AI everywhere, results not yet
At IPEM Global 2026, AI came up in every conversation. Ask for the results, though, and the room got a lot quieter.
One panel put the gap into words: capital is pouring into AI, but operational impact and value creation haven’t caught up. Almost every investment story now includes AI. Very few can point to a line in the P&L that it has changed.
The research says the same. MIT’s The GenAI Divide study, published in 2025, estimates that about 95% of generative AI pilots have no measurable effect on profit and loss. Its authors, who reviewed more than 300 public deployments, don’t blame the models. They point to tools that don’t know the business they operate in, don’t learn from use, and sit outside everyday workflows. They also found shadow AI almost everywhere: in nine out of ten companies, staff use personal AI tools daily, while fewer than half of those companies have bought official licenses.
For investors, that is a warning worth taking seriously. The technology works. What most organizations lack is the foundation to make it pay.
Why most AI pilots stall
Picture a common scenario. A company launches an AI assistant with plenty of fanfare but never gives it access to internal knowledge. It writes well, yet knows nothing about the company’s own products, policies or clients. After a few weeks of vague answers, people go back to the old way of working. The pilot is declared a technical success and quietly shelved.
Across regulated industries, we see the same handful of blockers: concern about client data and intellectual property leaving the building, the combined weight of GDPR, the EU AI Act and DORA, unclear access rights that invite shadow AI, knowledge spread across disconnected systems, swings between hype and skepticism, and token-based pricing that makes the business case hard to defend.
A better model fixes none of these. What fixes them is a clear decision about where AI runs, what it is allowed to see, and who keeps control.
The question we brought back
That is the question we kept returning to after IPEM: as AI becomes part of how decisions get made, where will it actually run, and who will own the infrastructure behind it?
This is where sovereign AI comes in. It isn’t about building your own foundation models or turning away from global innovation. It is about keeping control in four areas:
(1) Data: sensitive information stays protected and is used only within the organization’s own rules and regulatory obligations, whichever model touches it.
(2) Knowledge: contracts, policies, analyses and the know-how of employees are often worth more than any database. AI makes this knowledge usable at scale, which also makes it easy to lose if it flows into external tools unchecked.
(3) Decisions: every AI-supported outcome should be explainable and traceable, so someone can verify it and take responsibility for it.
(4) Operations: the organization, not the vendor, decides where each AI workload runs, based on risk, compliance and business value.
Why we built KVARK
KVARK, Egzakta’s sovereign agentic AI platform, grew out of exactly these questions. It is built for organizations where data sensitivity and regulation leave little room for compromise.
Early on, we found that standard servers couldn’t run demanding models efficiently inside a client’s own environment. The answer was high-density, liquid-cooled GPU infrastructure, paired with the software layer on top. The result is a single system:
100% on-premise, with zero data leakage. Everything is processed within the client’s internal domain, on liquid-cooled hardware behind enterprise firewalls. KVARK is air-gapped by design and can run in fully isolated environments.
Designed for GDPR and EU AI Act compliance. Auditability, transparency and access control are part of the platform, not add-ons. KVARK follows existing permissions, so the AI never sees more than the person it works for, and the same governance rules apply to AI agents.
Predictable total cost of ownership. Fixed licensing can reduce TCO by up to 50% over five years, and liquid cooling delivers cloud-class performance without cloud bills or per-token surprises.
Hardware and software delivered together. The platform arrives pre-integrated and tuned, ready for productive work from the first day.
Our deployments have taught us something that has little to do with GPUs. The slowest part of any rollout is getting processes, permissions and data in order. Clean data and well-defined processes and access rights decide how quickly AI starts delivering value, far more than the choice of model.
That foundation matters even more as assistants give way to AI agents that complete multi-step tasks on their own. AI agents need clear limits, human sign-off on decisions that matter, and a complete record of every action. We designed KVARK with that shift in mind.
Europe’s opportunity
When generative AI went mainstream in late 2022, it seemed its future would be decided by a few large technology companies. For European organizations, that raised uncomfortable questions. How much of your core business should rely on technology you don’t control? What happens if a provider changes its prices or terms, or gets caught up in geopolitics?
Europe’s answer is taking shape through regulation such as the AI Act and DORA, and through major public investment in AI computing capacity. The message is clear: AI is now treated as strategic infrastructure. For companies across the region, including mid-sized banks and enterprises, sovereignty doesn’t require a massive build-out. A few modern GPU servers can keep the most sensitive workloads in-house, while shared European capacity handles large-scale training and peak demand.
Models can be bought. Capability can’t.
If 95% of pilots fail, the more interesting question is what the other 5% do differently. According to MIT, they embed AI in real workflows, connect it to their own data, and let it improve in their own context.
Models are quickly becoming a commodity. Better ones appear every few months, and open-weight models now rival proprietary ones for many business tasks. What can’t be downloaded is an organization’s ability to put AI to work on its own knowledge, under its own control.
The winners of the next AI investment cycle will be the ones who own their infrastructure, keep control of their data, and can show real results. Europe has the opportunity to build that capability on its own terms, and at Egzakta, we intend to be part of it.
