AI

AI has reached a strange moment.

On one hand, it feels like it’s everywhere. New tools, new agents, new promises of autonomy appear almost daily. On the other, many organisations still struggle to turn AI investment into sustained, scalable value. The problem isn’t a lack of technology. It’s a misunderstanding of what enterprise AI is, and what it needs to succeed.

One of the biggest misconceptions is the tendency to equate AI with consumer-facing tools like ChatGPT. That first wave of generative AI was horizontal by design. That is, it was accessible to everyone but is shallow in its understanding of real business complexity. Useful and impressive? Yes – but fundamentally limited.

Enterprise AI is a different story. Real value comes from vertical AI such as agents designed to perform specific activities within real business contexts, using defined skills, tools and data. This is where businesses get real return on investment. Not through generic assistance, but through replacing or changing concrete parts of how work is done.

But this transformation won’t happen overnight. It will be gradual, layered and deeply iterative. We are already moving away from the idea of “human in the loop” to “human above the loop”. This basically means that people will supervise, orchestrate and control multiple agents rather than interacting with one at a time. This shift alone will have profound implications for how organisations think about work, accountability and leadership.

The End of Prompt Engineering?

A year ago, prompt engineering was labelled “the job of the future”. In reality, it barely had time to exist.

Today, most high-quality prompts are generated, refined and optimised by AI itself. In enterprise solutions, prompts are rarely visible to end users at all. Instead, they’re embedded inside applications, multiplied across agents, and continuously improved behind the scenes.

What replaces prompt engineering is something potentially far more durable: context engineering. The real challenge is no longer how to phrase a question, but how to assemble the right context — data, documents, tools and constraints — so models can reason accurately and stay grounded. This requires a deep understanding of systems, data quality and business logic. It is one of the reasons why the role of AI engineer is still relevant, while narrower titles are fading quickly.

Where AI Investments Quietly Fail

Despite all the headlines about AI bubbles and infrastructure costs, the biggest source of wasted investment lies in something far more mundane: fragmentation.

Many organisations look to direct AI in too many directions at once: different technologies, different vendors, different methods, all driven by isolated initiatives. This “heterogeneous box” approach is useful for learning early on, but only briefly. Left unchecked, it produces solutions that don’t scale, can’t be maintained and fail to reuse data or capabilities.

At some point, difficult choices must be made. Organisations need a coherent AI foundation that addresses data quality, scalability, security, sovereignty and trust. Two or three proof-of-concepts can teach you a lot. Ten creates too much complexity.

Autonomous Agents Need Structure, Not Freedom

Agents are undoubtedly the dominant trend in AI today. But the industry’s obsession with “autonomy” is misleading.

In practice, trust is the limiting factor, because every agent can – and likely will – make mistakes. When multiple agents interact freely, errors start to compound. Without a robust structure, hallucinations multiply and confidence collapses.

The solution isn’t more autonomy, but better structure. Workflows, business rules and formalised processes provide essential guardrails, but they must be reinforced by continuous monitoring, trust metrics, and human oversight above the loop. A realistic balance today looks closer to 80–90% structured workflow with 10–20% autonomy. This mix delivers productivity gains without sacrificing control or trust.

Fully autonomous agent swarms may seem like a tidy marketing narrative, but they are not a reliable enterprise reality yet.

Start with People, Not Processes

If there is one place to begin, it’s not deep process automation. It’s the employee experience.

An Employee 360 assistant that spans HR, IT service desk, process guidance and ticket reduction delivers immediate, measurable value. More importantly, it normalises AI in everyday work. Such familiarity builds trust, and trust facilitates deeper transformation later.

Trying to automate core processes before employees have lived with, and understand, AI almost always leads to resistance. But people don’t resist technology – they resist uncertainty.

The Real Risk: Trust

Trust is the hardest problem in enterprise AI, and the most uncomfortable to confront.

Traditional applications can be tested against known scenarios. AI systems interact through natural language, across unpredictable questions and sensitive domains. Ensuring outputs are safe, compliant and appropriate requires multiple layers of validation. This includes testing AI with AI.

Ignoring this challenge doesn’t remove the risk. It simply delays the moment when it becomes visible.

What Could the Next Three Years Bring?

Agent engineering will almost certainly become more standardised, with shared registries for tools, data and capabilities. Trust will become measurable, driven in part by regulatory pressure from global governments (such as the EU Artificial Intelligence Act). Application development will accelerate dramatically through natural-language-driven approaches, even if code becomes more disposable.

Organisational culture will experience a much slower change. Technology will continue to move faster than people, governance and risk appetite – and that gap will be the real constraint.

 No Longer Just a Tool

Describing AI as “just a tool” may help many to rest easier at night, but it doesn’t reflect reality. In many roles, AI is more like a partner. Agents are being managed like team members and leadership models are already adapting.

The most important question, then, is not which AI solution to deploy, it is: how AI can reshape services, operating models and markets.

Only once that ambition is clear does the technology conversation truly make sense.

 

AI is moving from process optimisation to deep business reinvention. But it won’t change everything at once, with bottlenecks such as trust, governance and speed of organisational change. Winning will require a strong mix of vision, agility and managerial commitment.