Remember when dashboards looked back at you dumbfounded that you achieved all of those numbers? Or when they proudly sorted data based on the few possibilities there were? These were useful functions. You could generate reports on stock positions, salary expenses and more to help you manage supply chain, plan resources or establish budgets for the quarter. As much as these were helpful, they were purely diagnostic. The events causing those numbers had already happened — stock positions had moved, expenses were due. It was relegated to the good senses of the humans using these applications to stay the course and utilize the numbers and figures on screen for the best benefits.

Applications have evolved and things have changed remarkably since. Applications can learn and understand where you could go, what you could do, who you could meet and even what you might like to eat. All of this, as you may notice is prognostic. This puts businesses in an enviable position. They can now understand customer behavior actively based on data and deliver personalisation at scale. Not only that, applications can predict business-relevant events ahead of time and help leaders prepare for outcomes.

Every year, applications seem to do more things that only humans could do before.

Evidently, businesses are keen to apply artificial intelligence to all of their solutions. According to a 2017 IDC report commissioned by Salesforce, this will be a big year for AI in terms of adoption. Over 40% of businesses surveyed globally claimed they will adopt AI over the next two years. In fact, by the end of this year, it’s expected that “75% of enterprise and ISV development will include cognitive/AI or machine learning functionality in at least one application.” But in this quest for businesses to become more intelligent, they will need to take a step back and ponder how all of this will be possible so they don’t miss the wood for the trees.

Here are five principles from my own personal experience of deploying Vymo, an AI-enabled enterprise assistant, across 50-plus clients.

What Volume And Variety Of Data Are Your algorithms Learning From?

Google’s Research Director, Peter Norvig, famously claimed, “We don’t have better algorithms. We just have more data.” The general consensus is that it’s useful to have as much data as possible. As a rule of thumb, Yaser Abu-Mostafa, a professor at Caltech, recommends “roughly 10 times as many examples as there are degrees of freedom in your model.”

Also, it is advisable to keep models simple so that they do not “overfit” (e.g., when a model correlates extremely tightly for a limited set of data points, explaining idiosyncrasies in the data set rather than the underlying trends that would be generally true).