Finance for AI : 2019 Year in Review

Review of 2019 in Finance for AI

2019 has been a really interesting year to be a finance manager in AI. The industry continues to change and innovate at a rapid pace, but the practical implementations have not seemed to reach a popular boiling-point. There is still no flashy consumer-ready iPhone-esque implementation of AI. However, there are certainly tons of behind-the-scenes machine learning implementations gradually improvements our services and impacting our lives. In Finance, this has looked like a lot of seed investment across the industry which I believe will soon start to sprout into large consumer visible impacts in 2020. Here is my recap of the AI innovation in 2019 from a novice perspective, and some of the key financial implications which may result in 2020.

2019 AI Trends

What: in 2019, all of the large players in language models and AI began to develop and release really big AI models based on a new style of model called the transformer. It is based on a Google model called BERT (Bidirectional Encoder Representations from Transformers). A technical review is here, but the summary is that this a very popular new method to allow huge amounts of data (350M parameters) to be used to train an AI model. Similar transformer models including Elmo, RoBERTa, and GPT-2 have also been developed.

To what end: These models have provided breakthrough results in language as measured by scores on key industry benchmark tests (ex. GLUE). More importantly, computer performance on language understanding tasks has jumped to near-human parity (language understanding, Chat-bot reply, and reading comprehension). The technology underlying chat-bots and text based AI models have made huge progress in 2019. This will continue as the newly improved tech will be integrated into more and more apps to simplify conversations and understanding between bots and agents.

Impact on 2020: there will be a large increase in the number of auto-”you-name it” capabilities. Where emails, documents, press releases, AP articles, twitter news breaks will be automatically generated by these large models (BERT, GPT-2). Additionally, capabilities of chatbot and reply agents will continue to augment our daily interactions: ex. Talking to more realistic bots in call centers.

What: The future use cases for vision systems are infinite as this is the key cognitive ability of humans and is the core of much of our interaction. Vision AI is not new (US post office has been using character recognition to sort mail since the 60s), but deep learning enabled capabilities are starting to bubble up from the lab into more and more enterprise use cases. A lot of this surge in development has been enabled by investments in tech hardware, software and models for autonomous cars, but the use cases outside this space are equally interesting.

To what end: Use cases taking vision from cool research to user implementation have begun to spin up in a big way in 2019. Some in-market examples:

Impact on 2020: We are still in the earliest stages of vision system integration in enterprise. As systems like camera-enabled lane-assist blend into the background of most new cars, new vision implementations will continue to diffuse into all areas of retail and manufacturing, producing huge wave of productivity gains.

Finance Impacts and Trends

While that is a novice summary of the AI industry in 2019, I am not particularly qualified to talk the tech. I am, however, qualified to detail the really interesting impacts I believe this tech has from a finance perspective.

Progress in AI based language and vision are key enablers to automating repetitive and non-value tasks across all sectors, with the goal of improved efficiency and enabling more high-impact work. With the ability to recognize text, objects, and basic conversational logic the number of tasks that can be taken on by a robot increase dramatically. I see this less as replacement and more as augmentation and believe the potential for productivity improvement is a huge financial opportunity. For example, with little to no capital investment, organizations can automate huge manual processes:

  • Automated invoicing systems — based on computer vision pictures of paper invoices converted to digital copies. An AI combs through invoices, turns them into digital records (like phone based check deposit), then kicks off a workflow to pay it based on the due date listed in the file.
  • Automated call center — calls with humans are transcribed using Speech recognition and records are saved and key phrases are extracted using Text analytics. These key phases label the call as positive, negative, or group them into certain call type categories. That record gets auto-saved with the name of the caller in your CRM system. A live dashboard of sentiment is created.

These large transformer models and new vision techniques share one thing in common: they require huge amounts of data and compute to run. Data and compute are the most expensive part of AI and the rate of growth in data availability and compute required are both growing exponentially. If the formula for a good AI = data + compute + people, the cost of compute and data are a 3:1 ratio on people despite the high cost of machine learning PHDs.

Compute is the king of cost here. Large AI models today run primarily on hardware called Graphical Processing Units (GPUs), similar to the high end chips running a gaming console or PC. The overall cost of an individual AI model can be very big to start. However, with the new transformer based architectures, the size of models is doubling every 3.4 months. Because the cost of GPUs scales linearly with model size, the cost of developing AI becomes very unsustainable very quickly.

ROI Focus:

Cost increases were a major feature of 2018 and 2019 AI space and large companies have highlighted cost as a growing concern. With exponential growth in input cost, the time has come to have a more meaningful discussion on ROI. In this space it is important not to take a traditional view of ROI as the payouts will not often be near-term and tangible (Does it pay out to make an AI dungeon-master in a game of D&D 10% more lifelike? Maybe). Therefore, implementing a down-stream impact or strategic ROI framework is important for investment decisions.

While costs are currently growing at a growing rate, there is some hope of cost reduction in the form of improved hardware and software efficiency. The drivers of these efficiencies are primarily enabled by AI techniques applied to the process of AI cost efficiency. One great example of using AI to make AI cheaper is by driving efficiency through transfer learning. Transfer learning is a process to develop AI models using less labeled (expensive, human labeled) data and use the output of one AI model to train a new AI model.

An oversimplified explanation: AI models typically learn from seeing many examples of A or B labeled by humans. After seeing enough human labeled pictures of a cat, the machine recognizes cat. In transfer learning, the model trained to recognize cats is used to teach a model which recognizes lions. The industry has found that the AI systems generalize pretty well and the new model requires much less expensive labeled data (and compute) to get up to speed. Therefore, after building an expensive model to translate Spanish to English, the same model trained on a lot of random Portuguese words (plus some labeled words) provides a cheaper starting point with decent accuracy.


I am excited for AI Finance in 2020 as I believe a lot of the leg work and seed investments in the last few years will turn into impactful implementations for end users. Watch out for cool implementations of transformer models (GPT2 + “ you name it”), big growth in RPA implementations, and computer vision integration with manufacturing. As a new years resolution, I will try to keep updating this blog with interesting learning, industry trends, and AI Finance topics throughout 2020.

Miami University Alum. Microsoft - Finance & Accounting.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store