These are excellent questions. I've conversed with Appen's Investor Relations team on this topic and summarised some thoughts below for you.
"1. The medium to long term necessity for large volumes of annotated data required by major AI players (Google/FB/Microsoft/Tesla, Amazon, etc)? Ie, how far off do you feel we are from seeing developments in automated annotation supersede Appen's services and impact revenues in a meaningful way? Also, developments in AI and ML such as Auto ML-Zero are rapidly paving the way for development of learning algorithms without the need or use of significant data sets - do you see this being a factor that may impact your revenue forecasts over the medium to longer term (3-5 years)."
-- Firstly, it is important that we as investors distinguish between Appen's services and which ones are potentially irrelevant/relevant to AI/ML developments. Large search and social media companies rely on the subjective human feedback provided by Appen's global and diverse crowd to improve their core algorithms related to ad placement and core product features. For these applications with search and social media companies, the developments in AI and machine learning that you refer to are in my view not encroaching on Appen's territory. This is because the relevance data Appen provides is an important input for the product engineering teams to ensure that features are unbiased, reflective of the current global and local context, and perform to customer expectations. "There is a direct link between Appen's relevance data and the core business model of Appen's customers, and a persistent need as end-users and context evolves. Appen sees a pathway for growth as the digital ad market matures and competition for improved ad and product performance intensifies."
In terms of image recognition for autonomous vehicles, speech interface systems etc. as you correctly point out there are definitely developments in this space exploring how to build AI with less labelled data. Appen's stance is that, at this stage, the quality of models that do not use labelled data typically lags models built using high-quality training data, hence why we don’t see large scale adoption of these methods. It is also important to note that Appen itself is using AI in the human data labelling process, as that has the dual benefit of lower cost and higher quality. "By lowering Appen's per-unit economics and improving our quality we expect that the current deep-learning models that require training data will remain attractive for many applications."
Lastly, It’s worth noting that some of the AI/ML developments around Auto ML are likely to have a positive impact on the demand for labelled data. Auto ML is a technology that selects the best algorithm and configuration for a given AI application, rather than providing the ability to create applications with less training data. These types of technologies democratise AI development – meaning that you no longer require a data science background to build models. Access to AI talent is a current bottleneck in the industry. Removing this will likely create a greater need for training data as there will be many more companies with AI models in production.
"2. If you compare Appen, especially to a business like Xero or Amazon or Microsoft, (all which have exceptionally low marginal costs for each customer added), do you feel that there's anything the business can or should be trying to do (or is currently doing) to reduce it's cost base (average marginal cost) per project? I'm not suggesting Appen's business model is bad by any means (I'm long after all), but to me, it's always seemed like more of a niche recruitment operation than a true "tech" business. Keen to get your comments/thoughts around that."
This is a really interesting question. In my eyes, there are a number of different reasons why companies like Xero trade on a significantly higher revenue multiple than Appen does (e.g. 20x sales versus 5x sales). The top two factors in my view would be long term margin potential and the recurring (or otherwise) nature of revenue.
Xero notably has a world-class revenue retention rate. However, although it goes unnoticed, APX also has a surprisingly high level of revenue retention and this is shown in the customer cohort analysis. It is not "technically" a majority-SaaS business, but it acts like one and I personally believe this is not priced into the valuation. The reason it acts like a SaaS business is that training data must be refreshed regularly or customers run the risk of 'model drift'. If APX was valued like a true SaaS company on multiples north of 10x sales, the share price would be north of $50 per share right now. I think we will get there in time.
In terms of the unit cost per project, I think Appen is making many moves in this space to increase scalability and reduce unit cost. This includes using AI in their data labelling process to increase the productivity of their crowd (volume and quality) and also applying AI to Appen's internal operations to deliver large scale high-quality data faster. Appen is very much a tech company at heart. I don't personally view it as a niche recruitment operation. Examples of tech within APX include leveraging AI to improve crowd recruitment and automatically matching crowd workers with tasks best suited to their capabilities. To achieve these benefits Appen is investing in engineering and their technology, including Appen's crowd recruitment and management platform and their data annotation platform.
T.E.P.
APX Price at posting:
$23.46 Sentiment: Buy Disclosure: Held
(20min delay)