The worst waiter in the world - no action = no value = no sense

These are additional shownotes. For the full transcript of the episode, scroll down

Getting started is the hardest thing, I must say. Which story to lead with? Well, I like this one, simply because it is the most important one. AI and data science are just means to an end, so the most important story cannot be about AI itself: what it is, how it behaves, what to look out for,... It should be about the fundamental goal you are trying to achieve. And for me that’s easy: you want to create value, and to do that, well, you need to take action. Therefore, building AI systems where there is any ambiguity about the actions they generate, or where the impact of those actions is reduced, doesn’t make sense.

This baseline is surprisingly controversial for some. I often hear arguments like: “But we cannot just trust those AI systems, right?” Or: “But they are too expensive to build!” True, those are all attention points, but the essential message is the shift of the burden of evidence that was mentioned in the podcast. If you are truly committed to transforming your business and believe that the future is digital, then digital should be the norm. Trust, for example, should be established by creating proper support systems and governance. It is easy to come up with reasons not to go straight-through digital, but keep in mind: your competitor might do it. Or some startup might do it and turn into Amazon, who thought that sales and delivery could be automated. Executives are paid to prepare their organization for the future and cannot afford to be too conservative. More and more, this means they have to ruthlessly identify bottlenecks that stand between the organization and the value, and then eliminate them. And those bottlenecks are not the humans, but yes, they are often the human steps in our processes.

One simple but concrete example is found in document handling systems. The AI system first extracts the text from potentially scanned documents, and then extracts information to identify and execute the appropriate action. If this process is fully automated, often most customers can be helped in the blink of an eye, and at substantially reduced costs for the organization. However, all the value of such a system can be destroyed by insisting on one simple thing: that it only serves as ‘insight’ to a human, or even just that a human checks whether the action is appropriate. This would

require the human to read the document, causing delays and again increasing the human workload, limiting the scalability during busy times,… Organizations should realize there really are better ways to create trust and checks and balances in such a case, and consequently putting the burden of evidence on the human element might avoid the destruction of value.

In practice, the focus on driving actions is often missing in various ways. In the past, I’ve often been asked to build systems, but when I asked how they would be used exàctly, people couldn’t tell me. Or organizations fail to adapt their processes around the solution, leading to suboptimal business outcomes because the actions are not executed optimally. Surprisingly, it is often left to those building the AI system to clarify its later use, even though they are often technical experts with little context knowledge, and they will most certainly not be the ones using it in their operations. As a result, solutions can become stuck in the proof of concept phase because they do not fit the exact need, because nobody got the buy-in from later users,… At the center, this is something good leadership could solve by instilling an action-oriented mindset in the organization, and by creating a governance that focuses on practical actions.

Another symptom of not critically considering the role AI plays in your organization’s activities is the lack of measurement of the value of your AI. I have met with dozens of organizations who could not tell me the value they reaped from their AI systems. Sometimes this is due to problems in measuring, but more often it is simple lack of measuring, and even more often than that, it is because the organization doesn’t have a definition of value. And creating such a definition, well, that really starts by understanding how the systems lead to action. In this respect, it is revealing that the exact value of dashboards is often shrouded in mystery, and post-deployment value measurements are rare. The root cause of this is often the lack of understanding of how exactly the dashboard has impacted operations – something which is inherent to the ‘insights’ paradigm.

The mere number of dashboards might even provide a measure of data maturity: an organization with too few dashboards may be immature, but one with too many may also be immature. Many organizations suffer a real overload of dashboards. I once worked with a mid-sized organization that started counting their operational dashboards, but simply stopped when they reached around 8,000, with no end in sight. And nobody could really say what the value of all these dashboards was, or even whether they were still being used. And contrary to what people think, dashboards do not come cheap; in another organization, the total cost of ownership of an average simple dashboard was estimated to be between 30,000 and 40,000 euros. And if ‘citizen data science’ is seen as a cheap workaround, think again. Because data provides just another language to describe reality, and like any language it has many subtleties that may not be understood by everyone. In every organization it is easier to get data that is lacking in at least one respect, than it is to get correct, complete, up to date, relevant data. And every data analysis or even simple aggregation still requires interpretation, yet another problematic hurdle for most. Exactly because citizen data science is cheap it is used for less important matters, and fewer safeguards are applied in the creation process compared to e.g. the development of AI algorithms. Typically little effort is taken to adapt or monitor operations appropriately, to educate users, to follow up on quality of output, maintenance costs are not anticipated,... As a result, the tools created through citizen data science risk becoming “zombies” - numerous, toxic zombies that slowly poison and weigh down the organization. In the end, they may just create stronger and more pervasive noise for those trying to read the signals to steer operations effectively.

Despite these previous comments, dashboards and business intelligence (BI) are vital tools for successful organizations. While I may have seemed negative about BI, my intention was to highlight that its true purpose is not always kept in mind, and that it should not be the default design pattern in a future-proof organization. The proliferation of dashboards is often due to a limited view of what it means to be “data-driven”. It is not enough to ‘include some data in your decision process’. A future episode may delve deeper into this, but at a minimum, it should also involve a reflex to exclude personal assumptions and interpretations from the process. In this sense, dashboards should strive to be objective, presenting hard facts and measurements rather than attempting to convey insights, which are essentially readings and interpretations of facts. Usually the analysis is not rich enough for such interpretation anyway, giving questionable guidance concerning needed actions. And besides this, BI merely strives to better inform tasks executed by humans, meaning many value-generating mechanisms (like reducing throughput time, reducing human workload,…) do not or almost not come into play. With AI, we should really aim higher with our ambitions, and strive to build systems that can execute the tasks themselves.

Finally, there is one more benefit to having actions driven directly by AI systems: it creates clarity about responsibilities and makes it easier to provide safeguards. When processes involve both human and AI contributions, governance can become complicated: what is each’s responsibility? Where and how should safeguards be implemented? With AI it is clear that the needed safeguards should be built by technical specialists with a specific scope in mind. However, if that scope only covers part of a process, governance that was good originally might become insufficient once the human steps of the process start to evolve in interaction with the AI. As a common example: when AI systems only provide suggestions to humans, the assumption is that the humans will inherently double-check the AI output. This will decrease the emphasis on creating other potential safeguards, but typically over time humans will learn to trust the AI and lose their critical reflexes, effectively leaving the process without any meaningful safeguards at all. This is just one manifestation of the so-called ‘automation paradox’, a very common paradox when humans collaborate with automated systems, often with grave consequences.