Takeaways 

• Despite USD 30–40 billion spent on enterprise GenAI, 95% of AI implementations produce no measurable return. The failure is rarely about model quality.

• AI value migrates upstream, the same way it did in cloud and mobile, toward whoever defines the defaults everyone else operates under, which is why AI sovereignty has become a board-level priority.

• Procurement defaults, pilot extensions, and infrastructure renewals are quietly setting the AI architecture every enterprise will live with for the next decade. The decisions get made in the boardroom without anyone recognizing them as architecture decisions at all.
  
  
  

Companies have invested  $30-40B in enterprise GenAI. Can you guess how many of them see real ROI? 5%. For a few years, the question surrounding AI was whether people were willing to adopt it or not, now, things have changed. As the stats prove, companies are already investing in AI. In fact, a recent survey shows that 88% of respondents said that their organization is using AI in at least one business function. So, if the question is not about implementation anymore, what is it? 

It’s sovereignty. The main barrier to scaling and seeing ROI is being able to have an AI model that can retain feedback, adapt to context, and improve over time. And you can’t do that if you don't govern your AI, if you’re fully dependent on a vendor. Let’s get inside this new question.

From “Can we implement AI?” to “What happens after AI adoption?”

Here’s a thing we all know: AI has evolved from being a tool that only a few groups use to a general-purpose capability built across different sectors and across different operations as well. We see marketing, operations, engineering, and any team using it. Despite this broad use, AI projects are stalling. A recent MIT report showed that only 5% of AI projects reached production with measurable P&L (Profit and Loss) impact. This is supported by another study from S&P, which stated that 42% of companies abandoned the majority of their AI initiatives in 2025, up from 17% the year before. 

If the use of AI is limited to a team member asking an LLM a quick question to speed up work, that’s not creating real, long-term business value. For example, that same generic LLM can’t be scaled by the company if the company needs to, it’s not tailored, it can’t be improved, it doesn’t retain context fully, it doesn’t adapt to workflows entirely, and, for sure, it can’t learn from feedback over time. It depends fully on an external vendor.

What seems to be a list against generic AI tools is actually proof of why we’re no longer discussing implementation or even the willingness to use AI: we’re talking about a major challenge, AI sovereignty, owning the AI our teams use.

The failure is not owning what governs your AI

AI sovereignty is an organization’s ability to govern its own AI capabilities, including the data it runs on, the models it depends on, the infrastructure that hosts those models, and the protocols that coordinate everything in between. A few years ago, the AI sovereignty question was happening mostly in government offices, framed around national strategic interests. But now it has moved into private sector boardrooms.


A McKinsey report showed that 71% of respondents consider AI sovereignty as an “existential” concern or “strategic imperative” to their organization's goals. Another report even said that 38% of enterprise leaders lack trust in AI security, while 33% fear vendor lock-in. Other McKinsey analysis also projects the sovereign AI market to reach $ 600 billion by 2030. The reason behind these numbers is that the implication of AI sovereignty for any company running AI is direct. 

As in every previous technology wave, value migrates upstream, toward whoever defines the defaults that everyone else operates under. Cloud, for example, rewarded the companies that built the infrastructure instead of those running workloads on it. The same happened with mobile, it rewarded the operating system owners, not the app developers competing inside their ecosystems.

And now we’re seeing the same. Any company operating fully dependent on infrastructure it doesn’t control, with orchestration logics designed elsewhere, governed by protocols set by another company, is sitting downstream of where value lives. Every output gets shaped by whoever is sitting upstream, and every workflow embeds constraints chosen by a vendor. This way, an organization stops running its own AI and starts renting access to someone else's, executed from inside its own walls. This is why AI sovereignty is essential.

Three decisions you’re already making about AI sovereignty 

Sometimes these kinds of decisions don’t look like decisions themselves. They happen across procurement defaults, pilot extensions, and infrastructure renewals. But they shape what an organization can do with AI for years. Here are the three most common: 

1. The infrastructure your AI runs on

Workflows, contracts, and team skills accumulate around whichever stack got chosen first, and the cost of migrating them later compounds quarter by quarter. If you don’t actively decide on this, forecasting what happens next, you might get locked into your vendor. 

2. The orchestration layer that coordinates your AI.

This is the layer that connects AI models to data, agents, and existing systems. Long story short: that's where the operational logic actually lives, and it gets locked in fast. Choosing how AI is orchestrated today shapes how every future agent inside the organization coordinates and escalates. This layer must be yours to own.

3. Building internal capability.

Full independence isn’t necessary, but full dependency is something to avoid. Every quarter you spend running on someone else's AI without building your own internal AI capability is a renewal to live under outsiders’ decisions. The internal AI capability is crafted across different actions: owning the intellectual property, governing your AI, embracing your own methodologies, and implementing change management. 

None of these three decisions shows up as a moment that demands a vote in the boardroom. They get taken silently in renewal cycles, for example. The AI infrastructure your company will live with for the next decade is being set right now.

The implementation era of AI is closing. The AI sovereignty era is opening, and the companies that come out ahead will be the ones that recognize the shift in time to act on it. Where are you going to stand?

FAQs about AI sovereignty

Why is AI implementation no longer the central problem?

AI implementation has been answered at scale. 88% of organizations now use AI in at least one business function. The unresolved problem is what happens after adoption. MIT Project NANDA's The GenAI Divide found that 95% of enterprise AI pilots produce no measurable return despite USD 30–40 billion invested, and the cause is structural rather than technical. Most organizations are running AI on infrastructure, models, and protocols they do not control. The bottleneck is now control.

What does "owning the AI operating layer" mean?

Owning the AI operating layer means controlling the coordination infrastructure that governs how AI models, agents, data, and protocols interact across the organization. It goes beyond using AI tools. Structural advantage in the AI economy concentrates at this layer rather than at any single model or application. A company that owns its operating layer sets its own defaults: which models it uses, how they orchestrate, and how its data flows. A company that doesn’t own this layer inherits defaults set by its vendors. At Arionkoder, this principle sits at the core of how we work, keeping that operating layer in the client's hands.

What AI decisions are companies making right now?

Three decisions taken between 2025 and 2026 will constrain enterprise AI options for years: the infrastructure your AI runs on, the orchestration layer that coordinates it, and whether to build internal AI capability or stay vendor-dependent. These decisions rarely arrive with formal recognition. They get made through procurement defaults, pilot extensions, and infrastructure renewals. Once workflows, contracts, and team skills accumulate around a choice, the switching cost makes reversal impractical. The architecture set today is the one most companies will live with for the next decade.

How can a company know if it's positioned upstream or downstream of AI power migration?

The diagnostic is operational. When a competitor deploys a customer-facing AI solution at scale, can the organization respond within 18 months, or do the responses depend on infrastructure, model, and orchestration decisions someone else already made? Companies positioned upstream control the defaults that shape their AI outputs. Companies positioned downstream operate within constraints chosen by vendors. The simplest way to answer the question is to audit how many layers of the AI stack (models, orchestration, data) are under direct organizational control.