The responses to our “The AI Peak Hype” post have been both polarized and engaged! We’ve had a number of requests for more detail and real-life examples; here, we’ll do just that.
To recap the previous post and provide the setting for the next phase of commentary:
- The current technology for AI isn’t there yet.
- Current “breakthroughs” in AI are operating in very specific domains.
- Even if progress exceeds expectations, there is no way it can live up to the hype.
What should an AI entrepreneur do?
- Raise fast – hype cycles tend to crash fast too
- Have a real business
- Utilize proven technologies (e.g. conventional neural networks and LSTM)
What should an AI investor do?
- Focus on the next-gen of AI technologies (specifically, reinforcement learning & adversarial networks)
- Avoid AI solutions looking for a problem
- Stay domain specific
The overarching theme from our first piece is that domain specificity is key in understanding current strengths in AI and how to think about potential commercial applications that can develop into viable business models. More overarching approaches are still very much best left to academic and research disciplines. Companies and investors need to keep their head focused on very specific domain applications that have the ability to add value and take advantage of well understood AI dynamics. To that end, we’ve focused this follow on post to a very specific domain and sub-domain applications of AI and dissected what’s on the cutting edge, what’s on the “bleeding edge” and what remains pure hype.
In looking at specific domains where AI shows promise, one feature to consider is the extent to which the industry has invested in technology to drive productivity gains and the level of which the industry has become digitized – meaning specifically that activity has shifted towards digitized formats and transactions handled by computers. This creates data, oftentimes reams of it, which gives AI an opportunity to come in, “eat the data” and automate decisions or do things more efficiently or intelligently.
The financial services sector is an area which has consistently been one of the biggest investors in applied technology spending. Most trading in the world has shifted to the electronic realm and could no longer take place without computers/digitized data and investments in machine-driven trading processes have accelerated over the past decade – a portion of which was highlighted in Michael Lewis’ book Flashboys which recounted the world of high-frequency trading. Trading is an area where vast amounts of data, both current and historical, are accessible to prime AI learning algorithms. In addition to trading, numerous aspects of the financial sphere have become increasingly data and algorithm dependent, from insurance to lending to traditional corporate finance. And, separate from customer-facing/trading situations, financial institutions have also invested in technology to handle increasing amounts of internal processes and functions, which provides further opportunity for AI-driven applications.
We’ve chosen four areas of the Financial Services sector for examples. In each, we’ll highlight some leading AI company in the subsector, discuss what we think the technology can achieve now and in the near-term, and what we feel is further out.
Subsector 1: Insurance
The cutting-edge: Chatbot-powered user interfaces. For example, Lemonade, the NY-based insurance start-up, has raised a huge war chest and is using tech to better the slow world of insurance. One of its key innovations is the one-line chatbot interface for handling everything from customer service to claims. This is in place and working today.
The bleeding edge: Actual AI-led chatbots. Most of that interface at Lemonade still has a person behind it. However, natural language processing, powered by convolutional neural networks, continues to become more effective. We think it’s safe to say scripted chatbots will be leading insurance interfaces within months or years.
The hype: Image recognition for claims handling. Some companies are testing this now. For example, multinational insurer Agaes is using technology developed by fintech startup Tractable. The insurer is leveraging technology that seeks to analyze images of cars which have been involved in accidents and make a decision whether a claim is valid, typically within seconds. There’s a lot of optimism that just snapping an image will allow instant confirmation of claims. But this is a very high level of abstraction. It will be years for image recognition to advance from distinguishing chihuahuas/muffins to actually making high-level (and high-value) judgment calls.
Subsector 2: Asset Management
The cutting-edge: Alternative data. Smart companies are using big data technology to consolidate, process and ultimately supply data that professional investors use to make decisions.
The bleeding edge: AI powered stock-picking. It’s no secret that the machines are eating Wall Street and investing better, faster and a greater scale than traditional traders. New York’s own Numer.ai is a tech-take on quant investing, creating a crowd-sourced, AI-powered hedge fund. Generally speaking, some of the most common techniques used by AI-powered funds are multivariate regression and supervised machine learning. So far, it’s unclear if quant funds are over or underperforming vs. traditional funds, but the trend is strongly towards AI-powered investing (and soon).
The hype: Generalized anything. AI stock picking works well because AIs perform better on narrow domains. For instance, it’s (without hyperbole) millions of times easier for Facebook to correctly identify a photo if it only needs to choose from your several-hundred friends. Stock picking is a narrow domain by itself – developing cross asset macro hypothesis about the world is not. So while machines are already beating humans at certain narrow tasks, we’ll need some general intuition running the bots for a long while.
Subsector 3: Lending
The cutting-edge: Predictive credit scoring. There’s no shortage of companies using AI to provide an improved version of the credit score for both individuals and businesses. Some examples include Zest Finance, Underwrite.ai and Demyst Data. This is actually quite hard to do (the technical challenge is mastering the design of the very subtle features that predict credit risk), but in our opinion AI solutions are now doing a more accurate job than traditional credit scoring techniques. Expect to see much more of it in the very near future.
The bleeding edge: Expanding beyond scoring. Some of the world’s best consumer lenders are using AI to decide what action is most cost-effective if you miss a payment. Even make walking maps of the
The hype: A human-free loan process. We see isolated examples of this, and a lot of boasting that it will be happening soon. But in practice, there’s too much at stake to relinquish the human underwriters anytime soon.
Subsector 4: Corporate Finance
The cutting-edge: Natural-language processing of financial data. This essentially enables users to ask questions in plain English that can establish connections, or correlations (obvious or obscure), that can influence financial markets. For example, “When Netflix beats earnings, how do Amazon shares perform the next day“, or “Which Apple supplier’s share price goes up the most when the company releases a new iPad?” (answer: OmniVision – which makes the sensor in the iPad camera). Kensho, a Cambridge-based AI startup with a long list of top-tier investors including Goldman Sachs, NEA, Google and Accel Ventures, has staked a leading role in this field. It sells itself as giving “the masses” access to the kind of computing power heretofore reserved for the likes of top hedge funds like Bridgewater Associates and Renaissance Technologies.
The bleeding edge: This is the perfect example of a Centaur at work – the combination of human mind and machine analytical firepower, also called “AI-human hybrid” intelligence. As companies like Kensho continue to increase in efficacy, we’ll see a number of the typically time-consuming analytical roles be automated and enable greater productivity.
The hype: The deal machine. There’s been some rumbling about AI being able to identify good M&A targets, etc. This is admittedly closer than the other “hype” sections in this article, but we’re still talking broad domain, generalized AI to achieve things like this with anything close to human-level precision.
Let’s not lose sight of what AI is – it’s really just statistics. The principles being applied aren’t that much more advanced than what you learned in STATS 101; they’re just applied iteratively and at scale.
Are there some groundbreaking engineering achievements being made? Absolutely. But it will be quite some time before we figure out how to assemble a bunch of linear regressions into stronger forms of intelligence. And then there is the issue of whether or not machines can achieve “consciousness” … we’ll leave that to another day.
Co-authored with Ryan Janssen. For more information contact us at firstname.lastname@example.org.