Although many software packages with built-in artificial intelligence (AI) are readily available on the market today, a specific customer solution still requires a lot of traditional handicraft. However, these specific systems allow AI to actually be used for those business processes in which they drive the most benefit. The question is: which business processes are the most suitable for applying AI?
AI is not that smart, but it is very fast
Now that the dust from the current AI storm has settled somewhat, we can see what its importance will really be. In addition to all those beautiful, interesting, and fascinating results to emerge from AI laboratories, the first commercial applications are beginning to emerge. These applications may have a narrow field of application, but they do perform quite well. They primarily conduct data analysis and pattern recognition, for example interpreting images and text, finding patterns in figures, and running intelligent searches. AI successfully diagnoses medical scans, interprets application letters and CVs, and tracks down fraudsters.
The value of these applications lies mainly in the fact that computers can do things faster than humans. Whether AI really does the job better is still the question, although AI is less likely to overlook things the way people sometimes do. The general trend is for AI to replace boring and repetitive human work. Examples are: reading thousands of pages of police reports in search of cold cases, interpreting medical scans, and predicting customer behavior at retailers.
“The information society resulted in the mechanization of information processing. By using machines (computers) a lot of manual paperwork will be taken out of your hands.” (Floriaan Hornaar)
What applies to all these cases is that real human intelligence is not imitated in its full breadth. Only limited tasks are automated. I see AI as the next step in mechanizing work. AI is not advanced enough to take over entire processes. IBM explicitly states that AI must support people (augmented intelligence) and not replace them, although it is tempting to automate entire processes to gain the most efficiency benefits.
AI is conservative
Currently, AI is mainly based on machine learning and big data. Without large amounts of data, it cannot recognize the patterns on which to base its interpretations and predictions. But data is about the past – the analyses are based on the past – and past results cannot provide guarantees for the future (especially when it comes to exceptions) because exceptions do not occur enough in the dataset. This implies that we can use AI mainly in those processes in which we do not want to see creativity or improvisation. It also implies that AI is not that good at discovering new trends and developments early on. Moreover, it is difficult to teach AI something new if, for example, we want to use new and innovative processes. We may have to repeat the learning trajectory of AI all over again.
“Many conventional AI systems are merely machine learning, or neural networks, or deep learning. They’re good at handling large sets of data but lack situational awareness or the ability to navigate around missing or incomplete data. They get stuck.” (AJ Abdallat, CEO of Beyond Limits)
And herein lies the crux of the matter. AI is not difficult because of the technology; AI algorithms are freely available. But training AI systems takes almost as long as teaching humans. This is true not because computers learn slowly, but because we first have to compile the “educational package” for the computer. And this package consists not of a few sample books, but of a whole library that must first be checked for suitability. Integrating AI systems into the business processes also requires attention, but that’s the case for every new technology.
AI must have added value
Where can we best use AI today? The best thing is to keep it simple. The investments in implementing AI must be earned back. The higher the investment, the more difficult it is to guarantee a decent return. Many business cases around AI projects are difficult because the investments cannot be recouped. They remain in the laboratory phase and that is a pity. How does this happen? Many pilot projects involve technically and functionally interesting applications, but these applications have little added economic value. For example: how much money can you save if you apply AI in recruitment and selection? At most, you can save a few FTEs. That doesn’t amount to a decent ROI for your AI project. But didn’t we just say that AI will support people and not replace them?
“Machine learning is cool, but it requires data. Theoretically, you can take data from a different problem and then tweak the model for a new product, but this will likely underperform basic heuristics. If you think that machine learning will give you a 100% boost, then a heuristic will get you 50% of the way there.” (Martin Zinkevich, Google)
The solution can be found in two areas. The first involves applying AI to new, innovative activities – things that your organization does not do now, activities that cannot happen without AI. This sounds very exciting, but, as any innovation project, it is also ridden with risks. In some sense, you could consider yourself lucky if you find a promising innovation with AI is the core of the solution.
The solution can also be found by applying AI to the primary bulk processes, such as medical diagnoses, claims handling, or fraud detection. AI does not have to be all that efficient because frequent small savings add up to fairly large amounts. Tom Rickert, partner at Next World Capital, claims that AI, used as a tool, can realize an efficiency improvement of two to five for the process step where the AI is applied. (When fully automating gains up to hundred times can be achieved.) In our case for tooling, AI becomes interesting only when it is used frequently – so basically, only within the primary processes of an organization.
After all, starting with the application of AI we must first have a clear picture of the added value – or actually customer value – of the system we want to create. Anyone who thinks they can buy the golden AI goose may be in for some very unpleasant surprises. There are so many myths surrounding machine learning and AI, you first have to study carefully what AI really is and know your own processes. Only then will you know where you can best apply AI for the best results – with the greatest value for your customers.