In 2018 we started an AI project at REALONAUT, long before every managing director was hearing the word AI in every other meeting. The idea: software would recognise from photos what condition a flat was in. The project was funded by the FFG, and our own software supplied the images.
Technically it worked. The neural network learnt what it was meant to learn. Even so, we did not continue it in 2020.
The problem was not the algorithms
An AI is only as good as the examples it learns from. For it to assess the condition of a flat reliably, it needs a great many images that an expert has rated beforehand. We would have had to pay for those ratings, image by image, from people who really know about flats. That was not something we could finance.
To this day, that is the most common reason AI projects fail: not the technology, but the data and what it costs to make it usable.
Three questions before any AI project
Before you put money into an AI project, you should be able to answer these questions:
- Do we have the data? Not just any data, but enough of it, in good quality and with permission to use it.
- Who tells the AI what is right? Someone has to rate examples and later check whether the results are correct. Who is that, and what does their time cost?
- What happens when the AI gets it wrong? A wrong suggestion in an email is annoying. A wrong decision about money, people or safety is expensive.
If your provider does not raise these questions themselves in the first conversation, you should be on your guard. An impressive demo answers none of them.
What I took away from it
I consider AI one of the most important tools companies are getting their hands on right now. But today I look first at the data and the people behind it, and only then at the technology. That order would have saved us a lot of time back then. It can save you a lot of money.