Winning AI Strategies – Part 1: Startups

This article is the first of two on winning AI strategies, focused mostly on startups.

So let’s get very tactical and talk for a few minutes about how best to successfully implement AI. There are three groups of players in the AI field: Big tech, startups and established companies. Each has their own approach to AI strategy and implementation.

AI is core to the success of big tech

Big tech includes Google, Facebook, Amazon, Apple and Microsoft. AI is a strategic imperative for these companies, and they are competing aggressively in this field. They have the largest number of AI programmers and most sophisticated tools, and have built their businesses around a core of AI technology. These companies have a lock on data for billions of users, and this data fuels their AI engines for consumer products. In addition, most of these companies offer AI platforms as a service to business customers. In 2017, these companies spent $22 billion on AI-related M&A, and this appears to be accelerating. It’s worth noting, some Chinese companies like Baidu, Tencent and Alibaba are also becoming significant AI players.

We’ll explore the specific strategies of each of the big tech players in subsequent posts.

Startups are nipping at the heels of big tech

While big tech has made major advances in AI, startups have been innovating like wildfire. There are over 3000 AI startups in 2018 worldwide, of which 40% are in the US. The US AI startups raised over $9 billion that year. While these AI startups may lack access to the consumer data of big tech companies, they have made up for it in attracting the best talent and developing novel algorithmic approaches. Many startups focus on building business AI solutions where their customers provide the necessary domain expertise and data. Increasingly, these startups build their AI solutions on top of the very platforms provided by their big tech brothers allowing them to innovate and scale more quickly.

Startups have the luxury of being very focused and nimble. But time is their enemy since competition is forever at their heels and resources are limited. This is even more true in the AI context, and in some ways harder to manage.

Successful startups solve specific real-world problems by marrying deep domain expertise with machine learning talent

In order to stay focused, it’s important to have clarity about what real-world problem you’re solving. This usually requires domain expertise to understand what’s not working today, as well as sufficient technical understanding of what AI can achieve. Having a metric for measuring success, and a specific goal post will help keep the company focused.

For example, focusing on robots to pick up objects for packing boxes at a warehouse is a real-world problem. As mentioned before, performance can be measured by mean-picks-per-hour, and achieving a rate better than humans would be a clear goal. Covariant is a startup focused on this very problem.

To survive and thrive, startups need a scalable approach to acquiring (and labeling) quality training data

Startups need a well thought out and scalable approach to data to make sure they can deliver on the promise of AI. Quality data is the lifeblood of machine learning. In most cases a startup will need to partner for access to this training data. Scraping or snagging data may seem appealing to fledgling startups, but many prototypes never scale because they don’t have a way of getting legitimate data as the company grows, or the data quality degrades as they expand. Best to see this coming as soon as possible by casting a broad net and sampling data from multiple sources.

Also, startups need to think about how to scale their data cleaning and labeling processes. Likewise, data security is also something that should be addressed from the start, especially if the data is especially sensitive or valuable to hackers.

Competitive advantage for an AI startup comes
rarely from its algorithms

The first principle for AI implementation is don’t reinvent the wheel and think carefully about what should be your technical secret sauce. The good news is that there are a number of powerful platforms that startups can leverage to test and even launch their AI solutions. Using these cloud-based tools will allow the startup to seamlessly scale with its business. Examples include Google’s TensorFlow, Facebook PyTorch, and Microsoft’s Cognitive Toolkit. Plus there are open-source libraries, such as Keras, to extend these tools even further. And with tools like AutoML, even the model selection and configuration process may leverage the wisdom of the crowd.

At the same time, it’s worth thinking about where you might need to have a distinctive technical approach. Sometimes this is driven by specific performance requirements or applications. Where connectivity is a constraint, edge AI that works on users’ devices may be required.

Another principle of AI implementation is to walk before you run. In other words, find ways to add value quickly in the early stages of the implementation. This can be in the form of a technical proof of concept or a set of business case studies. Even better yet, adding value to customers out of the gate can provide your startup with some early revenue. Along these lines, designing the system to have a human-in-the-middle can often speed development and be helpful to shake out corner cases. In any case, since you’re developing a learning system, once you get to a minimum viable product, the AI can continue to learn and evolve to get better over time.

Startups have an advantage attracting and retaining the best AI talent by sharing upside

It’s also important to have a plan to attract and retain the right talent. Machine learning and data science expertise is less scarce than it was a few years ago, but getting the right kinds of people for your team is still challenging. Think more engineering and less research. Also, depending on the solution, don’t be afraid to bring on people in related fields like psychology and linguistics, since diversity of thinking, as well as people on the team, tend to yield more robust AI models.

Even startups need to think about ethical issues
related to AI early

Startups need to establish ethical boundaries early. At the very least it’s worth setting out ethical principles for the use of AI. This prevents problems down the road, and helps build a strong ethical culture for your startup from the ground up. Also, depending on the nature of the application, the establishment of failsafes may be critical to avoid harm, both expected and unexpected.

There are over thousands of AI startups across almost every industry, but execution will likely be a big part of determining winners from losers.

In our next article we discuss winning strategies for established companies, and how it differs from startups.

Note: this article is based on a lecture given by me at UC Berkeley, Haas School of Business, as part of the “Business of AI” online course released December 2019