Much of 2025’s AI news chased the latest generative model releases. As the foundation model market matures, that hype will matter less in 2026, but will there still be noticeable shifts?
Absolutely. Independents like OpenAI are scrambling to diversify, turning assets into hard tangibles such as data center bricks and chips, while giants like Google leverage their vertical stacks, from bare metal to massive interaction channels, backed by adjacent businesses that fund it all, challenging NVIDIA’s dominance as well.
Meanwhile, clever innovators could still rewrite the scaling laws that define the link between model quality, compute, and data. And Chinese models such as DeepSeek and Qwen are gaining momentum in the open-source space. Yet for now, the foundation model market is trending towards a broader range of models that are better, cheaper, and faster, and is clearly on a path to further commoditization.
AI Regs Not Going Away in 2026
The political mood has been how to harness AI for economic growth, leading to calls for less AI regulation.
Some have foreseen major rollbacks on the most influential set of AI regulations, the EU AI Act. When the European Commission proposed amendments under the ‘Digital Omnibus on AI Regulation Proposal’, there was speculation that 2026 would see AI regulations watered down or even shelved.
Reality check: Organisations operating in the EU still need to be prepared for the AI Act to guide the development and deployment of AI.
Simplification is a positive goal, but the Act’s core principles, such as its risk-based approach with proportionate requirements for prohibited, high, limited, and minimal-risk systems, remain intact.
The proposal mainly delays certain deadlines rather than removing them, extends exemptions from SMEs to small mid-cap enterprises, centralizes oversight of general-purpose AI models within the AI Office, and shifts AI literacy from a requirement to a strategic recommendation for limited and minimal-risk systems.
While these measures are being negotiated, there’s a risk of unintended consequences if timelines remain unchanged and the Omnibus proposal isn’t adopted swiftly. In that case, existing legislation will take effect. To minimize uncertainty, the EU must act quickly, or organizations will need to play it safe and plan according to current rules and deadlines.
On the GenAI ROI Debate: Hold Your Nerve
The second half of 2025 saw a backlash against claims of AI’s impact. Studies and commentators reported that enterprises were seeing little, if any, ROI from generative AI implementations.
So, as we enter 2026, we push through a veil of disappointment, if not disenchantment, with AI. This is a natural reaction to the pumped-up hype and noise. Loud, exaggerated claims won’t accelerate adoption in 2026.
Yet 2026 will be a transition year. The AI projects that matter are those delivering transformative results, major enterprise programs where AI is woven into processes to improve service outcomes and operational efficiency. These initiatives don’t wrap up in a few months; they require 12–18-month timelines for full-scale rollout and adoption.
They’re not about quick fixes promoted by some vendors but about ensuring foundational work around data, processes, and change management is thoroughly done and dusted.
Agentic AI Rolls On and On
So where can we find the real money and action in 2026? It is in the application layer, the layer of downstream tasks where this AI is applied and where the value is created – or lost. This is where AI gets embedded into the often-unglamorous reality of business strategy, enterprise architecture, and workflows. Enter the ‘advent’ of agentic AI.
But simply unleashing herds of agents won’t solve your problems. At best, they will be ineffective; at worst, they will turn into an angry mob. Did you know that in leading agentic benchmarks, even the best agents resolve less than half of the tasks? Yes, these benchmarks are intentionally hard, with ambiguous goals and underspecified tasks, but isn’t that the grim reality of any enterprise?
The real successful agentic use cases in 2026 will need to be robust, predictable and reliable. They’ll focus on design-time uses, such as the ideation or design of applications, with business problem owners in the loop. This applies to new applications, but the real opportunity lies in transforming legacy systems with AI, freeing up cycles for real innovation.
For run-time cases, success will come from combining the creativity of agents with the predictability and repeatability of reliable data and agentic tools, such as workflows and business rules. These agents must operate in the confinement of a case, where the required data, context and state is provided, but only within allowed access and actions.
Automated AI ‘eval’ benchmarks will be used to add more science into agent development, enabling agents to recover and adapt in real-time, and continuously monitor agentic systems.
Finally, real successful agentic use cases will require transparency and explainability for different audiences, from the top down to the most granular level.
What is keeping AI From Becoming Autonomous Intelligence?
As we look beyond 2026, the real challenge isn’t just scaling models or embedding them into workflows; it’s moving toward AI as true autonomous intelligence. This requires breakthroughs that go far beyond today’s large language model-driven systems.
Early cybernetic systems, even in the 1950s, were situated forms of intelligence: embedded in environments where they could gather their own data, make decisions, act on them, and, most importantly, react to feedback and learn from it.
By contrast, today’s large language or text-to-video models, despite their impressive capabilities, are passive couch potatoes by those standards. They are spoon-fed massive amounts of curated training data, disconnected from the real-world environments where they need to operate, and lose their ability to learn once deployed.
Agentic systems are a first step toward greater autonomy, but they must actively acquire the right data, perform goal and intrinsic motivation-based reasoning, form long-term memories, and continuously learn from feedback to adapt to dynamic conditions. They also need to interact with humans, agents, and other systems. This requires understanding other “minds”, beliefs and intentions, developing abilities to coordinate, negotiate, argue, balance interests, and build reputation and trust. And they must do all of this safely and reliably, with ethics, norms, governance, and transparency built in.
The next frontier is building AI that is adaptive, self-improving, socially aware, and capable of negotiating complex environments—while remaining accountable and aligned with the needs and values of humans, other agents, society and the business.
