Paper in the digital age (and soon the age of the metaverse but that will be fodder for another post) sounds like an oxymoron, but paper consumption in the United States has increased by 126% over the past 20 years. The average business consumption of paper increases by 22% annually. A significant amount of business is still done on paper (think consumer loans, mortgages, automobile purchases, home sale contracts, invoices, receipts, etc.,) which requires firms to dedicate resources towards ingesting and digitizing the data to fit into downstream digital workflows and systems. As we wait for the promised land of an all-digital paperless world, in the immediate here and now organizations still need to deal with a lot (and a growing amount) of paper. Can technology help streamline and optimize? Is it cost-effective?
Let’s dive in.
Entity Extraction
Technology has come a long way during the last couple of decades for efficiently digitizing paper. While we’re not yet at the nirvana stage, where any document with any print quality can be ingested and produce a perfect output that can be processed into other systems without any human review, providers such as Google Cloud have made significant investments and progress towards building and providing machine learning/artificial intelligence tools that can significantly boost productivity. There are many use cases for digitization, the one I will focus on in this post is entity extraction (also known as key-value pair extraction) because this is the predominant use case within enterprises.
So what does entity extraction mean?
Entity extraction is a text analysis technique to automatically (without explicit coding) extract data from text, classify and organize it. A simple real-life example of manual entity extraction is that of an employee of a mortgage originator (typically a loan processor or loan officer assistant) who receives W2 and 1040 documents from a prospective borrower, extracts key data (such as name, address, income, filing status) from these docs and plugs that data into a loan origination system.
Google Cloud provides sophisticated tools to automatically ingest documents and produce an output that can easily be presented and fed into other applications. For example, some fields and values that can be extracted from the above docs are:
- Employer Identification Number: 12-1234567
- Wages: $50,000
- First Name: Samuel P
- Spouse social security number: 432-19-8765
- Filing Status: Married filing jointly → yes, checkboxes can be detected and values extracted
Here, the Employer Identification Number, Wages, etc., are known as keys, and the 12-1234567, $50,000 data points are known as values hence the term key-value pairs. These are just a couple of examples, Google Cloud can handle just about any structured, semi-structured, or unstructured document and extract content and meaning. While these sample data points have relevance within the context of a home mortgage workflow a few additional workflow examples (and validation rules) are as following:
- Driver License accompanying an application for unemployment: Does the name and address on the license match the name and address on the unemployment application?
- Notary Seal and signature on document: Has this document been notarized within the last 30 days (if the financial institution has an aging requirement for notary signature)
- Automobile Title: Is it a legitimate automobile title? Is there any such thing as a 1965 Honda Pilot?
Google Cloud Platform
Google Cloud Platform provides pre-trained document processing models (for example lending, procurement, utility) where Google Cloud has optimized and tuned the model and also ability for customers to fully define, create, train and operationalize a custom model. Data (lots of it and more the better) is oxygen for any machine learning model. For example, say your business has thousands of signed and executed contracts from customers that you’d like to digitize and extract fields from (contract term, signatory, signatory title, effective date, etc.,). You can define a custom machine learning model and train and tune it for digitizing and extracting fields of interest and store these within a structured location, such as a database. SpringML has built tools and solution accelerators that can be used to rapidly craft custom solutions for clients and decrease go-to-production timelines.
Return On Investment
In many organizations, digitization of paper is often a ‘cost’ so oftentimes a commonly expressed sentiment is “This is all great but what is this going to cost and how quickly can I recover the cost of my investment” In other words show me the money and what is my ROI?
Having been in Corporate IT for more than 20 years and a buyer of technology products and services I can relate to this sentiment. Costs vary for multiple reasons including scope, existing organizational technology infrastructure, resources, awareness, maturity, etc, but below I make an attempt to do a back of the envelope calculation in the spirit of John Maynard Keynes quote “It is better to be roughly right than precisely wrong.”
To be clear, this is not intended to be a proxy for a detailed financial model with parameters that will be unique to each organization (training, onshore/offshore resource mix, licensing cost, hiring, etc.,). Think of this as a means to address the basic feasibility question – does this thing have legs and should I invest more time and resources into fleshing this out? In general, while productivity-enhancing technology such as document processing has a positive ROI, especially at scale, that might not be true for every situation.
Technology Costs
| Item | Qty | Annual Cost | Remarks |
|---|---|---|---|
| Cost for pages scanned | 12,000,000 | $780,000 | 4 Million documents, each document having 3 pages = 12 Million pages. |
| Supporting technology cost | $470,000 | Includes data, storage, networking, machine learning and other costs. | |
| Total | $1,250,000 |
ROI Projection
| Item | Qty | Annual Cost | Remarks |
|---|---|---|---|
| Time to manually digitize each document in mins | 20 | $13,500,000 | 4 Million documents, 20 mins per doc, $10 per hour labor cost. Rounded up to the nearest half million |
| DocAI + manual review in mins | 10 | $6,750,000 | Half the time = half the labor cost |
| ROI Year 2 onwards | $5,500,000 | ROI = Existing manual costs to digitize less ongoing technology costs. ROI for Year 1 will be lower than ROI for Year 2 due to one time implementation costs. |
All assumptions and calculations are entirely hypothetical but nevertheless are loosely anchored to real use cases. This does not represent any price or effort commitments for any products or services from either Google Cloud or SpringML
Boiling it Down
Folks love numbers – but the above basically boils down to the quantitative fact that at scale, it is highly likely that ROI is positive even if one layers in additional costs of training, hiring specialized staff, etc.. Also note that I’ve applied a labor rate of $10 per hour which is extremely conservative (and unlikely in the age of inflation when starting wages at Target are $15). Plug in higher labor cost numbers and the ROI will correspondingly increase.
SpringML’s Value Proposition
You anticipated this right! So here we go
SpringML was founded in 2015 and is a Google Cloud Partner. We have more than 250 implementations. Don’t let the ML in our name throw you off! We have chops in multiple areas and are ready to partner with you on your cloud journey for a variety of use cases – cloud migration, data management, analytics, visualization, machine learning, contact center artificial intelligence and document processing to name a few. Contact us and let’s discuss how we can partner. Your document volumes might be different than the example here but there still might be a positive ROI – we can get creative together.
