You are a digital business that verifies consumer identities, you have just finished integrating an ID Verification vendor, time to flip the switch and go live. You give it a few days and notice a high number of customers cannot go through the verification process. Consumers are predominantly good actors, so this is frustrating. You cannot get them through the verification funnel and are bleeding business daily. Digging through your data you see the customers have been rejected because the document is not supported by the vendor. This document type was issued only recently by the government authority. You go back to the vendor and ask, how soon before the new document type is supported? Days? Weeks? Months?
Document forensics on government issued IDs is the cornerstone of an ID Verification solution. IDs are verified for authenticity, either by using automatic techniques (Auto) relying on Computer Vision and Machine Learning technologies or an army of manual teams (Manual) with varying levels of expertise, that are typically setup offshore.
If you are a business seeking a solution to verify consumer identities, you probably ask the following questions of an ID verification vendor:
- How many documents does your solution support?
- How does the vendor count the number of documents that have been issued around the world? There is no canonical standard for documents issued, not even U.S.
- There may be negligible or no difference between two issued IDs (simply a reissue). Determine if the vendor is double counting.
- What is the coverage in various geographies?
- Once you get past the breadth, consider the depth of support for a document. How effective is the company at finding IDs that have been tampered (physical and digital forgeries)? Be sure to take multiple scenarios into account – for example, photo, text, signature, background.
- What are your processing times and the associated SLAs (Service Level Agreement)? The solution could be Manual, Auto, or hybrid which means that Auto failures waterfall over to Manual.
While all of the above questions are important to ask your ID verification vendor, there is one question is often asked too late, or never: What is rate at which new documents are onboarded or what is the document onboarding velocity?
This is because there is a tremendous amount of churn where older documents go out of circulation, and newer documents are issued by nations and states constantly. The churn on an average is between 20%-30% per year and when a new law like the REAL ID act is passed, there is a sudden influx of documents put into circulation.
With this constant churn and sudden influx, some vendors react and adapt more quickly than others, because of many reasons:
- Some solutions have a poor system design. If a solution is designed well, the onboarding of documents should be strictly “content” update as opposed to a “code” update. Here “content” means that any parameters, models, data generated that is specific to the new ID type (issue) is kept independent of the code. Therefore, the code can deal with new ID type being onboarded in an abstract fashion. Inferior solutions have content that is dependent on the new ID type intertwined with the code, and because of this very tight coupling, each new ID type must be treated as a special case with customized code. In these bad designs, one must go through a code release cycle to provide support for newer ID types, which can be a lot more time consuming.
- Deep Learning solutions are now ubiquitous; however, naive implementations are data hungry. Therefore, if a solution that onboards a new document must train or retrain models that require many hundreds or perhaps thousands of examples of a new IDs in circulation, it is a huge challenge. This is because the IDs have PII (Personally Identifiable Information) and the vendors have strict contracts with the businesses like you, regarding retention. It is extremely hard to quickly harvest large amounts of data and even if possible, there must be infrastructure to quickly label the data, which itself is a large challenge when dealing with PII. Algorithms must be more sophisticated, using techniques such as few-shot learning (learning from few examples), generative networks or hybrid classical computer vison-deep learning methods.
- One other approach vendor might take is to send unsupported documents to manual review until there is support for automatic processing. While there are issues with SLAs with Manual, one must also ask the questions such as where is my data going geographically? Is it vulnerable to leakage on getting there? Is the data provided by your business used to train a model that will now benefit your competitor?
- Vendors that already serve diverse geographies (not just US and a few EU regions) tend to have solutions that are more sophisticated. Typically, one runs into many technical challenges in onboarding a new geography, for example the vendor may have to deal with paper documents (that are not rigid), or deal with lower case letters on the id, or unusual font types. The versatility helps the vendor adapt to a new ID type quickly, as they have seen it before.
As a business seeking an ID Verification solution the one question you should ask the vendor is “What is your document onboarding velocity?” It is important to dig deeper into this issue. Do you have a system design that allows you to onboard new documents as “content” update independent of “code” update? Also, do you have an algorithm that can bootstrap with a small number of ID images? Are you relying on manual teams to onboard unsupported documents, if so, what are the SLAs, and are there controls offshore to avoid leakage? Lastly, if the vendor relies on training on your production data, and if given permission to use it, will the models be used exclusively for your benefit, or is there a preferred pricing you can negotiate if it would benefit all the other customers? Asking all these questions will get you your answer to the question of how quickly a newly-issued document is supported.