AI-first Product Development and Digital Transformation

In an article in The Wall Street Journal in 2011, Marc Andreessen shared why he thinks “Software Is Eating the World”. He was right; in the past decade, we saw the mass adoption of software in every aspect our lives; from search, shopping and travel to health, finance and beyond. This led to the generation of an unprecedented amount of data. Such large datasets, when paired with powerful (and relatively cheap) compute power, allowed scientists to explore new frontiers in AI. For instance, some of the old algorithms (e.g., neural networks) showed “superhuman performance” in domains such as computer vision, natural language processing, and reinforcement learning, to name a few.

This feedback loop of “more data >> better AI >> more mass adoption / funding, and repeat” seems to have enough fuel to continue all the way to achieving Artificial General Intelligence — a big and near-impossible milestone for mankind and machines. Therefore, it is difficult (if not impossible) to imagine software without AI. Not dissimilar to, but far bigger than, how hard it was to imagine serious software without some form of a database in the past decades. As a result, verticals such as health and finance — which are going through digital transformation at scale — should embrace AI-first product development, in order to achieve AI-first digital transformation.

AI-first products are apps and software services that employ machine learning to inform and assist their users.
Apple’s iOS for iPhone is a familiar experience that delights users with both inform and assist features. Map’s push notification is a good example of informing (i.e., giving you the information you need, when you need it), and Mail’s unsubscribe feature is a simple example for assisting (i.e., helping you to actually carry out actions). While such features appear to be simple, single experiences, behind them are multiple machine learning models. For instance (and of course, if we oversimplify), in order for the Map experience to be 90% accurate, you will need each of the underlying predictive models (e.g., Am I in a traveling car? Where am I going? What is the traffic like? What is the ETA? Do I want to see this push notification?) to be far more accurate than 90%, from basic laws of multiplication of probabilities. This high bar for model accuracy has become a lot easier to meet in the past few years, increasing the feasibility of an AI-first strategy.

Of course, this definition goes beyond adding a few “AI features” to an old software product, and rather implies a rethink / reimagination of the offering / UX in light of what AI can make possible; AI is a key driver of the design from the beginning, rather than entering the product roadmap later. Despite the great potential for AI-first digital transformation in many (if not all) industries / verticals, there are multiple challenges that corporations will face in achieving success with AI. The three main ones are 1) talent, 2) data, and 3) product-market fit. Therefore, the article will finish with some thoughts on each of these.


A true AI-first strategy requires top talent in machine learning, design, engineering, product and strategy. On the other hand, attracting such talent away from the top AI destinations like Apple, Google and Amazon is increasingly difficult. Even the tech giants sometimes struggle in competition with startups and have to “acquihire” small teams away from the competition. If you are wondering why, maybe you haven’t watched this interview with Steve Jobs, or read this article in New Yorker. Or perhaps you haven’t come across the plethora of online content on the 10x effect that great talent can have on a product’s success.

As a first step towards the AI-first journey, large corporations need to embrace the right vision, culture, and incentive mechanisms to attract, retain and develop scientists, engineers, designers and product strategists with AI expertise.

Successful AI-first products are developed at the intersection of three key skill pillars; building a team to cover all these areas should be the first and most important objective.


Without the right data, conditions for success for AI do not exist. But what is the right data, and how should a company define its optimal data strategy for its AI-first journey? In theory, many old verticals — such as health or finance — have historical data. Such data are usually scattered across many systems and their linkage and standardisation is proven to be a huge challenge — evident from the famous anecdote that “80% of a data scientist’s time is spent on data cleaning”. Even when linked and standardised, they are not granular enough or well-structured for machine learning usage because the original purpose of such data was simple reporting / MI dashboards.

Two guiding principles can increase the chance of success for a data strategy:

  1. Data should be collected to train the AI. If you can train the machine to become intelligent with the data, you pretty much can do everything else with that data. This also means that if you don’t have proper AI scientists around the data strategy table, you are increasing the risk of failure for your data strategy; they are the ones who can help with the “what’s the right data for making the machine intelligent?” question.
  2. There is no such thing as “the right data strategy”. Your strategy’s power lies in its agility and speed in responding to mistakes, and correcting course.

Product-market fit

Advances in AI do not automatically (or, necessarily) translate to useful products and services that people want. The main reason products and startups fail, is because they don’t meet customer needs in a way that is better than other alternatives; this is the lack of “product-market fit”. To borrow Dan Olsen’s words (i.e., “problem space” vs “solution space”), while AI can be a great tool in the solution space (the HOW question), product strategists and designers are needed to map the problem space (the WHAT question).

Starting in the problem space first (or, solving the “what” question first), can lead to higher chance of product-market fit.

To avoid the product-market-fit trap, corporations need to get out of the building (or GOOB), as Steve Blank put it. Furthermore, they need to learn ways to reduce the cost of failure, to enable the increase in frequency of innovative bets. Embracing methods like GV’s design sprint is one way for corporations to start creating fast prototypes and test them.