From AI tool to a new workflow

Many companies understandably start with the tool. They test ChatGPT, Copilot or a specialist solution and then look for possible use cases. I recommend doing it the other way round:

Take a real process that currently consumes time, involves several people or repeatedly leads to questions and handovers. Then assess where AI could genuinely improve it.

It is also worth defining KPIs from the outset (measurable indicators used to assess success). These might include time spent, quality, specialist effort, error rates or lead times. Only then can you later judge whether the new process is actually better.

An AI tool can be tested quickly, but the underlying process does not necessarily change with it. Business processes involve different roles, hierarchies and often several systems. This creates dependencies, handovers and clear responsibilities.

An AI tool therefore does not automatically improve the workflow. Real process improvement only happens once it is clear where AI should intervene, what data it needs and how it fits into the existing process.

Reviewing the existing process raises a number of questions: Who does what today? Where do repeated questions arise, and how much time do they take? Where is knowledge repeatedly gathered from scratch? What genuinely requires specialist expertise, and what is mainly repetitive work?

Only then does it become interesting to decide what AI can take over or prepare. This might include research, a first draft, structuring existing information or comparing different documents and data sets. It is worth looking closely at where AI adds significant value and where it does not. Deciding deliberately not to automate a step is also part of the learning process. At the moment, AI capabilities are evolving faster than many organisations can test them.

“Human in the loop” has almost become a mandatory phrase. But the underlying questions still matter: Where exactly do we need people in the process? Where do we still need human judgement and decision-making?

The key is to distinguish routine checks from decisions that genuinely require expertise. Does a specialist really need to review every intermediate step, or is it enough to step in at a critical point? Where might sampling be sufficient?

That depends heavily on the process, regulatory requirements and governance. With sensitive data, questions around privacy, information security and data residency also come into play. Where do we need expertise and experience, and where can a system work independently in advance? Context and accountability are decisive.

Instead of planning a large-scale AI rollout straight away, I recommend starting with a manageable pilot. Subject-matter experts need time to experiment, make corrections and understand why something works — or why it does not.

This does not just create a new AI-supported process. It also builds practical knowledge within the team around the technology itself.

During the pilot, the previously defined KPIs are reviewed and documented. Only this comparison shows whether an interesting AI experiment has actually resulted in a better way of working.

As an experienced IT project manager, I connect business and IT. I help set up the project, structure requirements and support implementation across development, the Product Owner and your team.