From models to agents
In a rapidly shifting digital landscape, artificial intelligence continues to redefine industries, business models, and even research methodologies. In recent years, there has been a marked shift from traditional large language models (LLMs) to what’s now being called agentic AI. Unlike the static, prompt-response interaction of earlier LLMs, agentic systems are given tasks and determine how to execute them autonomously. This evolution enables AI to break tasks into sub-components, select tools, evaluate results, and self-correct, representing a new phase in AI autonomy.
For example, the internal AI assistant developed by Prosus, named Toqan, saw hallucination rates drop from 2.5% to 1% after introducing agentic logic. That reduction is not just statistically impressive, it’s functionally transformative. When AI begins to approach or surpass human-level accuracy in complex tasks, entire workflows can be reimagined. In 2024 alone, Prosus saved over 1.7 million working hours – equivalent to 1,000 full-time employees – by using AI assistants. This doesn’t translate to layoffs, but rather ‘invisible hires’ that free up staff for higher-value work.
The path to effective AI adoption
Yet, these gains are subtle. They don’t occur in grand shifts but in micro-bursts of daily productivity – quicker document summaries, improved code review, better research insights. It’s a new paradigm where productivity is fragmented but cumulatively powerful.
Despite the successes, the journey wasn’t linear. A critical misstep early on was the assumption that users would naturally adopt AI tools without guidance. In reality, adoption requires active engagement: training, incentives, and workflow integration. It takes approximately six months for users to move from basic usage to full proficiency. The key takeaway? Organisations must invest in education and support to enable AI-driven change. The best technology in the world will fall short if users don’t know how – or why – to use it.
Infrastructure versus application
While infrastructure companies like Nvidia have captured the lion’s share of AI market capitalization, attention is now shifting toward the application layer. Just as the early internet era was dominated by Cisco and other infrastructure giants before being overtaken by software and services, AI may be poised for a similar inversion. Enterprise adoption data backs this up. In 2025, a sharp uptick in enterprise subscriptions to tools like ChatGPT – which could be the ‘Uber of AI’ signals a broader shift. Consumer behaviour may be poised to change on a global scale.
The Chinese model developer DeepSeek challenged assumptions about the cost and scale required to produce competitive AI. With comparatively low budgets, DeepSeek released models whose performance was within 3-5% of industry leaders like OpenAI and Anthropic. While not necessarily superior, the implication is profound: top-tier AI capabilities may no longer be reserved only for the well-capitalised elite. Open-source communities are also narrowing the performance gap. For many use cases, these models are already good enough, creating opportunities for innovation outside traditional powerhouses.
AI and the future of work
Y Combinator’s recent call for full-stack legal startups – companies that don’t just sell AI to law firms but reinvent the law firm itself – is a clear example of AI’s disruptive potential; if traditional players don’t adapt quickly, they risk being replaced rather than enhanced. This ethos underpins the urgency of broad AI adoption. Organisations and individuals who opt out risk stagnation. Every job is changing. The only question is whether individuals will shape that change or be shaped by it.
The opportunity lies in measured urgency through embracing AI while maintaining clear-eyed judgment about its limitations. The winners will not be the fastest adopters, but the most intentional ones – those who know when to fail, when to double down, and how to bring their organisations along with them.
