From model to system: Three AI developments to act on now

Behind all this jargon lies a fairly straightforward transformation. Organisations are no longer focused solely on models, but are shifting their attention to the broader system around them, data, workflows, governance, people, and decision-making.

This is an international shift that is becoming increasingly visible in the Netherlands. Conversations in boardrooms are no longer about whether to ‘do something with AI’, but rather about how to implement AI so that it becomes as fundamental to business operations as an ERP system.

1. AI as a core process

Over the past two years, many Dutch organisations have followed a similar path. Microsoft Copilot licences have been rolled out, internal chatbots implemented, and training programmes launched to help employees get started. This delivers an initial, tangible gain, many professionals save around 15 to 30 minutes per day on repetitive, routine tasks.

However, individual productivity turns out to be a poor measure of business AI success. The effects are difficult to quantify and often seem modest at first glance. The real value emerges when AI is embedded structurally in the processes that influence revenue, cost, and risk. At that point, the focus shifts from questions like “how many hours do we save per day?” to questions such as:

• What proportion of our cost base and revenue is directly or indirectly affected by AI today?
• Which decisions are we demonstrably making faster or better with AI support?
• Which outdated ways of working have we actually eliminated?

Operating at this level immediately makes an AI roadmap more strategic. In practice, many roadmaps still resemble a long list with dozens of isolated use cases per team, each with its own technical requirements and governance arrangements. Stacking up pilot projects mainly just adds complexity.

To gain control, organisations need a scalable, integrated roadmap with clear, organisation-wide criteria for data, risk, technology, and adoption.

Early signs of this approach can already be seen in larger Dutch organisations. For example, a team experiments with Copilot and translates those experiences into a reusable blueprint for other workflows and departments. Or a bank maps all high-volume work (from customer service to fraud prevention) and uses that insight to systematically drive transformation and improve risk management.

Such approaches make it possible to connect numerous single point solutions into a single end-to-end value chain.

2. From data perfection to ‘good enough’

For a long time, it was assumed that true, transformative business value would only be achievable once all company data had been brought together into a single, perfectly structured layer. In reality, data has only become more diverse and messy, structured databases now coexist with policy documents, emails, PDFs, videos, and countless other forms of unstructured information.

Generative AI thrives precisely on this mix of structured and unstructured data. Attempting to structure everything perfectly is therefore not only unrealistic, but often unnecessary. Fully cleansing and perfecting data quality frequently costs more than it will ever deliver.

In many processes, “fit-for-purpose” data is sufficient it is not flawless, but reliable and rich enough for the intended use.

Of course, there are important exceptions. In heavily regulated domains, data quality remains critical. A chatbot assessing complex financing applications against policy cannot afford inaccuracies – ‘good enough’ would be a risk. But in other steps in the same value chain, less than perfect data can still be used effectively.

It therefore makes more sense to prioritise process rather than starting with data. Start small, look at which three to five workflows are so important that they can be designed around the data that is already available?

Instead of a single, all-encompassing central data programme, this leads to a focused approach for each core workflow.

Risk and compliance play a key role here. They help determine where, for example, 80% accuracy combined with a human check already delivers significant value. Agents require a higher threshold, they should only be deployed when a false positive rate of below 1% is achievable.

This creates a more nuanced balance between speed, value, and risk.

3. AI adoption as both lever and risk in the digital divide

Process thinking is the foundation of any AI transformation, but success ultimately depends on adoption. No process design will hold in practice if employees lack the skills to work with the associated tools.

However, the rise of AI is widening the existing digital divide. Colleagues who are already digitally skilled adopt AI applications more quickly, experiment more, and achieve results faster. Those who struggled with earlier waves of digitalisation fall further behind.

As a result, AI adoption is not just a technology issue, but also a HR and diversity challenge: who can keep up with the new way of working and who cannot?

Organisations that want to stay ahead therefore combine several elements. They define their strategic AI ambitions clearly. They translate those ambitions pragmatically into processes, working concretely from key value streams. And they invest at least as much in adoption, upskilling, and change capability as in new technology.

Not only to realise more value from AI, but also to prevent the digital divide within the organisation from widening further.

This article was first published in BlogIT in Dutch.