The concept of artificial intelligence has been around for a long time. According to Wikipedia, the field of artificial intelligence research was founded as an academic discipline in 1956. The massive increase in the amount of data available since the birth of the internet and growth of sensors has created the opportunity to apply AI to a wider range of opportunities such as self driving cars, enabling virtual assistance like Apple's Siri and helping businesses operate more efficiently. It's the amount or data, increase in processing power, new math models and drive for new businesses and/or efficiencies that makes AI so compelling.
This is important to finance & accounting professionals because technological advancements will enable the financial tech stack of the future to be created and you need to be preparing now.
Artificial intelligence is the broader category of using computer science (eg. math and processing power) to receive input and make decisions. Encyclopia Brittanica (yes, it still exists - online) defines Artificial intelligence (AI) as the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. Intelligence beings have the ability to learn and adapt.
Machine Learning is a sub-set of Artificial Intelligence that allows the computer to learn from a distinct set of data inputs without being explicitly programmed. A system that uses boolean rules (if-then statements, etc.) requires programming to create the categories of data and rules to execute against while a machine learning can fit an algorithm to the data and then use that algorithm to categorize and organize new data inputs automatically.
Deep learning is an evolution of neural network models (think neurons in the human brain) to learn using larger networks with more layers - hence the name. This requires more data, processing power and advanced modeling for many machine learning applications. Deep learning typically excels in applications where there is abundant data and processing power.
This is important because Machine Learning is the component of AI that is creating real improvement opportunities in the financial function NOW.
Most of the focus of applying advanced technology in the F&A organizations is on machine learning given the types of data sets this group works with. Specifically, structured data like general ledgers, structured accounts and labelled transactions.
The purpose of machine learning is to develop one or more mathematical formulae (algorithms/models) which can leverage historical data to interpret data sets - could be predictions, recommendations or clustering. These models themselves don’t typically change unless a machine learning engineer pushes a new model, however the outputs of the model will change as the data inputs change.
For example, your Netflix recommendations may be different throughout the day. Even if they have not updated their recommender model, many of the model inputs have changed: their catalog of shows, your behavior (you’ve watched something), or other people like you have changed (they’ve watched different things). While the model hasn’t learned anything new, its knowledge is nuanced enough to give appropriate recommendations as the world around it changes. This is ideal for financial data like matching reconciliations.
The output from a machine learning system generally falls into one of two categories: classificationwhich predicts a discrete set of class labels, often with a probability for each (i.e. “this flower is a rose”, “there is a 57% chance of rain this weekend”, “these two transactions are a match”) or regressionwhich predicts a continuous quantity (i.e. "this customer will spend $1,200 next month", “we predict 2.4” of rain next month“).
Over time, models may be updated based on new and changing patterns as external data systems evolve. This can result in flexible and low-maintenance systems which require little care and feeding. A well built machine learning system has the ability to effectively adjust and updated to changing conditions, much like you or I, and in this way can help augment many manual, human tasks.
It is important to understand that machine learning models don’t magically update themselves, but respond to different data inputs. This means that your models need to be updated as the nature of the data changes on a schedule that aligns with these changes. In the account reconciliation process for example, if the structure and type of data you get from the two accounts being reconciled don’t change over time, and the model is accurate at matching then model updates are not necessary. If you change one or both of the sources of the data (new bank, different credit card gateway, etc.) then the model should be updated.
To give this discussion context, imagine, as a fun example, that your middle daughter is a junior in high school in the Computer Science club and wants to use machine learning to predict the color of the sky for any minute and day of the year. This is a simple example that could be solved with basic statistics, but helps illustrate the approach. One of the first decisions she will have to make will be whether she will be using a supervisedor an unsupervised approach.
Supervised learning is learning by example. A machine learning algorithm is provided “ground truth” data - data where the thing you are trying to predict is known in some form - like for a certain period of time. A model learning from that data can then be used to make predictions when the answer is not known - like for a future period of time. The fundamental problem in supervised learning is to learn a model that generalizes to new data (is not “over-fit” to the examples provided).
This initial set of data that a model learns from is often referred to as a “training set.” A good training dataset is a representative sample of historical examples and includes both the original raw input values, and the corresponding output value (the correct answer) for each set of examples.
Structured data, such as numbers, dates, and strings, can be represented by rows and columns, unstructured data cannot. Examples of unstructured data include images, audio, videos, e-mails,, and word processing documents—essentially, things stored as files. Unstructured data tends to be much larger and take up more storage than structured data. Both structured and un-structured data needs to be interpreted and cleaned by the machine learning process.
In our sky-color example, our student might decide that a supervised approach seems appropriate, and collect structured data over the month of March that includes the color of the sky for each day and minute (perhaps over a random sample throughout the month). She further decides to limit the number of predictions to four colors: blue, orange, purple, and black. This set of observations would be her training set. This is a very simple example with only one set of “features” - the colors. Machine learning models perform better with more, different features. Every column of data in a reconciliation file is treated as a unique feature to be evaluated.
Once the system “learns” from the training set, it can make predictions about new, unseen data, based on the historical examples it has been provided. It may continue to improve over time as each new prediction is monitored by something or someone (a human) and corrections are signaled to the system. For example, a history of incorrect predictions may be incorporated into an additional “lesson” for the model’s algorithms. Imagine, for instance, what kind of predictions we might expect if our sky-color predictions had only been taken at midnight (or noon) of every day; a new, updated “lesson” might be in order. There is also an issue or danger with using too narrow of a data set that could cause bias in the output. In this example, will the model work in months if we only sample March or only in one location? The training data needs to be as un-biased as possible given the desired output.
Unsupervised learning differs from supervised learning in that there is no “ground truth” to supervise the model. Instead, unsupervised learning is about finding (often hidden) structure in data. Customer segmentation - grouping customers into natural clusters based on common behavior - is a good example of unsupervised learning.
It is important to understand that finance and accounting professionals primarily live in the supervised learning universe due to the highly structured “ground-truth” data available via ERP and other systems.
A recent survey of financial professionals by the AICPA and Oracle described opportunities for F&A teams to harness technology to dramatically improve business operations and generate positive revenue growth in three ways:
A.I. and M.L. have had significant impacts in many industries and are becoming more mainstream in their applications. It is time for finance and accounting teams to invest consideration cycles in this incredibly powerful technology to move the F&A organization into the next level of efficiency and driving business performance.
We hope this white paper has helped you understand more about the technology and how it can impact your organization and business. Contact us at firstname.lastname@example.org if you would like to learn more about our AI-powered, enterprise-strength matching reconciliation engine.
“Artificial Intelligence”, Encyclopedia Britannica, https://www.britannica.com/technology/artificial-intelligence
“Cousins of Artificial Intelligence”, Seema Singh, Towards Data Science - https://towardsdatascience.com/cousins-of-artificial-intelligence-dda4edc27b55
“Machine Learning: An Introduction”, Sujeet Kumar Jaiswal, https://blog.sujeetjaiswal.com/machine-learning-an-introduction-de88d85ebc5d
“Agile Finance Unleashed: The Key Traits of Digital Finance Leaders”, Association of International Certified Professional Accountants and Oracle. https://www.oracle.com/corporate/pressrelease/agile-finance-study-011719.html