Scotiabank is the third largest bank in Canada. In late 2019 its assets were valued at CAD998 billion. The bank has more than 23 million customers around the world, 89,000 full-time employees, and over 1,000 branches.
Scotiabank’s Treasury department manages the bank’s internal finances, including Central Bank reserves, interbank payments and transfers, intra-departmental financing, and internal projects.
The Treasury department came to Electric Brain with a straightforward problem. Treasury employees were wasting time gathering and consolidating data, an inefficient process that introduced an unnecessary margin of error. But the data they were collecting was required for financial activities. The question was simple: How to automatically free employees to focus on their real job. How to lower the signal-to-noise ratio so they’re undistracted from their core responsibilities?
The general answer is, let machines do what they’re good at, so humans can do what they’re best at. The specific answer follows.
Many employees in Treasury spent most of their workday collecting and assembling data from other departments. The data were from different sources in various formats: spreadsheets with varying structures; email threads with required data in body copy by different writers across time. Given the problem state, our objective was to pull relevant data regardless of format, and organize them for ease of use, so that we could define reliable specifications.
The situation led us to design a 3-part program.
Create the dataset from all sources
We needed a dataset of the problem. Working with Treasury, we pulled emails and spreadsheets from the archives and annotated and tagged them. This time-consuming step gave us a reliable dataset which illustrated the many content types to be extracted from the massive source files.
Write and run the extraction algorithms
A reliable dataset allowed us to create the core data extraction algorithm. We anticipated the need for feature preparation and fine-tuning to achieve the desired results. One algorithm was designed for email content. We used natural language processing techniques, such as word vectors and sliding-window analysis, to analyze the text and isolate transactional data. We achieved 86% accuracy in processing the email content.A second algorithm was written to handle the data in spreadsheets. This required a custom design. We developed preprocessing techniques which are specific to spreadsheets. For instance, the algorithm isolated various data groupings in a spreadsheet, and identified columns and rows that could be used to predict a feature.
Engage our A.I. in supervised learning
The above effort gave us a perfect result in cross-validation. This then allowed us to split the dataset 80/20, a traditional process in data science. This allowed the algorithm to perfectly predict 20% of data from the first 80%. The machine generalized what it learned from the 80% to correctly apply it to the other 20%. So, the system required no specific programming for any of the spreadsheet formats. And the entire extraction process was accurately learned by the machine from the supervised training examples.
Scotiabank Treasury integrated this A.I. into their standard processes. And now employees don’t have to try to be machines. Instead, they’re doing more of what people do best. They’re working on creative ideas and considering the human element in their business. Our A.I. has thus contributed to Treasury’s performance and culture. A.I. doesn’t have to be dramatic to be powerful.
We can do A.I. for you.
A PureFacts Company
400 – 48 Yonge StreetToronto CanadaM5E 1G6
647 884 email@example.com