Lately I have spoken to several founders who have told me that people have been asking them to do things with AI for their products/services. Clearly there is a hype, and a lot of attention is being focused around AI. In addition, there seems to be a wave of powerful investments in startups with AI in their name, which feeds the hype.
If I can give my two cents on the subject, based on my experience working in a startup that used AI to build products, I will say:
- Firstly, AI is a technology, just like smartphones or the cloud. Similarly, we could distinguish between two types of companies: those that develop technology and those that rely on technology to develop their product or service. I think the first step is to be clear about this, and to know where you stand as a company. But just as creating a cloud or a smartphone is within the reach of a few, developing AI and selling it as a product or service is within the reach of a few. So you are most likely a company that customises and uses AI for your product, but does not develop it.
(From here on I only refer to companies that use or want to use AI to build products) - Secondly, we need to ask for what use case we think we can use AI as a technology. AI is often used for: predictions, recommendations, chat bots, speech-to-text, document classifiers, image classification, video analytics, among other things.
- Once the use case that can benefit from the use of this technology has been identified, it should be analysed like any other new functionality; Try to understand the value proposition and the ROI it can bring to the user. Along with implementation and operational costs and risks. However, the use of AI to implement our use case introduces some important particularities to take into account that can impact both the value proposition and the costs and risks:
- In the world of AI, there is no such thing as 100% accuracy; Unlike the traditional software world, where the aim is to implement use cases that work predictably 100% of the time. This point is particularly important because it implies a change of mindset with regard to the expectations we offer about our functionality. In the same way that Netflix’s movie recommender sometimes makes recommendations that are not aligned with our interests, our app can give us unwanted results. If this is not affordable, then it is better to rethink the chosen use case.
- Our use case may produce biased results. There have been several important cases where the models turned out to be sexist or racist, because of bias in the training data. One such case was when Amazon developed a model to automatically select the best resumes of candidates applying for its job offers. The model was detracting from women’s resumes because they used male-dominated resumes from recent years to train the model, and the model learned to penalise women. In addition to the problem of bias, there is also the problem of the explainability of the results. Today, in general, we have no tools to explain why AI gives a particular result.
- AI is powered by data, and this is one of the keys that can differentiate us from the competition. Do we have a good volume of quality data that can be used for model training? This gives a competitive advantage to companies that already have traction and users, because if they have done things right, they may have good data to train models on. However, depending on the type of data we need, existing regulations and standards protecting the data may prevent or complicate its use. An example is when a company developing a B2B product needs access to the personal data of its customers’ customers, in order to train the models. In this case, access to such data is not always straightforward.
- In the world of AI, there is no such thing as 100% accuracy; Unlike the traditional software world, where the aim is to implement use cases that work predictably 100% of the time. This point is particularly important because it implies a change of mindset with regard to the expectations we offer about our functionality. In the same way that Netflix’s movie recommender sometimes makes recommendations that are not aligned with our interests, our app can give us unwanted results. If this is not affordable, then it is better to rethink the chosen use case.
Finally, when embarking on a project using AI, it is highly advisable to talk to someone who knows the subject well and has done it before. This person will be able to help us analyse the use case and see if it can benefit from the use of this technology. What I do not recommend is to start looking for AI experts to hire on staff, without first seeking external advice. Because on the one hand, it may be that AI won’t add value to our business, and on the other hand, the range of AI specialists is wide and it’s easy to hire the wrong one. Just as when we get sick we first go to our family doctor and he or she refers us to a specialist, this would be the best way to cure our AI pains.