Nearly every organisation is considering doing something with AI by now, but the step from “we should probably do something with this” to a concrete, working application remains difficult for a lot of businesses. This article outlines a realistic approach, based on projects I’ve guided myself for organisations of various sizes.
Don’t start with the technology, start with the problem
The most common pitfall is that an AI project starts with a tool rather than a question. Someone read about ChatGPT or a specific application, and the search begins with “what can we do with this” instead of “where are we currently losing time or quality.” The latter is the better starting point.
In practice, this means: first map out which recurring tasks take a lot of time, are error-prone, or are simply tedious enough that nobody enjoys doing them. That’s often a better candidate for AI than an impressive but rarely used application.
Five steps to a working implementation
- Map out the bottlenecks. Talk to the people actually doing the work. They know exactly where time is being lost, better than any management overview could tell you.
- Pick one concrete starting point. Not five applications at once, but one well-defined process where results become visible quickly. Success on a small scale builds trust for the next step.
- Sort out data and privacy upfront, not afterwards. What information is a system allowed to use, and what isn’t? This isn’t a side issue: it determines whether an application can be deployed safely and responsibly.
- Build and test with real users. An AI application that only works in a demo isn’t an implementation yet. Let the people who will use it daily look at it early and give feedback.
- Aim for handover, not dependency. A good implementation ends with the team understanding how the system works and being able to adjust it themselves, not with a black box only an external party understands.
Common mistakes
A few pitfalls that significantly increase the risk of a failed implementation:
- Starting too broad, so nobody feels clear ownership of the outcome
- No attention to changing how people actually work: buying a tool alone changes nothing if nobody uses it
- Assuming AI is always right, without a human check on important output
- Only considering privacy and data protection after the system is already in use
Need help with the first step?
I help organisations translate AI into concrete, working applications that fit how they actually work, with attention to privacy and without creating lasting dependency. Read more about AI consulting, or get in touch for a no-obligation conversation about where AI could make the most difference in your organisation.
