The Top 5 “Recipes” That Give AI Projects a Higher Likelihood of Success

By Aman Naimat On Oct 16, 2018

For artificial intelligence (AI) to enjoy continued success and uninterrupted investment, it should be less academic and more practical, less esoteric and more straightforward. AI projects need to be grounded in clear-headed thinking and consistently aimed at producing net new value in business or society. In other words, we want to put the intelligence back into artificial intelligence.

We believe there are five core “recipes” for AI, intended to be used by executives, entrepreneurs, product managers, engineers and data scientists. Taken as a holistic “cookbook,” they represent a common language and fluency for developing successful AI solutions.

Start with Spreadsheets

Instead of an “AI First” approach (which echoes the “mobile first” mantra), a more important approach is being “Data First.” For every big data problem, there is always a small data solution. Billions of rows won’t always tell someone what they need to know; rather, most times answers can be found in very small sample sizes, which once validated can be expanded to billions of rows. It is more important to get to the right answers at low scale and then scale later.

Nowadays, many companies have the misconception that in order to be successful in AI, they must aggregate every bit of data across the entire organization. What’s happening instead is that it takes years to get all this data in one place, and yet even then, AI can’t be built on top of it. These types of projects drag on for years then eventually are canceled in a cycle we call Data Lake Hype. Building massive data lakes is the wrong approach to optimally utilizing AI. If you can get a human to look at small data and find the answers, then you can later double down by applying AI to the larger dataset.

Pick the Right Problem to Solve

For AI to be impactful, you should pick problems that are most lucrative and/or most meaningful to your business. If you try to address the wrong problem, you may never be able to solve it or it might turn out to be a problem you don’t want to solve. It is not very helpful to go looking for problems your data scientist (nothing personal since I am a data scientist myself) can work on as if AI was a magical potion to solve any problem.

Akin to starting a company, it is better to identify a juicy problem and then ask, “Can data science solve it?” Many problems cannot be solved through AI and for most it may be overkill; it is imperative to quickly identify when this is the case. You will save yourself a lot of time and money.

Perfect the Solution Now and Scale Later

Donald Knuth, the Stanford computer scientist considered the father of algorithm analyses, famously said, “Premature optimization is the root of all evil, if not most of it.” Don’t try to scale immediately; find the right answer first and scale later. AI and data lakes are more of a science problem than an engineering one. The ideal approach is to solve the problem quickly with limited data and features, using simple models such as decision trees or regression analysis.

If you already have some intuition on the problem, build a simple solution around that intuition for a quick win. Then, and only then, should you add more complexity, sophistication, and linguistics. In fact, most AI models show promise — or prove not to work — with less than a dozen features.

Avoid Black Boxes

For people to believe AI works, they need to be able to trust it. Trust is a key area closely linked to the adoption and success of AI products. Humans will take action only if they trust the AI models, and they will only trust the AI if it is transparent. Humans don’t trust a black box.

Let’s say you’ve chosen a solvable problem, picked the right AI model and, with the right amount of data, solved the problem. What we have found is that people will not believe or trust your model even if it’s 100 percent accurate (and it will probably never be 100 percent accurate). Perfect, or close to perfect, doesn’t really matter if humans don’t trust the technology.

So, how can we solve this problem?

Tie it to metrics that matter to users, such as revenue to justify your AI. Don’t invent some arbitrary score. If I’m looking for a new house, show me the range of offer prices that will most likely allow me to make a winning offer. Don’t invent scores that are disconnected from the problem or the real world.

Run backtesting with user data. Use historical results to show how the data would have provided the right answer every time.

Show all results, both positive and negative. Show me everything, including the failures. Humans are very interested in looking at what we did not find; this builds trust.

Provide transparency on the underlying features that drove the recommendation or classification.

Make Your Results Actionable

While businesses are adopting AI quickly, they can’t just add AI models and a data scientist on top of any business problem. AI needs to involve action and be incorporated into a larger workflow. Because without action, nothing will actually happen. I have run into infinite projects (including my own) where the end result is a recommendation or a rank.

The problem is that humans are too busy and too distracted to look at the billions of input from all sorts of sources. For AI to have an impact, it should take some action at some confidence. That’s how continuous value can be generated. Minimally we should think about how we can integrate the ML models to external systems that can take some action, send email, alert someone — some workflow that ties it to the real world.

A perfect example is two cars that I recently test drove. One gave me a blinking light when the car would be close to an imminent crash, the second actually braked the car automatically. Which car would you prefer to be in?