There is an I in both BI and AI, and the corresponding English word is “Intelligence”, which appears to have the same meaning.

Actually, it’s not the case. Even if we only discuss issues in the field of business decision-making, there are significant differences between the I in BI and AI.

Broadly speaking, tasks that rely on data to assist business decision-making can be called BI, and BI can be said to include AI. BI is the purpose, AI is the means, and BI can be achieved through AI means or other means. In this sense, the I of BI will be greater than that of AI.

However, the term BI has now been severely narrowed down. By BI, we often refer to interactive multidimensional analysis of data. That is to say, looking at statistical values from various dimensions, there are routine actions such as pivot, slicing, and drilling, and people may try to identify problems in business operations during the interaction process. For example, if a product has a large sales volume, drill down to see which region or sales method caused it, and then apply the underlying business rules for promotion.

To be honest, this type of BI is very simple in terms of technology, and the mathematical knowledge used does not exceed that of middle school. It is basically addition, but there are many conditions. Most business users can understand, while all database engineers can implement it. The interface of BI products is often flashy, and the use of good products is very smooth, but there is not much technical content (i.e. difficulty), and it can be done meticulously with a lot of effort (there is a lot of complexity). The technical difficulty of BI lies not in the interface, but mainly in how the backend can achieve high-speed response when the data volume is particularly large. After all, BI is an interactive operation scene, and it is useless to have a respond after a few minutes of waiting.

AI is different. In the field of business decision-making, the main focus is on model-building and predicting, which means using historical data to predict the probability of certain events (don’t think about stock trading, there is currently no publicly available mature AI technology that can continuously predict stock trends). For example, banks can predict the default probability of loan customers to determine whether to issue loans and loan interest rates, merchants can know whether packaging certain products for sale can increase the probability of users purchasing to determine sales strategies, etc. Obviously, this information has significant implications for business operations.

To implement these tasks, engineers need to have rich statistical knowledge and be familiar with various algorithms of machine learning. This technical content is very high, and there are not many people who can truly master it. There are many open-source packages available for model-building and prediction, which can also be used by ordinary programmers. However, if you don’t understand algorithms and simply use open-source packages, the models built are often poor, and you don’t know why, and there is no way to improve them. Because a good model requires a lot of data preprocessing and model selection work, and various algorithms also have a large number of parameters to set, some people are not even clear about what a normal distribution is, so of course, it is impossible to build a good model.

Even for business users who only use these models, in order to evaluate the effectiveness of these modeling predictions, they need to learn a lot of knowledge and concepts, otherwise they will not be able to understand them. AI does not have a very low threshold for either engineers or users.

So, AI and narrow BI are completely different, these two are far apart. BI’s I is a relatively easy IT technology problem (especially in the front-end), while AI’s I is a mathematical problem that requires N years of accumulation.

Many years ago, there was a famous beer diaper story in the industry, which was told by SAS company that putting beer and diapers together can increase sales. However, this story is often used by BI manufacturers to argue. In fact, the multidimensional analysis interface of BI cannot discover this pattern.

Nowadays, some BI vendors are also adding AI features to their products, which means adding some open-source packages. This situation can be considered a gimmick. Because if this manufacturer can do well in AI and the threshold itself is high enough, they can completely develop a product independently, without the need to hang on to a low threshold BI product.

Expanding the scope of AI further, not just limited to the field of business decision-making, such as the trendy generative AI, will have some correlation with BI. The goal of modern narrow definition BI is to enable business personnel to analyze data, using a smooth visual interface. If generative AI technology is used, machines can directly convert human language and text into data calculations, which can further reduce the application threshold.

But current technology is still difficult to achieve this., because the data structure of enterprise applications is often very complex, making it impossible to simply put it in the GPT prompt (far beyond its length), and the cost of doing fine tune is also relatively high, the key is the lack of materials. On the other hand, the syntax structure of SQL is only suitable for single table queries. When the association between multiple tables is not unique, there can be serious ambiguity, and AI does not know how to match them.

At present, the SPL team is researching this issue. Using DQL technology can solve the association ambiguity of SQL, and then using GPT technology to layer natural language. Combined with SPL’s deep understanding of data structures, it is possible to make correct and sufficiently complex natural language queries. At present, we have achieved basic Chinese query capabilities and are currently developing English.