Where are the Unicorns in Consumer AI?

Over the last few years we’ve seen a large number of Enterprise startups raise significant capital to apply machine learning to solve business problems. According to Pitchbook data there are currently over 10 non-listed B2B Unicorns in the US and Europe that can credibly make the claim that machine learning is core to their service offering for the Enterprise, while there is not a single B2C company above a $1bn valuation. This has left me wondering: where are all the consumer companies building machine learning enabled services for you and me?

There are many reasons for this scarcity in applied ML for consumers, which I’ll expand on in the next section. Digging deeper it also becomes clear that we are about to see a new wave of consumer facing applications that are “machine learning first”, where “AI” is the product. A number of factors are now coming together that will enable large and lasting outcomes in ML-enabled consumer applications. An exciting, while still challenging space for a new generation of Deep Tech entrepreneurs.

Paradoxically, commercialization in machine learning did initially occur in the consumer space. It just happened under the hood. The largest tech platforms fully embraced machine learning, advanced research, and helped to mature the developer stack. Most of all, it hyper-charged their consumer offerings without the customer ever knowing.

Some examples: Google started to introduce deep neural networks for speech recognition in 2012 (the same year they managed to recognize cat videos on Youtube) and its RankBrain now powers their core product, Search. From two deep learning projects back in 2012, Google is now pursuing more than 1000 [src]. Optimized routing on maps, translations or image recognition to improve image search results are just a few examples of their successful applications in the wild. Similarly, Chinese search giant Baidu has ramped up its activities and hiring in order to advance their products through machine learning. Amazon built their voice-based search platform Alexa with neural networks powering their speech recognition.

The large content platforms then went to increase relevance of personalization and content recommendation by introducing deep learning. Facebook, who have progressed academic research with their work on DeepFace, a system that uses neural networks to identify faces (achieving 97.35% accuracy when they published their results in 2014), are using deep learning for content recommendation on users’ feeds, ad targeting and more. Spotify refined its music recommendation engine, which was largely relying on Collaborative Filtering, with deep learning to include content-based recommendations [src]. Similarly, Netflix have boosted their recommendation system, as well as other features that improve customer experience and aid retention [src].

Enabling factors for consumer AI applications are now in place

Outside of B2B and established companies, the stars seem to be aligning for consumer AI startups. Why? We now have the enabling technologies, infrastructure and entrepreneurial knowledge in place to overcome the four major obstacles that have held consumer applications back from successful commercialization:

1. Data gap. The big four platforms (GOOG, AMZN, FB, AAPL) are sitting on most of the structured and available data on consumer behavior today. Their network effects are strong gravitational forces that keep users and data off new solutions, by having the inherent value of their respective services increase with scale, thus widening the gap for any company that wants to compete. New entrants are met with this harsh reality (limited access to data) and general doubts about feasibility (slim chance of being funded). What is changing: There is an increasing amount of data on consumers outside of the big platforms, enabled by tech adoption, and accelerated by on-device storage capabilities and decreased cloud storage costs. Regulation is now also starting to restore the consumer’s power over their own data (e.g. the EU directive PSD2 for the financial sector), which helps to free valuable data that entrepreneurs can build applications on. Furthermore, and a bit of a circular reference to this article, will more and more entrepreneurs naturally recognize intelligent ways of building a business that clear hurdles of data access, as knowledge about the possibilities and constraints of ML-enabled models permeate through the market. And there is now a larger subset of entrepreneurs that are able to adopt machine learning in a commercial setting as the developer stack has significantly matured.
2. Value gap. Unlike consumers, B2B customers are willing to pay in advance. We see many companies happily compensating new startups for their R&D (e.g. co-development through paid pilots) because the problems that can be solved with machine learning are immensely valuable to them and have not been solvable in the past. This is not the case in the consumer space, where even if a small number of visionary early adopters exist, their share of wallet is rarely sufficient to fund a business off the ground. Consumer loyalty is fickle and even early adopters disappear quickly once it’s apparent that there is a significant gap between value and vision. Gratification has to happen now. No utility, no usage. What is changing: Along with founders becoming more aware around selecting models that drive significant and immediate consumer utility, we are seeing more investors that recognize the opportunity in consumer AI. They are increasingly knowledgeable (also read: comfortable) in picking successful data acquisition approaches that won’t fail to scale, which will allow entrepreneurs to fund data acquisition and build a working MVP prior to a public release.
3. Technology gap. Building teams to apply state of the art deep learning has been an expensive endeavor, while the enabling technology on the consumer end hadn’t caught up to allow full or partial training on-device. What is changing: Apple’s latest iPhone generation, including its dedicated A12 chip is a factor of 10 more powerful to compute neural networks, making it feasible to do training on-device for the first time. More and cheaper storage, on-device and in the cloud, as well as increased availability of high bandwidth connections further enable ML-applications. Additionally, we’ve seen many advances in the developer stack that have lowered the hurdle to apply machine learning: from Google’s TensorFlow to Amazon’s SageMaker, as well as countless open source libraries and startups offering developer tools. All this is now making it feasible to start a business with deep learning and machine learning in general for non-academic backgrounds. It further reduces set up costs and leaves more resources for data and user acquisition.
4. Brand gap. Machine Learning is not a selling proposition in the consumer world. “Brand recognition” is still low with 23% of consumers recognizing the term (this number is from an Australian study, the only credible research I could find). Not particularly surprising for a field where, anecdotally, even VCs are often not able to distinguish between the terms machine learning and deep learning. And even if the concept and benefits are broadly understood by the consumer, “artificial intelligence” still does sound scary, dystopian, and out of control . Technology is a means to an end, so this won’t change. What is changing: successful Consumer AI companies require teams that have deep tech DNA, but are also highly capable in building products with a consumer branding focus, while understanding and engaging in “digital native” marketing to scale. We’ll increasingly see teams that recognize this and combine both core capabilities in their organization.

Connie Chan, investor at Andreessen Horowitz, recently shared her thesis on AI-first mobile companies (read her post here) where she mentions the few examples of companies that have already achieved commercial success with deep learning at the core of their products. Most notably TikTok (Chinese social video app, that acquired and merged Musical.ly last year and reportedly sees over 500 million monthly active users. Plug: our portfolio company Dubsmash is the second largest player in this category, while building on a different product focus.) and Soul (Chinese dating and social matching app). To some extent, these examples show that ‘true’ consumer AI is not only possible, but possible and a huge opportunity as well.

Where to expect the next AI Unicorn

Don’t get me wrong, there remain a number of “death traps” for entrepreneurs that want to build machine learning enabled applications for consumers. First and foremost, will entrepreneurs fail when building low utility, low frequency solutions. This sounds obvious, but is even more so the case in a machine learning environment (→ value gap). Similarly, difficulties lie in trying to solve heterogeneous customer preferences with complex labelling requirements and features (= large amount of data points necessary), combined with the need to build up a database from scratch. Another route to fast insolvency is entering a market with advanced substitutes and incumbents (e.g. a service for music discovery would have to compete with Spotify, Shazam, Soundcloud) or relying on predictions that have to be computed based on behavioral data from popular social networks.

Despite (or arguably because of) those difficulties, there is significant potential for value creation in this space. Machine learning is helping humans to cater to the most fundamental of needs for self-maintenance and self-preservation and it will give us superhuman strengths in extending our cognitive abilities as individuals and mankind as a whole. The only question is how much do we let the machine into our lives and how much data do we allow it to collect. Because this will constitute the only limiting factor for entrepreneurs in the end.

To size the opportunity of machine learning enabled consumer applications for the coming decade, an analogy to the emergence of Mobile Apps might help. With the iPhone (iOS ecosystem), and Android, we’ve seen new and disruptive business models, but also successful startups building applications where existing and large incumbents from the Web had missed the opportunity to adapt. The same will happen with consumer tech and AI. We’ll see new disruptive business models, but also proven models that will get a makeover through new tech capabilities.

At Heartcore we are therefore bullish on the emerging “Consumer AI” sector and are certain that we’ll see an increasing number of great teams working on this opportunity. To make this a bit more tangible let’s dive into four areas of value creation with obvious potential for large ML-enabled consumer tech businesses to be built.

• Discovery. In the short term I expect most of the AI-first innovation in the consumer space to center around recommendation and decision support. These companies are cutting through the noise in a world of abundant choice. Should we call them Meta Search 2.0? After all, most of the categories that are big in Meta today will be subject to an overhaul in UX, driven by an application of deep learning. Dropdowns, forms, categorization and other UI features of the Web 1.0 era are failing with diverse consumer preferences and longtail inventory. Looking for a job or restaurant, buying apparel, optimizing insurance coverage, planning a holiday; these are all areas where we will see an update in the way we discover products and services. Discovery services will be native to mobile, and potentially to voice-based interfaces in the long-term. Capital efficient data acquisition strategies will be crucial here, which is why I am expecting a large share of eventual failures, next to a few unicorns that will have solved complex discovery problems.
• Cognition. Machine learning will allow us to extend our cognitive abilities for complex decision making and problem solving, so far as to have them being done for us in the background without us being actively aware. Smart assistants will decide for us what to buy, experience, and even carry out tasks for us. Amazon, Google, Apple and Baidu are all in the race to build the next gatekeeper platforms with general personal assistants. But there will be also other specialized assistants that solve complex problems for various domains that go beyond today’s rule based and static solutions. From investment allocation decisions to planning weekend activities to reading through and answering emails, we’ll be able to get more done and better.
• Diagnosis. Advanced recognition of complex patterns and interdependencies through deep learning will reach consumers in the Health sector. New entrants will not only diagnose better, but also improve how treatments and medicine are prescribed and applied. A natural obstacle (and opportunity) for entrepreneurs lies in the regulated availability of data in the healthcare sector. While it will remain dominant for most HealthTech companies to reach consumers through payers and providers (e.g. due to efficiency or regulatory restrictions on marketing), we expect some new category leaders that market directly to consumers. One example is our portfolio company Kaia Health who are building a “Digiceutical” to treat back pain (machine learning is powering their video based assistant to make real-time posture corrections during training sessions). An important player to watch in the sector will be Apple who are emerging as the dominant platform in Health with its increasing penetration of on- or close-to-body sensors (Apple Watch, iPhone and AirPods).
• Augmentation. The most obvious sector to profit from recent consumer tech advances that allow endpoint or on-device training is the Augmented Reality / AR space. This enables new possibilities to adapt what we see and hear. Neural networks will be utilized for image, object and face recognition, the creation of virtual objects, as well as the manipulation of images and sound. Entrepreneurs can now build advanced applications that help us navigate the real world, e.g. by giving additional information on physical objects, people and points of interest, or allow people with physical disabilities to see better or hear better. This is an exciting field, but it’s early on. One reason: other enabling technologies like earphones and and smart glasses are lacking necessary product maturity and adoption, so it will take a longer period of time until we’ll see disruptive models here.

At Heartcore we are actively looking for the next ML-enabled consumer tech success in these areas (and the many more we have yet to recognize). If you are an entrepreneur building consumer applications around deep learning please reach out to us through this typeform. I would love to chat about your vision, technology stack and data acquisition strategy.

PS: Thank you Deepka, Max and Saket (from Lafamiglia) for your insightful comments and thoughts on this.

📝 Read this story later in Journal.

🗞 Wake up every Sunday morning to the week’s most noteworthy Tech stories, opinions, and news waiting in your inbox: Get the noteworthy newsletter >