Neal Riley
Published on 5 October 2026
From models to systems: what changed in AI over the past year?
AI is no longer just answering questions; it’s starting to do the work. Explore what this shift means for business leaders.
A year ago at OxGen25, we were asking what AI agents might become. A year later, we have started to see what happens when they actually work. Models have become more capable, but more importantly, what they can do is considerably more usable.
AI systems can increasingly work with tools, inspect their own output, recover from mistakes and continue towards a goal. In other words, we are moving beyond AI that simply answers questions towards AI that can help complete work.
For business leaders, that changes the conversation. The question is no longer just: which model should we use? It is increasingly: what work should we delegate, what evidence should we require, and what must remain under human control?
Capability isn't the same as adoption
One of the lessons of the past year is that AI capability and organisational adoption move at very different speeds. A new capability can become available almost overnight. Turning that capability into something an organisation can use consistently and responsibly takes considerably longer.
Businesses still have people, processes, budgets, procurement cycles, risk controls, customers and legacy systems. AI may move at the speed of software, but organisations don't. That gap between technical capability and organisational readiness is where much of the real work now sits.
Organisations benefit most from AI by understanding where it genuinely changes how work gets done and building the right environment around it, rather than simply adopting every new model first.
From chatbots to agents
The rise of AI agents is an important part of this shift. A conventional chatbot responds to a request. An agent can take a task, use tools, act on what it finds, check the result, and decide what to do next. It is a different relationship between people and computers. But greater autonomy also means greater responsibility.
When an AI system can take action, organisations need to be much clearer about what it is allowed to do, what success looks like, what evidence it needs to provide and when a human needs to step in. These aren't futuristic questions. They're management questions.
What should organisations do now?
The move from chatbot to agent shouldn't be understood simply as “more capability means more productivity”. A better way to think about it is: more capability means more possible work, more possible failure, and a greater need for structure.
There are four questions business leaders should be asking.
1. What work is suitable for delegation?
Start with work that has clear inputs, outputs and acceptance criteria. The more clearly success can be defined, the easier it is to understand where an AI system can take responsibility. That doesn't mean every suitable task should be automated. It means being deliberate about where delegation creates value and where human judgement remains essential.
2. What permissions should the system have?
An agent shouldn't be given access simply because it is technically possible. Leaders need to consider what an AI system may read, change, send, purchase, or publish, and what it should never be able to do without human approval. As systems become more capable, permissions become part of the business design, not just an IT or security issue.
3. What evidence is required?
If an AI system produces an analysis, recommendation or document, what do you need to know before acting on it?
Organisations should define what sources, checks or supporting evidence an AI-generated output needs to provide. The more consequential the decision, the stronger that requirement should be, which is particularly important as AI-generated material becomes more abundant. When content is cheap to produce, knowing what to trust becomes more valuable.
4. Where must responsibility remain human?
Delegation does not remove accountability. It changes where judgement is applied.
If generating code becomes easier, architecture and security matter more. If content becomes abundant, judgement and credibility become scarcer. And if agents can perform more tasks, deciding which tasks they should perform becomes a core management skill.
The model is only part of the system
One of the biggest shifts of the past year is that the model is no longer the whole story. What increasingly matters is the system built around it. A model, combined with the right tools, data, environment, verification, and human expertise, can be far more useful than one considered in isolation, which is why comparing AI purely on benchmark scores can be misleading. The best model isn't necessarily the one with the highest score. It is the right fit for the work you need to do.
For leaders, that means thinking beyond AI procurement. It means thinking about workflows, permissions, governance, skills and accountability. It also means recognising that not every organisation needs to build everything itself. The question is where genuine capability and control are essential, and where reliable access to someone else's capability is sufficient.
More capability means more judgement
Perhaps the most important implication is that AI doesn't remove the need for human judgement. It changes where that judgement is needed.
The next stage of AI won't be defined by model capability alone. It will be defined by how well organisations turn that capability into systems they can trust, govern and use. The more useful questions for business leaders are:
- Where should AI change things?
- How quickly can we adapt?
- And where must human judgement remain firmly in control?
Those questions will shape the next phase of AI adoption—and the organisations that ask them early will be better placed to make AI genuinely useful, rather than simply available.