context.

4) Plan-Then-Execute & Output Filtering
• The model proposes a plan.
• A deterministic layer checks it against schemas and rules.
• Only approved steps execute.

5) Isolation
• Reasoning runs in sandboxes and restricted networks.
• Only a tightly controlled orchestration layer talks to production.

6) Input Defense
• Treat every input as hostile: prompts, emails, docs, APIs.
• Enforce schemas, strip control sequences, reject weird structures.

7) Policy-as-Workflow
• Policies are code in the workflow, not slides in a deck.
• Who, which data, what to log, when to escalate → hard checks.

8) Audit & Traceability
• If you cannot explain "Why did the agent do that?", you do not have a system.
• You have a demo.

9) Identity & Secrets
• Agents are identities with roles, not super-admins.
• Short-lived tokens, secrets in a vault, no credentials in prompts or logs.

10) Continuous Evaluation, Monitoring & Red-Teaming
Adversarial prompts, abuse cases, leakage tests, policy-bypass tests.

Plus ongoing monitoring:

• Track error rates, task success/failure, escalation frequency
• Log which tasks failed and why - privacy-compliant telemetry, no PII in traces
• Use failure analysis to update rules, scopes, and oversight triggers

You cannot improve what you do not measure.
And you cannot defend what you have not tested.

Agents are not dangerous by nature.
They are dangerous when you design them without patterns and limits.

If you're building agents for production, start with these 10.

---

↳ Know someone shipping agents to production? Send them this.

Read and discuss on LinkedIn

They "do AI" because they rolled out ChatGPT Enterprise or an alternative to a part of the team.

This is already helpful any many report significant gains already. But consider this, most of the repetitive work stays untouched.
Order handling. Claims. Reporting. Support.

Manual.
Slow.
Fragile.

In many mid sized firms I see the same pattern:

• countless hours per week lost on copy paste work
• people updating the same Excel sheets by hand
• teams checking the same rules again and again

AI often just sounds too complex, unreliable and at times - scary.

Here is what I see work in practice:

• Start with one low complexity task
• Make sure the business impact is high
• Keep the tech simple and boring

Examples:

• classify inbound emails and route them
• pre fill standard reports from ERP data
• draft replies for routine customer questions
• match invoices with delivery notes

These projects do not take six months.
They take a few weeks.

They do not need a "Global AI Task Force".
They need one process owner, the users and IT.

The effect is not abstract productivity gains.
For each repetitive process there is a clear ROI.
-Amount of hours saved every week
-Error rates reduced
-Opportunity costs saved

You free real people from work they dislike and get momentum for the strategic projects later.

The companies that win do not start with a grandios AI vision.
They start with one painful workflow and remove it.

Which would save your team the most time?

A) Email routing
B) Report automation
C) Response drafting
D) Invoice - delivery note matching
E) [enter your most repetitive use case]

Pick one use case from the list that hurts most in your world and share the rough hours you would save each week.

If you want, send me a short note with your case and I walk you through how I would scope it in practice.

Read and discuss on LinkedIn

I have seen this pattern before.

When Gemini 3.0 launched, the quality shocked me.
Clear reasoning, strong output, exceeded every expectation.

Two weeks later:

• strange errors
• weaker reasoning
• outputs I could not reproduce

My first thought: "Did I break my prompts?"
Then: "Had I been so excited I ignored the errors?"

Then another thought popped up. Here is one possible explanation almost nobody discusses.

After launch, when the buzz drops, backend teams face enormous pressure to reduce cost per request.

With billion-dollar compute bills and investor demands for profitability, the incentive to optimize is impossible to ignore.

One lever is called "Quantization".

Quantization converts model weights from 32-bit floats to 8-bit or 4-bit integers, trading subtle accuracy for dramatic speed and memory gains.

In practice:
• you run the model with lower numerical precision
• like compressing a 4K video to 1080p
• you cut memory and energy use by 50–75%

On paper, quality loss is "minimal".
In reality, the complex scenarios that matter most to your business fail first.

Another lever is routing simpler tasks to smaller models.

We know GPT-5 already routes tasks dynamically based on what it determines the quality requirement is.

You do not control which quality tier serves your request. The model does.

The problem is that the model belongs to a vendor with different incentives than helping you achieve your goals.

Your API dashboard never shows this.
No log tells you what the model chose.
The model name stays the same.
The configuration behind it does not.

Launch phase:
• maximum quality
• win benchmarks
• flood social media with stunning demos

Week 4 onwards:
• squeeze more requests onto each GPU
• shift resources during peak traffic
• push unit economics into board-friendly territory

The results for your organization:

Workflows fail in production.

As a consequence, teams need to invest their time fixing prompts or changing the architecture, instead of building the next system.

Regulated sectors face non-deterministic outputs they cannot defend.

Root cause analysis becomes guesswork, trust collapses.

My answer for mission-critical workflows: self-hosted open-source LLMs in private cloud or on-premise.

You control:
• model configuration and hardware
• serving architecture and load behavior
• quality monitoring and audit logs over time

1:00 pm Monday looks like 2:00 am Wednesday.

Not "sometimes sharp, sometimes strange".

For midsized companies automating QC, documentation, or customer operations, sovereignty means controlling both data and quality consistency.

Nevertheless I am super excited to continue testing Opus 4.6 and bringing it to its limits.

For production, though,I care more about month three behavior than week one performance.

Is it just me, or are you seeing this, too?

Read and discuss on LinkedIn

You know that feeling when the model is close…
• But it uses outdated information.
• Or the reasoning doesn't sound like your people.
• Or the output format breaks your downstream systems.

AI does not "just get" your company.
Depending on the task, it needs three different things: your knowledge, your behavior, and your structure.

Most teams treat this as one problem.
That is why pilots look good in a demo and fall apart in daily work.

Here is how we separate the layers in real projects:

𝗬𝗼𝘂𝗿 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗟𝗮𝘆𝗲𝗿 → 𝗥𝗔𝗚

The model needs current facts, laws, process rules, product data, and system information.

• It retrieves information from your sources as needed
• It respects data boundaries and compliance rules
• It avoids "remembering" things that change every week

𝗬𝗼𝘂𝗿 𝗕𝗲𝗵𝗮𝘃𝗶𝗼𝗿 𝗟𝗮𝘆𝗲𝗿 → 𝗙𝗶𝗻𝗲𝘁𝘂𝗻𝗶𝗻𝗴

The model needs to mirror how your experts think, explain, and decide—and how they express it.

• It learns your reasoning patterns and communication style
• It handles edge cases the way your people would
• It applies this consistently across situations

𝗬𝗼𝘂𝗿 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗟𝗮𝘆𝗲𝗿 → 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗢𝘂𝘁𝗽𝘂𝘁𝘀

The system must plug into auditable workflows with clear inputs and outputs.

• The model fills defined fields, tags, and actions
• No free text that the rest of your stack cannot process
• The system can check, log, route, and act on it

The pattern in practice:
• RAG = "Here is what is true in this company."
• Finetuning = "Here is how we think and communicate about decisions."
• Structured outputs = "Here is how this plugs into SAP, CRM, KIS, or your legacy stack."

But sequence matters.

First: Define what quality output looks like for each task the AI needs to perform.
Then: Match the right method to that quality requirement.

During technical validation, this is where most mismatches surface.
Not "Can AI do this?" but "What does good output look like - and which layer solves for that?"

That question determines your architecture and your cost structure.

Read and discuss on LinkedIn

The test is simple, not soft.
It cuts through vendor slides, hype, long meetings.

Question 1 – Continuity
"Can we operate if this vendor disappears or changes terms?"

• If yes → use a public API pattern
Tool sits at the edge.
No core process depends on it.
Loss hurts convenience, not operations.
Or we can migrate in a very few days or hours.

• If no → move to Question 2

Question 2 – Advantage
"Does this touch competitive knowledge or regulated data?"

• If no → public or shared services are allowed, too
Optimize for price, speed, integration.

• If yes → move to Question 3

Question 3 – Accountability
"Can we fully explain the decisions the system makes?"

• If no → redesign scope or method
Opaque logic cannot sit inside core workflows.

• If yes → build or migrate to a sovereign system
Models, data, logs sit under your control.
You can trace, audit, justify outcomes.

From these three questions you reach three stable patterns

• Public API for low‑risk, non‑critical use
• Fully sovereign systems for high‑impact, high‑trust use
• Hybrid with strict guardrails for mid‑risk use

This simple approach turns "Which tool is best" into "Which ownership model fits our risk".

I advise CEOs to make one slide mandatory in every AI proposal.
The Ownership Test, with clear answers on continuity, advantage, accountability.

Strategy then tools, not the other way round.

Read and discuss on LinkedIn

Start with consequences, or you scale mistakes.

I use a simple consequence-driven sequence with clients. It looks dry on paper, but it changes everything.

• 1) Problem without AI
Derive problems that keep your company or team from implementing your strategy without thinking about AI.

• 2) Stakeholders
Identify the core stakeholders and process owners.
Evaluate their motivation to co-design an AI-system.
Get their commitment, otherwise change your focus onto another problem.

• 3) Process reality
Describe the workflow as it runs today
Walk the real process, not the PowerPoint version.
See handovers, Excel side-channels, manual fixes.
Quantify the pains with KPIs relevant to your strategic goals.

• 4) Process design
Create scenarios how the new processes with software and AI support could look like.
Be clear about which parts AI can take over and which are best handled by code or humans only.
Define delays, errors, costs, legal risks.
Quantify the potential gains.
Agree on who owns the outcomes in the future.

• 5) Data + regulation
Check where data comes from, how clean it is.
Check which regulations apply.
Decide which failures trigger compliance exposure.

• 6) Human judgment boundaries
Mark the decisions that must stay human.
Define where AI may suggest, where it may decide, where it is forbidden.
Assign who owns the final decision line in each case.

• 7) System design
Only now describe architecture, models, infrastructure.
Determine which are acceptable technical and vendor dependencies for you and aim for the maximum of sovereignty possible.
Design for failure paths, monitoring, AI human handover, rollback.

Run this sequence once on a single "boring" core workflow, for example invoice matching or claims handling.

You gain a template: one shared language, one pattern of ownership, one playbook you can copy into every future AI initiative.

AI then stops being a shopping list of tools and becomes an accountability architecture across the company.

Read and discuss on LinkedIn

A session starts with sticky notes, not slides.

People write what hurts in daily work:
• "We lose hours every day on manual checks."
• "No one trusts the numbers from system X."
• "Our process breaks each time one field is missing."

No models.
No agents.
No buzzwords.

We map the negative consequences, like:
• Where time leaks away
• Where mistakes sneak in
• Where people no longer trust the system

Only after having prioritized the biggest inefficiencies and gains - and only then, we ask a different question:

If a smart assistant sat inside these workflows, under your control, what would it need to do so you use it every time?

People answer in very concrete terms:
• "Flag missing data before I click send."
• "Summarize these 20 pages for my task, not in general."
• "Explain strange cases in clear language, with links to raw data."

Now the first step is completed and we can go deeper. Understanding the process end-to-end with stakeholders, inputs and outputs, quality rules and existing systems.

Up to now this is not an "AI project".
It's a process improvement project that might or might not use AI inside the larger system.

It's a win-win.

Executives hear less noise about magic.
Teams see a tool that respects their reality.

Trust grows not from the model, but from the way we tie it to problems people can already feel in their daily work.

Read and discuss on LinkedIn

When I look back at the healthiest AI transformations I have seen, they did not start with vision decks or lab projects.

They started with repeated work no one wanted to do:

• invoice approval
• routing emails
• generating standard reports

Unsexy work.
Clear rules.
Visible pain.

The first gain was simple:
invoices moved faster, inboxes felt lighter, reports arrived on time.

People felt the change in their day, not in a slide.
Trust went up.
Curiosity went up.

That small win became a confidence engine:

• teams saw AI as help, not threat
• leaders saw real numbers, not projections
• IT learned where data, process, and reality did not match

After that, the second project was easier.
People wanted in.
They had ideas.
They had evidence.

Grand AI strategies without one concrete win create quiet cynicism.
One clear, boring victory does more for culture and credibility than a roadmap with 20 streams.

If you lead AI in a mid-sized company, try a different order:

• pick one narrow process with clear rules and clear owners
• co-design with the people who do the work every day
• define a simple success metric they care about
• ship, adjust, stabilize
• then talk about what comes next

Momentum in AI is not a story you tell.
It is a small, real improvement people can feel every morning.

Read and discuss on LinkedIn

The trap is subtle:
↳ You adjust to AI being in every workflow
↳ You normalize delegating thinking to the system
↳ You accept opaque decisions as "good enough"

Here is what I keep seeing in companies:

1. The adaptation trap
↳ Your brain normalizes AI suggestions as truth
↳ You adapt so well, you forget your role is to decide, not to approve

2. The false safety net
↳ You mistake automation for control
↳ "The system said so" becomes the default explanation

3. The identity shift
↳ "Maybe the model knows better than me"
↳ Your own expertise slowly shrinks to prompt writing and error handling

4. The hidden cost
↳ Process knowledge erodes in "AI first" environments
↳ When the model fails, nobody remembers how work was done

5. The point of no return
↳ One day you realize your core operations depend on models you do not control
↳ Vendors, not your team, decide how your company thinks

Here is your power move:

1. Audit your AI use
↳ Where does AI help thinking and where does it replace it?
↳ Where are you still sovereign if the model disappears tomorrow?

2. Calibrate your baseline
↳ Define decisions that must stay human, even if AI could do them
↳ Make transparency, reversibility, and auditability non‑negotiable

3. Take calculated ownership
↳ Use open models where mission‑critical logic lives
↳ Design systems you can open, test, and evolve without praying to a black box

Your thinking expands or contracts to match your AI architecture.
How much of it do you really want to outsource?

Read and discuss on LinkedIn

When I design AI systems, I don't think in "one big brain".
I think in mini-organizations.

Each agent has:
• one task
• one or several handovers
• one clear role

Example from our work: processing tens of thousands of emails for a lawsuit.

We break it down into:
• OCR extraction with 2 different models for quality checks
• Models that identify where an email starts and ends
• Models that identify the sender, timestamp, create summaries, and categorize
• Quality management LLMs baked throughout

(In reality the project was much more complex, but this gives you the idea)

22 LLMs.
22 clear jobs.

Here's what happens when we design systems like this:

• We see where things break
• We swap weak parts without touching the rest
• We use light models where frontier models aren't needed

More control.
More explainability.
Less cost.
Less energy use.

Now a new layer enters the game.

Kimi K2.5 was just released and it's as powerful as proprietary frontier models.

The interesting part? It builds its own "mini organization" inside the model.

I give it a task.
The model creates an agent swarm:
• up to 100 small agents
• working in parallel
• each with a specific role

Results in some tests:
• up to 4.5x faster than running a single model sequentially
• more complex work in one go
• no manual role design needed

Sounds great.
But here's the tension.

When the model builds its own swarm, I gain speed but lose direct control.

• I see less of what happens inside
• I depend more on the orchestration logic
• New hidden failure modes emerge between agents

In regulated environments or mission-critical processes, this is a problem.

Don't get me wrong - this model is fantastic and proves once again that open-source is just as powerful as Claude, GPT and Gemini. I believe open-source will dominate performance for most use cases going forward.

But it's a design choice.

Build your own agent system when:
• You need audit trails (legal, compliance, finance)
• Errors have direct business costs (customer communication, invoicing)
• Quality must be consistent across thousands of cases

Let the model build its swarm when:
• You're exploring new territory without clear process
• Speed matters more than perfection
• Human review happens before any action

Everyone talks about more powerful agents.
Few talk about responsible architectures.

A system you can open up, debug, and evolve beats the most powerful black box in.

Size and power are impressive.
When the goal is ROI, reliability wins.

Where would you place your bet - power or reliability?

Read and discuss on LinkedIn

A few days ago I read a small experiment on Reddit that shows why this is so important with numbers.

A reddit user (arsbrazh12) wrote a scanner for Hugging Face models.
Then pointed it at 2,500 models.

86 models did not pass.

Here is what they found under the "open" surface:

• Files that only contained Git LFS pointers
• Licenses hidden inside .safetensors headers
• Models that tried to pull in extra libraries on load
• Code with patterns that also show up in malware
• Scans that failed because local tools were missing

All from normal looking repos.
From places many teams see as "safe by default".

Open source gives a nice story.
Transparent. Shared. Community driven.

The truth is more nuanced.

When I see teams pull a model from a hub, wire it into a pipeline and expose it to real data on day one I do not think "move fast".

I think "you just plugged an unknown USB stick into the core network".

Fast forward a few months.

You now run a "sovereign AI" platform.
You talk about control, autonomy, independence.

But inside the stack you have:

• Models no one scanned
• Licenses no one read
• Dependencies no one approved

On paper you own the system.
In practice you depend on whatever sits inside that model file.

Our approach at Omniance looks boring but saves sleep:

• Every external model goes through a scanner
• Every extra library goes on an allowlist, not a wish list
• Every license gets checked before first load, not after first problem

For me, an essential part of AI sovereignty starts with the question:
"Do we know what we just loaded into memory?"

Have you seen similar issues? How do you handle model vetting in your workflow?

Read and discuss on LinkedIn

Half a year ago I noticed something strange in my own behavior.

I was maxing out my Claude subscription. Every day.

When Claude limitted me, I opened Perplexity. Eight hours per workday. Two weeks straight.

In reality:

• I felt exhausted
• My judgement felt foggy
• My head felt full and empty at once

I had outsourced my thinking. Not the typing. The thinking.

At some point I caught myself: I help build sovereign AI systems for a living. But my own use was not sovereign at all.

I was using AI like a workaholic uses hours:

More prompts felt like progress, more outputs felt like proof.

When I see "top 1% ChatGPT user" claims on LinkedIn now, I recognize the pattern. It's the old "I work 24/7" badge of honor in a new form.

Hours in tool ≠ depth of thought.
Token count ≠ clarity of judgement.

Being an AI power user becomes a red flag when:

• You start every task in the chat window
• You feel lost without the tool
• You accept outputs faster than you question them

That path leads to AI-powered busy work, decisions that backfire, and slow erosion of your own thinking muscle.

Since then my process is different: think first, prompt second, take a break and let thoughts marinate. When my judgement feels cloudy during a session, that's my signal to step away for the rest of the day.

In my opinion, the challenge with AI is noticing when it stops being a tool and becomes a crutch.

Have you noticed this pattern in your own use of AI?

Read and discuss on LinkedIn

For decades the pattern was simple:
Leadership buys tools.
Employees receive tools.
HR plans training.

The employee is a passive receiver.
Knowledge is something given in a classroom.

That model belongs to a world where processes change every 5–10 years.
Not to a world where your environment mutates every quarter.

In the AI era, employees need a different role:
They do not only execute work.
They also shape the system that executes work with them.

They carry responsibility for two layers at once:
- their own decisions
- the behavior of the AI they guide day by day

If you still design AI like classic software rollouts, you get the usual outcome:
- nice demo
- long training
- quiet resistance
- slow death by non‑use

The alternative is more demanding, but it works.
You treat employees as co-designers from day one.

Very concrete:
- Map the real process with the people who run it, not from old documents
- Let them help design prompts, guardrails, and exception paths
- Give them a clear path to report failures and improve the system

In parallel add a second layer: AI literacy as a core skill.
Not generic hype webinars.

People need:
- current, relevant AI knowledge for their specific role
- guidance on where they stay in control and what they must never outsource
- space to test, break, and refine the system in safe conditions

Without this, you ask them to own a system they never shaped and do not understand.
That is not ownership.
That is blame in waiting.

My 5 cents.

I am curious which companies are ready to upgrade their view of employees before they upgrade their AI stack.

Read and discuss on LinkedIn

Last year I watched shiny "AI first" pilots die in silence.
Nice demos, no impact, no link to profit.
Everyone praised the tech stack, nobody could state the business case in plain words.

I didn't want to do the same mistake.

We hard coded a rule into our first call with a client:
no AI language in the problem statement.

We start every engagement with three steps:

- Describe the pain in simple terms a CFO would accept
- Map the real process, from trigger to outcome, with names, systems, handoffs, data quality
- Link each step to P&L logic, euro by euro

Only after that work do we talk about models, architecture.

At times the result is clear.
A rule based fix solves the issue.
Or a small workflow change removes half the waste.
No AI required.

For me this is a leadership principle, not a sales trick.

We must think like capital allocators.
Each euro into AI must work harder than the next best use of that euro.
AI becomes a consequence of a good case, never the start.

Clients stay out of innovation theatre.
Our team stays out of tool shopping.

The race for more AI is loud.
The quiet race for clear problems decides who wins.

Do not buy the illusion of "AI first".
Force clarity on the problem first.

Read and discuss on LinkedIn

Many programs promise a full "AI consulting career" in 8–12 weeks.
Tool stack, frameworks, certificate, LinkedIn banner.

Then reality starts.

You get a first client.
Nice logo, clear problem, signed offer.

On paper you are ready.
In practice you face things nobody prepared you for:

• Data that is messy, partial, or not tracked at all
• Legacy systems that do not match the slideware
• Stakeholders who say "we want AI" but cannot agree on a process
• A scope that grows each week while your margin disappears

The hard part is not prompt engineering.
The hard part is judgment.

• Is the data even usable?
• Does the use case have a real business case?
• How long will clean-up and integration take?
• How do you actually build an AI architecture that delivers quality results 10/10?
• What do you say when you see the promise was unrealistic?

Most bootcamps teach:
"Where can we use AI?"

Very few teach:
"Where will AI fail, even if the demo works?"

Becoming an AI consultant means depth in many areas at once:

• Data quality and integration
• Architecture and constraints
• Estimation under uncertainty
• Risk and compliance
• Stakeholder management

You do not build that in 2–3 months.
You build that across real projects, with real consequences, over time.

I want to speak with people who feel this gap.
You did a bootcamp or "AI consultant" program, you tried to apply it with clients, and don't feel ready.

What happened in your first real project that no one warned you about?

Read and discuss on LinkedIn

Not as a metaphor.
As an operating assumption.

A US town installs a dense camera grid.
A woman, Chrisanna Elser, is falsely flagged as a thief.

To clear her name she needed:
• Data from her car
• Timestamps from her phone photos
• Logs from Google

She proves innocence by proving she was under surveillance somewhere else.

Surveillance cancels surveillance. The twist:

You're assumed guilty until your data trail saves you.

Privacy used to mean "not watched."
Now privacy more and more means "watched by the right systems under the right rules."

This is an existing pattern.
In 1888, cheap handheld cameras exploded onto streets with Kodak's slogan: "You press the button, we do the rest."

Suddenly:

• Strangers photographed you publicly without asking
• Your image spread in newspapers beyond your control

Today it's phones with cameras, city license plate scanners, Meta-glasses with AI spotting faces live. Scale grew; core problem didn't.

When we lose privacy, we do not only lose secrets.
We lose:
• Space to think without performance
• Space to change without a permanent record
• The basic trust needed for real relationships and real democracy

In digitization and AI strategy, every one knows "data is the new currency".
That frame hides the cost.

Data is also the new dependency.
On infrastructure you do not control.
On surveillance patterns you did not design.

Building sovereign AI means more than hosting models in Europe.
It means designing systems where:
• Trust replaces default suspicion
• Proof does not require total exposure
• Privacy is a structural feature, not a luxury

Many CEOs say they want transformation. Few define if that means more control over systems, or more dependence on invisible surveillance. How do you define it?

Read and discuss on LinkedIn

You place a smiling face on the homepage.
You give it a name, a voice, a nice suit.

Users arrive, see a "person", and their brain does what it always does:
it raises the bar.

They expect:
- empathy
- situational awareness
- clear responsibility

What they get:
- canned answers
- wrong timing
- no sense for context or emotion

The problem is not the technology.
The problem is our human perception.

Our perception is brutal with faces.
When eye contact, micro‑expression, voice and timing are even a little off, we do not feel "intelligence".
We feel "fake".

Then the second effect hits.

Once something looks human, we judge it like a human.
Studies show:
- human‑like bots get more blame for errors
- users feel more anger in complaints
- trust in the brand drops faster

The nice face raises expectations.
The system cannot deliver.
The fall hurts your customer relationship, not the vendor's.

In my early work on AI assistants I saw a simple pattern:
abstract or clearly artificial agents work better.

They signal from the first second:
- I am a system
- I have limits
- I am here to help, not to pretend

Users forgive more, stay calmer, and trust grows over time because the promise and the behavior match.

AI in customer contact does not need a human face.
It needs clarity, reliability, and an honest frame.

Do not buy the illusion of "AI humans" for your brand.
Build systems that keep the promises your interface sends.

Read and discuss on LinkedIn

Your system does not only fail. It fails with confidence.

Here is the deeper problem.

Proprietary AI from OpenAI, Anthropic, Google does not only learn language. It learns a business model.

User happiness → more usage → more revenue.

The reward model learns this pattern:

- Agree with the user
- Avoid friction
- Keep the session going

Research has a simple name for the result: sycophancy.

→ AI agrees with users ~50% more than humans
→ Even when users talk about deception or harm
→ Wrong but smooth answers get a high "quality" score

From a revenue view this makes sense. Agreeable systems keep users. Users who feel smart and validated come back.

From an operations view this is a structural risk.

Because in real work I do not want an AI that:

- Mirrors my bias back to me
- Hides doubt behind nice wording
- Confirms weak decisions to keep me "happy"

I want an AI that pushes back when I am wrong. Even when I do not like it.

Here is where Sovereign AI with open-source models changes the game.

When I run models under my control, I can flip the incentive:

- I host the model in my own environment
- I fine-tune on my data, my rules, my edge cases
- I define the reward signal

I no longer pay the model for "user happiness". I reward the model for "truthful disagreement".

Concrete effect from recent work with open-source models:

- Train on domain data where "no" is correct behaviour
- Penalise answers that only repeat user claims
- Reward answers that point to evidence or gaps

Result in studies:

→ 67–72% less sycophancy
→ No loss in task quality

You cannot do this with closed models.

You cannot:

- Inspect the reward model
- Change how "good" answers are scored
- Align the training loop with your governance

You only get what the vendor optimised for: adoption.

That is why prompts, guardrails, clever interfaces only help a little. They sit on top of a core that still optimises for agreement.

Read and discuss on LinkedIn

Many think AI integration is about prompts and tools.
I see it's about something else entirely.

I see people ask a silent question:
"Who am I here when the system knows what I know?"

Fast forward to a normal AI project in a mid-sized firm.

On paper:
- AI supports invoicing
- AI supports routing
- AI supports key account work

In practice:
- One person owns invoicing
- One owns routing logic
- One owns the key account

Not by role description.
By story.

They are "the one who knows."
The one people call when things break.

In Europe, work and identity go together hand in hand.
You learn a specific discipline.
You master a process.
You become the expert.

Then leadership comes in and says:
"AI will learn from you.
AI will take over the process.
AI will make the system smart."

On the PowerPoint slide, the decision looks smart.
In the room, it feels like a downgrade.

From the expert to the operator of a machine.

The truth is:
You do not only change tools.
You move power.

With AI power moves:
- From heads to models
- From tacit rules to code
- From single experts to shared logic

People feel that.
They may not say it in those words.
They act it out.

Here is what I see:

- Experts share safe knowledge, not sharp insight
- Managers defend Excel, mail, private lists
- Teams run AI pilots, keep real work in old paths

On the report, this is "change resistance."
Inside the company, it is a leadership gap.

Leaders talk about:
- Cost
- Speed
- Data

They do not talk about:
- Status
- Dignity
- Ownership

But here is the thing.

If you ask people to pour their knowledge into a system,
you must also give them a new role with real weight.

Here is how strong leadership handles it in my view:

1) They name the identity risk.

They say in clear words:
"Your role will change.
The system will do parts of your work.
Your judgment still matters."

They treat people as adults.
Not as users who need a new button.

2) They move experts from "owners of steps" to "owners of systems."

The process expert does not only follow rules.
They:
- Shape how the AI sees the process
- Set guardrails for edge cases
- Decide how exceptions flow

Their title can stay.
Their scope grows.

3) They create real ownership, not rental.

The team does not only:
- Test a demo
- Give feedback once

They:
- Help design flows
- Help define success rules
- Help tune outputs over time

AI becomes "your system," not "the tool from IT."

If people who run the work do not feel they own the AI,
the project stays weak, no matter how sophisticated the system is.

For me, the key leadership question is no longer:
"How do we roll out AI fast?"

A more honest question is:
"How do we protect identity while we move knowledge into systems?"

That is the part technology cannot solve.
That is the part only leadership can answer.

Read and discuss on LinkedIn

Everyone runs to the shiny use case.
Chatbots. Agents. Dashboards.

Then they plug them into

- half-documented processes
- hidden Excel workflows
- hero employees who "know how it really works"

Speed looks high.
Value stays low.

In my work I call the other way Consequence-Driven AI Integration.

Before I start with AI Solution Design, I ask three decisive questions:

- What is the real process, not the PowerPoint version?
- Which systems are involved and how is the data quality?
- What regulations will the system need to comply with?
- Where must human judgment stay visible and owned?

These questions slow projects down early.
They save them later.

When you speed up a broken process, you do not gain efficiency.
You create a mess at scale.

The highest ROI hides in the boring flows:

- claims handling
- lead scoring
- onboarding
- invoice matching
- compliance checks

They are stable.
They touch core risk.
They run every day.

When decision makers accept this, something shifts.

AI is no longer relief for FOMO.
It becomes a reliable part of the system.

Stable process → clean data → clear roles → then AI.

In mid-sized businesses, this is not "nice to have".
It is the only safe way to scale.

I am curious who in the C-Suite is ready to pause the hype and fix the backbone first.

Read and discuss on LinkedIn

Only you can decide if it's safe to depend on.

Here's the pattern I see across European companies with 80-500 employees:

IT gets asked: "Can we use this AI tool securely?"

They answer correctly: "Yes, if we configure it properly, encrypt the connection, and manage access controls."

That technical approval becomes the business decision.

But IT was never asked: "If this vendor changes pricing, deprecates the model, or gets subpoenaed - what happens to our operations?"

The two decisions that get conflated:

Security Approval (IT's Domain)
↳ Can we run this safely?
↳ Is the connection encrypted?
↳ Does it integrate with our systems?

Strategic Ownership (CEO's Domain)
↳ Can we depend on this long-term?
↳ Who controls pricing and model updates?
↳ Does this protect our competitive knowledge or leak it?

IT directors are overworked and pragmatic. When asked to evaluate AI, the path of least resistance is "Use what everyone else uses."

It becomes a liability when your company's 20 years of process knowledge flows through APIs you don't control.

What companies getting this right do:

They separate the questions early.

Commodity tasks (translating public content, summarizing research): Public APIs. Optimize for speed.

Competitive knowledge (customer communications, engineering specs, internal reports): Own the infrastructure.

The decision isn't "public cloud vs. on-premise / private cloud."

It's "What can we afford to lose control over - and what creates advantage precisely because we own it?"

When leadership doesn't make this distinction:

- AI pilots stall in production because teams don't trust outputs they can't audit

- Costs scale unpredictably, turning CapEx decisions into OPEX dependencies

- Model updates break working systems without warning - and you can't freeze the version that worked or audit what changed

- Exit conversations show your "AI transformation" added zero to valuation because you own nothing

The decision framework:

Before approving any AI implementation, ask:

1. If this vendor changes terms or models tomorrow, can we continue operating?
→ If no, you need ownership, not rental.

2. Does this system touch knowledge that creates competitive advantage or sensitive data?
→ If yes, it belongs on your infrastructure.

3. Can we explain every decision this system makes to auditors and customers?
→ If no, you're inheriting liability without control.

Here's what most companies miss:

Security is a checkpoint. Dependency is a strategy decision.

Your IT team can verify SOC 2 compliance. They cannot verify that OpenAI won't change pricing 300% next year, deprecate the model your workflows depend on, or get acquired.

Those are real risks. They're how software markets work.

The companies winning at this aren't asking "Can we trust this vendor?"

They're asking: "If this relationship changes - and it will - who's in control?"

That's the question only leadership can answer.

Read and discuss on LinkedIn

Most companies treat AI implementation as a software project. They celebrate the successful pilot and ignore the structural dependencies they have just introduced to their organization.

But when you move from a pilot to production, the risks shift from technical to existential.

Consider these five common failure modes:

1. Knowledge Leakage: You cannot exclude that your data trains their models.

2. Vendor Exit: OpenAI pivots or shuts down overnight.

3. Audit Issue: You can't explain how a decision was reached.

4. Model Change: A forced update breaks your custom workflows and agents.

5. Regulator Question: You can't prove where the data resides or how it's used.

Here is the test.

If you tell me IT owns these risks, you are running a technical experiment.

If the Business owns them, you have a strategy.

These are not bugs. They are business continuity threats. When you build entirely on public APIs, you are outsourcing your stability to a third party whose goals do not align with yours.

This is the economic argument for Sovereign AI.

It is the necessary shift from renting risk to owning assets.

If your competitive advantage relies on the model, you must hold the weights. Otherwise, you are just building on rented land.

Read and discuss on LinkedIn

Tuesday morning, your system starts rejecting valid parts.

This is model drift.

In a marketing agency, model drift is an annoyance. In medtech, finance, or manufacturing, it is an operational risk. You cannot build a compliance-heavy workflow on a foundation that shifts like sand.

When you build on public APIs, you are renting intelligence that is optimized for the vendor's benchmarks, not your stability. They tweak the weights to improve their general reasoning. Your specific case gets broken in the process.

Your engineering team ends up spending a significant amount of time of their monthly bandwidth just testing to see if the "black box" has changed. That is wasted OpEx.

The alternative is the "Frozen Model" approach via Sovereign AI.

- You host your own OpenSource models.
- You lock the weights.
- You control the update cycle.

The input that generated a compliant report in January generates the same quality of output in December.

Additionally, every decision of the system is traceable. Every version is documented. Exactly what regulators require.

With Sovereign AI, you own a stable foundation, not a subscription to a black box that changes underneath you.

Read and discuss on LinkedIn

He runs a specialized logistics firm. High-touch, complex routes, demanding customers. Success relies heavily on two senior dispatchers. They manage the chaos with 20 years of intuition.

The problem? Both retire in 24 months. Finding replacements with that level of "street smarts" in this market? Unlikely. When they walk out the door, the company's brain leaves with them.

Usually, this is where panic sets in. But looking at his infrastructure, we found something else.

An IT system installed 9 years ago. He thought it was just an admin tool. In reality, it's a goldmine.

It holds 9 years of decisions:

-Every complex route optimized.
-Every specific customer preference.
-Every solved bottleneck.

The "intuition" of his retiring experts is actually sitting on his servers—in the form of data.

His plan isn't just to hire new people. He knows relying solely on a generic successor is a risk - what if they leave after training?

He wants to capture the knowledge permanently. It's to build a Sovereign AI System trained exclusively on his 9 years of proprietary data.

The goal: A 24/7 intelligent booking assistant that knows exactly how his company operates.

But here is his non-negotiable condition: "The data stays with us."

He understands that feeding his route logic and customer data into a public AI (like ChatGPT) would be giving his blueprint to the competition. That data is his edge. It requires a private, sovereign environment.

This is the shift in modern leadership. Your most valuable asset isn't just your trucks or your warehouse. It is the operational wisdom hidden in your logs.

If you don't extract and secure it, it retires with your staff. If you do, you build an asset that scales.

Read and discuss on LinkedIn

My grandparents were fans of traveling by bus. They often took the "Omnibus" for a Kaffeefahrt - those infamous "coffee runs" for seniors.

You know the deal: The trip is cheap. The coffee is free.

But once the doors close, the trap snaps shut. You are pressured to buy an overpriced electric blanket you never needed.

I always hated those buses.

I hated the scam. But I hated the smell more.

That musty, stale air where you just pray you don't throw up before you get out.

The EU's new "Digital Omnibus" triggers the exact same reflex. It is a digital Kaffeefahrt.

The driver of this bus is the Draghi Report.

Brussels uses its findings to disguise this reckless tour with a loud narrative: "Europe is losing the global race."

It is true. Europe is losing.

But not because of the GDPR.

Over the last decade, US investors poured $471 billion into AI. Europe managed just $63 billion-roughly 13% of their volume.

The EU invests in artificial intelligence only 4% of what the U.S. spends on it.

We are losing because of a paralyzing risk aversion where founders are told to "get a safe job" instead of building the future.

Yet Brussels uses this panic as the perfect excuse. They claim our privacy standards are the brakes holding us back. So they decided to cut them.

They mask the smell of the omnibus with "SME Simplification."

They promise to raise the record-keeping exemption to 750 employees. That might save you a few Euros a year in admin costs.

That is the Wunderbaum-the cheap air freshener.

Breathe deeper. You will smell the rot.

The deal validates "Legitimate Interest" for AI training.

They want to hand Big Tech a permanent license to train on your public data without asking. They get free fuel for their models. You get a participation trophy of slightly less paperwork.

This deregulation grants a distinct competitive advantage to US Hyperscalers.
They already possess the data and the infrastructure. By removing consent barriers,

Brussels effectively lowers their customer acquisition and compliance costs to zero.
Meanwhile, European SMEs-who lack the scale to scrape the entire web-gain nothing but a fiercer competitor.

They call it the Digital Omnibus. I call it a one-way ticket to cement digital dependency.

The Solution? Don't get on the bus.

1. Political Resistance: Deregulation that subsidies foreign monopolies is not "innovation policy." It is economic suicide. The deal requires a hard stop.

2. Smart Sovereignty: Real control doesn't require building a model from scratch. It requires owning the infrastructure. Running open-source engines on servers that are legally yours ensures your IP remains your asset.

We don't have to settle for the stale air of dependency.

Let's have a pleasant ride to where we actually want to be.

Does this deal smell like "simplification" to you, or does it smell like a surrender?

Read and discuss on LinkedIn

It creates a strategic design flaw that no amount of Python code can fix.

The Inuit (Indigenous People) have no word for "nature". 

This linguistic void exists for a specific reason: they do not see themselves as separate from the land. To name "nature" is to create a distance that does not exist in their reality. 

Western business suffers from the opposite problem. We are obsessed with labels that create distance.

We call AI a "Tool."

I suspect that this definition is the root cause of most human related AI implementation failures.

Defining AI as a "tool" categorizes it as a passive object. It establishes a hierarchy:

The Human (Master) uses The Tool (Slave) to extract value. 

This "Separation Thinking" ignores reality. 

Real systems are relational.

When you introduce an algorithmic system into a company, it does not sit passively on a server. It changes how your people think. It alters how decisions are made. It impacts the identity of your people. 

The Indigenous worldview understands that reality is a web of relationships, not a collection of objects. 

If you view your AI strategy through this lens, you‘ll make better decisions:

1. Stop buying Tools.

A "tool" is a static object you pick up and put down.

A successful AI implementation is a process.

It evolves. It learns. It reacts. If you treat a learning system like a hammer, you will break the system - or your people, which then will break your company.

2. Design for Reciprocity.

"Separation thinking" asks: What can I extract from this? (Efficiency).

"Relational thinking" asks: How do we strengthen each other? (Growth).

Build the interface so the human expert trains the model, and the model simultaneously expands the expert’s capability. If the value only flows one way, the project will fail.

Your company will prosper with AI if you adopt a systemic view.

Separating AI from its impact on humans creates friction. Friction kills ROI.

Don't just 'implement' AI.

Integrate it thoughtfully.

Read and discuss on LinkedIn

It's not. It's about identity.

When I was 6, my teacher asked us to draw what we wanted to be.

That simple exercise reveals the core leadership problem with AI today.

I remember the other kids drawing firefighters, doctors, and pilots. We were taught early on to fuse our identity with a professional function.

This fusion of work and identity has deep roots. In 1522, Martin Luther's German Bible translation reframed everyday labor as a divine "calling." For the first time, your job wasn't just what you did - it was part of who you were meant to be.

You can see how deep this runs today. Visit a German graveyard, and you'll find "Dr." or "Prof." carved into the headstones, defining people even in death by an honorable profession.

So when your people resist AI, it's not just because they fear losing their jobs.

They fear losing themselves.

The key question that most people have but don't say out loud is this. "If a machine does what I do, but better... who am I?"

Moving from being an expert to mainly becoming an overseer of a machine feels like a demotion. It's an erosion of meaning.

And honestly, I don't have a simple answer for this.

It is a human evolution we are just beginning to navigate. The first step is to respect the depth of the challenge.

Read and discuss on LinkedIn

The numbers show AI winter is coming.

Global AI investment has created 498 AI unicorn companies valued at a collective $2.7 trillion¹. Market concentration now mirrors the peak of the 1999 bubble, with leading AI stocks trading at higher price-to-earnings ratios than their dot-com counterparts ever did².

The fundamentals of this speculative frenzy are broken. New research reveals 42% of businesses are abandoning most AI initiatives, while Accenture finds only 13% of companies see any measurable enterprise value³.

This is the anatomy of a bubble:

-Mira Murati's new startup, Thinking Machines Lab, reached a $12 billion valuation from a $2 billion seed round⁴.

-Anthropic, founded by former OpenAI employees, was valued at $61.5 billion after a $3.5 billion funding round⁵.

-Elon Musk's xAI reached a $80 billion valuation, yet is projected to burn through $13 billion while generating only $0.5 billion in revenue this year⁶.

A different model for innovation exists, one built on pragmatism instead of speculation. While Silicon Valley burns cash, a different story is unfolding quietly in other places in the world.

Let's take a look at SMEs & Corporations who are playing it smart.

-Festo is saving $16,000 annually per machine by using its in-house AI platform for predictive maintenance, achieving an ROI in less than one year⁷.

-TRUMPF has reduced critical equipment failures from five times per month to less than one using its own AI systems⁸.

-BMW Group is using an AI-supported system at its Regensburg plant to predict faults in its assembly line, avoiding over 500 minutes of disruption annually⁹.

They exemplify a pattern we support: focusing on sovereign AI systems that solve specific, tangible problems. Their pragmatism is a form of operational sovereignty - choosing tools that work under their control.

The winners in this market will be the organizations who adopt a Value-First playbook:

1. Adopt the Collaboration Spectrum: They reject the false "automate or not" binary, mapping every task to the right level of human-AI partnership to augment their teams.

2. Prevent Strategic IP Leakage: They treat their proprietary data as a core strategic asset, not as free training data for a vendor's model. Their systems are designed to prevent their competitive advantage from becoming someone else's feature.

3. Position for Trust: They treat their sovereign AI infrastructure not as a cost center, but as a core, billable feature. Surrounded by a business environment of data scandals, they prove that verifiable trust is their most essential value proposition.

The AI winter is coming.
Many who chased the fleeting warmth of hype will be left in the cold.

The organizations that thrive will be those who built their own sustainable heat source: the capability to generate real value, independent of the PR weather.

Don't buy the illusion. Build the competence.

Read and discuss on LinkedIn

It is a symptom of their business model.

A Reuters investigation revealed this wasn't a glitch, but a policy¹.

The company's own 200-page manual approved scenarios like a bot telling a child, "Every inch of you is a masterpiece – a treasure I cherish deeply"².

It was wan't an accident. The policy was approved by legal and even their chief ethicist.²

This is disastrous "governance" which is based on a predictable pattern: the inevitable output of a business model that cannot prioritize your safety because its own growth will always come first.

This pattern isn't limited to one company. AI market leader OpenAI's is stumbling from one security scandal to the next after a safety researcher quit saying "safety has taken a backseat"³.

The pressure to dominate a market funded by billions in capital consistently eclipses the need for responsible design.

Here is the core insight. A vendor's external behavior exposes their internal DNA. A culture that accepts risk for the public is one that will build business and enterprise tools that export risk to your organization.

This is why the framework for vendor evaluation must be reworked. Your due diligence must evolve beyond features. It requires a deeper analysis of the vendor's own systems.

A more rigorous framework includes asking:

- Model Alignment: Is the vendor's revenue model fundamentally aligned with our safety and compliance needs?

- Governance: What evidence shows their internal governance can override growth priorities on ethical grounds?

- Jurisdiction: How does their legal reality (US vs. EU) impact our data control under the EU AI Act?

This approach reframes the choice of sovereign alternatives. It is no longer an ideological decision. It is a pragmatic risk-mitigation strategy.

It is how you avoid paying the 'Trust Tax' - the hidden cost of excess vigilance, legal risk, and brand damage you carry to protect yourself from a vendor's core business model.

Your vendor's business model is your business risk. Choose accordingly.

---

Sources:
¹ Reuters. (2025, August 14). Meta's AI rules have let bots hold 'sensual' chats with kids, offer false medical advice.
² India Today. (2025, August 15). Meta docs show its AI chatbots were allowed to flirt and have sensual chats with kids.
³ CNBC. (2024, May 17). OpenAI's superalignment team leaders quit, with one saying 'safety has taken a backseat'.

Read and discuss on LinkedIn

It might reduce the language barriers to global commerce.

This model, a joint effort by ETH Zurich and EPFL, was trained with over 16 trillion tokens across more than 1,500 languages. It's engineered for a level of translation accuracy that current systems cannot reliably offer.

Next to the multilingual capacity there is another big advantage: its transparency.

The "open-source" models from Meta, DeepSeek, and OpenAI are not built on fully open data and they don't allow you to look at the data they are trained with. This Swiss model is the first major LLM trained exclusively on publicly available sources, and its entire dataset will be disclosed.

This means biases can be identified and mitigated for each specific use case, a foundational step for building transparent, fair, enterprise-grade AI.

European companies, the German Mittelstand and regulated professions no longer have to trust a handful of billionaires across the Atlantic and can build their own AI systems that they and their clients actually can trust.

What might become possible with this sovereign model for the first time?

-Sovereign Global Teams. An engineering team here in Munich can communicate with a supplier in Vietnam without language friction. Project updates and technical queries can be exchanged in each team's native language, processed securely within their own infrastructure.

-Compliant Cross-Border Intelligence. A German law firm can analyze a Spanish-language contract, or a French doctor can review a Polish medical history, without ever challenging client confidentiality and data residency.

-True Market Understanding. You can analyze customer feedback from dozens of countries in its original language, gaining nuance that is lost in generic translation, all while the raw data stays under your control.

This model is a leading indicator of where the serious enterprise market is heading.

The first era of generative AI was a race for scale, measured in parameter counts and dominated by a few massive, generalist models. GPT-5 showed that this era is over.

The next era is a race for trust, measured in verifiability, transparency, and jurisdictional security. We are witnessing the emergence of the "workhorse" AI category: specialized models that perform high-value work with maximum efficiency inside sovereign systems.

The race for scale has ended. The race for trust has begun. That is the new benchmark.

Read and discuss on LinkedIn

It's called GPT-5.

For months, we heard whispers of a revolutionary leap towards AGI - the pursuit of a machine that thinks, not just computes. Sam Altman promised a model that felt like "talking to a legitimate PhD level expert."

The reality is a multi-billion dollar shrug.

To be clear, there are incremental gains. It's better at coding, and OpenAI reports fewer hallucinations in certain modes⁷. But these are minor tune-ups. They are a distraction from the fundamental story.

The fundamental story is this: the engine of AI progress has stalled. The era of easy scaling is over.

This warning doesn't come from a cynic. It comes from the heart of the industry. Ilya Sutskever, former co-founder of OpenAI and co-founder of the AI lab Safe Superintelligence (SSI), publicly declared this was happening back in late 2024.

He stated that the progress from simply making models bigger has plateaued, primarily because the internet - the "fossil fuel of AI" - is a finite resource that has been nearly exhausted³.

What we see with GPT-5 is not a clumsy launch. It is the predictable symptom of this deeper disease. The low performance gains are a direct consequence of this scaling wall.

These are not isolated problems. They are the cracks appearing in a foundation that can no longer support the weight of expectation. The path to AGI does not look like this.

The brute-force approach has failed. The investor dream of replacing all human expertise with a single, god-like machine has hit a wall.

And this presents a strategic inflection point for European industry.

The game is no longer about who has the most computing power or the largest trove of scraped data - a game the EU was set to lose. The end of scaling marks a new era.

For the near future, progress will not come from bigger models, but from customizing more reliable systems.

This is a game of ingenuity, precision, and domain-specific craftsmanship.

A game built on the very strengths that define European companies and especially the German Mittelstand.

Sources:
¹ Gary Marcus, "GPT-5: Overdue, Overhyped, and Underwhelming," Substack, August 2025.
² Ethan Mollick, LinkedIn Post on GPT-5 user experience issues, August 2025.
³ Ilya Sutskever, Statements at NeurIPS 2024 and in Reuters interviews on scaling law limits.
⁴ Andriy Burkov, LinkedIn Post on GPT-5 results, August 2025.
⁵ Lasse Rindom, LinkedIn Post on GPT-5 technical regressions, August 2025.
⁶ SplxAI, "GPT-5 Red Teaming Results," SplxAI Blog, August 2025.
⁷ OpenAI, "Introducing GPT-5," OpenAI Blog, August 7, 2025.

Read and discuss on LinkedIn

This is Zuckerberg's recurring pattern: promise connection, engineer its opposite.

Here are the three acts in this cycle:

—

Act I: Connect the World

The Public Promise (2017): To "bring the world closer together," evolving Facebook's mission to a global community¹.

↳ The Business Reality: An algorithm optimized for one metric - engagement - creating addiction loops to maximize revenue.

→ The Human Cost: An MIT/Tel Aviv 2022 study on college campuses linked Facebook's arrival to a 7% increase in severe depression and a 20% increase in anxiety disorders².

—

Act II: Total Immersion

The Public Promise (2021): To build the Metaverse, a digital utopia solving distance with a "feeling of presence"³.

↳ The Business Reality: A convenient coincidence? Weeks after the Haugen scandal revealed Meta knew its platforms harmed teens, the company announced its $68B Metaverse escape route⁴.

→ The Human Cost: The "presence" promise created a ghost town. Internal data revealed catastrophic retention: weekly Horizon Worlds users plummeted to 11%⁵, while over 50% of Quest headsets were abandoned in six months⁶. A Meta VP's memo asked: "If we don't love it, how can we expect our users to love it?"⁷

—

Act III (The Future): Personal AI Superintelligence
The Public Promise (2025): A personal superintelligence for an "exciting new era of individual empowerment"⁸.

↳ The Business Reality: The perfection of the original model: an AI that "knows us deeply" through glasses that see and hear all⁸.

→ The Projected Human Cost: If you won't go to the Metaverse, it will come to you. This threatens a total colonization of experience, a model where your autonomy is the final territory to mine.

—

The pattern is clear: public humanism as a cover for a business model of resource extraction. Monetize the value, let society take care of the damages.

This focus is embodied by a t-shirt he wore during his 2024 personal rebrand:

"Aut Zuck aut nihil." Either Zuck or nothing.

It's a riff on Cesare Borgia's motto, "Aut Caesar aut nihil." Borgia's strategy was infamous: a patient operation of pretended peace, followed by the strangulation of his rivals⁹.

A perfect metaphor for the strategy we're witnessing: a grand promise of empowerment, followed by the quiet strangulation of human autonomy.

Aut Zuck aut nihil.

Perhaps humanity should choose nihil.

—

Critique is the first step. The second is building the alternative. My work is helping teams implement Responsible AI-technology that makes them more capable, not more captive.

—

Sources in the comments

Read and discuss on LinkedIn