Do focus on pain points and use cases where artificial intelligence can address previously unsolvable problems or tackle existing ones 10x better.
This point about looking for “10x” improvements as opposed to marginal gains is becoming horribly clichéd across the AI space, but it is worth addressing. Founders face several challenges in living up to this benchmark when Applied AI solutions often end up being incremental at best. To overcome this, look for applications where the core benefits or attributes of Artificial Intelligence technologies suggest a significant step-up in the category or current solution to which it is being applied. For instance, if you are working on a machine-learning bot using natural language processing to execute more efficient Q&A, think of the most common pain points in human-to-human interaction. Customer service can be painful for consumers and expensive for businesses, hence bots that enhance customer experience pose an obvious use case. Going one level down, traditional retail banks whose user interfaces are limited and customer service teams are particularly inaccessible, present a good opportunity for a bot that plugs into personal banking portals/data whilst providing additional value by helping you track your saving or spending habits.
Don’t become a solution in search of a problem.
Back to the old problem of tech push versus market pull, there is no easier way to spend a lot of time (and money) than developing cutting-edge AI technologies without first thinking about their applications. While many AI startups have been built and acquired (see below) based on technical capabilities and a passion for the promise of AI technology, they have had clear technical applications. That said, the next wave of AI startups will be industry specific and use-case first.
Do use ‘off-the-shelf’ solutions.
There are a whole host of open source libraries out there as well as affordable software offering machine-learning as a service. You can find a pretty good list here but, depending on what you are working on some of these tools can help get you started:
Microsoft Azure is the closest thing to machine learning as a service, providing an impressive suite of tools for you to build your own models and applications on top of existing libraries in a simple and intuitive way. Similarly, the analytics package built on top of IBM Watson lets you throw data at it and it throws back insight.
Scikit-Learn is an open source library for advanced analytics in Python, such as carrying out classification, clustering or regression. Combined with matplotlib, it can produce some powerful insights and visualisations.
TensorFlow is Google Brain’s (recently open-sourced) software that enables you to carry out deep learning on your data through graphically represented neural networks. To see it in action, check out some of these awesome videos.
Apache Spark combines your existing workflows in Python or SQL to carry out analysis on larger datasets. If you are working on tasks requiring natural language processing (parsing, sentence segmentation or speech recognition for sentiment analysis) Apache Open NLP also offers a great toolkit.
Don’t focus your time and energy on building AI infrastructure.
In the same way cloud platforms and services were dominated by giants such as Oracle, IBM, Amazon and Cisco, AI infrastructure will remain the preserve of leading engineers at Google, Microsoft and Amazon (not to mention an army of researchers working on open source solutions). If the solution already exists, why try to close the circle? Better to build on top of it.
Do come up with creative ways of creating and obtaining meaningful datasets.
Data is everywhere… at such an early stage you may simply be considering how to build the data architecture that supports possible AI applications. However, if you want to get cracking sooner, why not come up with more creative ideas around building datasets through APIs into open source or useful, paid-for databases. Better still, why not explore the potential for mutually beneficial partnerships and innovative business models that allow access to proprietary or hard-to-access data.
Don’t focus solely on the back-end build or massive proprietary datasets.
If you are focusing on this, you are probably late to the party… There have been some impressive exits following the first wave of research-driven companies (Deepmind and Nnaissence in particular) ‘acqui-hired’ before generating revenues, and the second wave of machine learning infrastructure firms bought soon after commercialising their products. However, the next wave of AI startups will be characterised by the disruption of verticals and value chains.
Do make a product out of your work and sell it time and time again in search of a scalable business.
Other investor blogs make a compelling case for solving industry-specific problems specifically by combining AI with subject matter expertise. Consider how you might build out your team to ensure you have the right balance of technical know-how but most importantly product, design and development expertise that can execute the Applied part of your AI offering. Startup principles of speed of commercialisation and product execution remain just as important, and your team has to reflect that.
Don’t get too comfortable on service revenues, it’s hard to scale a consultancy.
If you aren’t building a product, platform or SaaS offering, then there is a high chance you have fallen into the consultancy trap. Startup teams experimenting with AI are often a bundle of sought-after skills and capabilities. While you may be able to sell these at a premium or pose as attractive acquihires, you could go after a much bigger prize: Exponential growth.
Do look out for vertical niches that require the specialisms afforded by AI without compromising on venture scale opportunities.
Look out for industries currently leading the charge in AI adoption, which tends to be already highly digital and early-adopters by nature (high tech, telco, financial services) and think about categories where AI can have a transformative impact. For instance, Stanford University identified several including: Healthcare - to enable improved care / diagnosis through advanced analysis and robotics; Education - through tailored learning systems and analytics; Low-resource communities - particularly in planning, resourcing and scheduling tasks; and Entertainment - to create more interactive, personalized and engaging customer experiences. Pick your category and then focus.
Don’t go after opportunities that run the risk of commoditization by larger, well-funded incumbents or ultimately represent only a tiny market.
With low-level task-based AI and AI infrastructure becoming increasingly commoditized by large tech companies, it is even more important to take a step back and look for ‘venture scale’ opportunities in different verticals. Digging into the data on AI startups, two verticals with strong representation are fintech and healthcare, both of which stand out from Machine Learning as a Service (MLaaS) and the broader analytics landscape.
Secondly, something that most current AI disruptors (and acquirers) have in common is their focus on enterprise (B2B) applications, which is arguably where the need is strongest and the opportunity is greatest due to the sheer amount of data available to them. What might Applied AI offerings look like across a range of consumer offerings? As a thought exercise, try matching the attributes of different AI techniques to potential market applications and start to sense check these assumptions with potential customers.
To summarise, Applied AI represents a fundamental shift from a cohort of startups focussed on technology push ideas, excited by the promise of AI and ‘deep tech’ to one characterised by market pull i.e. putting the use case first. To boil down some of these opportunities and pitfalls, we thought you’d appreciate an emoji-powered 2x2.