Should Big Capture Operations Use AI Engines like Amazon Textract for Intelligent Capture?

Most of our large financial services clients are trying to introduce intelligent capture into their high volume capture operations. They want to deploy AI and machine learning-based capabilities for document classification, data extraction, and validation. Many of them wonder whether they should deploy relatively “raw” engines and tools like Amazon Textract, Google's Document AI, or similar AI tools rather than implementing potentially more expensive products and services provided by the more traditional capture vendors. Our answer is typically “no.” If you have a large capture operation and want to get intelligent capture into production quickly for anything other than simple workstreams, then you should not deploy Amazon. This post explains why.

Five Kinds of Intelligent Capture Solutions

First, Amazon Textract belongs to one of five intelligent capture categories, and each has its strength and weakness. They are:

  1. Mature “platform” vendors: Provide everything related to capture, including scan/input, workflow, and comprehensive reporting. Primarily consists of Kofax, IBM, and OpenText.
  2. Mature intelligent capture “solution” vendors with innovative offerings: Provides mature and next gen capabilities; some workflow and reporting, inherent or integrated IPA. Includes vendors like ABBYY, Adlib, Mitek/A2ia, and Parascript.
  3. New intelligent capture “solution” vendors with innovative offerings: Provides next gen capabilities, some workflow and reporting, inherent or integrated IPA. Includes AntWorks, Ephesoft, HyperScience, and Vidado.
  4. New “engine/tool” vendors with innovative offerings: Consists of AI generalists who have a focus on intelligent capture and provide individual engines or tools that can be developed or OEMed. This is the category that includes Amazon, Google, and many smaller AI engine vendors
  5. Vertical industry or application specialists: Consists of software or service providers with narrow focus on industries or applications, but no significant footprint innovative capabilities. Includes Conexiom, Informed and Glynt. Note that these specialists sometimes embed the “engine/tool” products.

Amazon Textract, Google, and other AI engines belong to the “engine/tool” category. To evaluate them, we first must define a set of assessment criteria.

Criteria for Evaluating Intelligent Capture Offerings

To evaluate any intelligent capture product or service we judge it against the following criteria:

  • Vendor criteria: Health, stability, footprint in financial services, partner relationships, and professional services.
  • Capabilities criteria: Breadth, depth, innovation, and performance in relevant functionality. There are two types of IC capabilities:
    • Primary capabilities: classification, extraction, validation, and conversion.
    • Secondary capabilities: IPA, process orchestration, reporting, and administration.
  • Deployment Readiness criteria: Time, effort, resources, cost, and risk to get solution into stable production in your organization’s ecosystem.

Now let’s use these criteria to evaluate whether the tool/engine products offered by Amazon and similar vendors are a good fit for large financial services capture operations compared to the alternatives.

Our Assessment of Engine/Tool Products Compared to the Alternatives

  1. Regarding vendor criteria, engine/tool products don’t have a significant presence in the large financial services vertical, or a reliable solution provider channel for that vertical.

Amazon and Google are healthy stable companies, but lack a focused customer footprint or professional services and partner focus in financial services.

Those financial services firms that Doculabs tracks who have worked with these tools have not grown them beyond POCs, “innovation teams,” or standalone workstreams. None have adopted them for production environments with large scale, complex processes, let alone for enterprise use across a wide array of workstreams. By contrast, many of the vendors in the other categories have significant numbers of customers in financial services with large, complex ingestion workstreams.

The new engines/tools are OEMed and embedded within ECM or BPM platforms (e.g. Alfresco, Nuxeo), and within intelligent capture suites (offered by vendors in all the other approaches) – but as advanced capabilities within software product offerings requiring significant preparation and configuration to be ready for deployment.

  1. The engine/tool category fails at breadth and depth of relevant primary capabilities.

Complex workstreams with a high variety of document types may require more than one intelligent capture approach and engine.

They require multiple approaches and modules, including OCR, computer vision, NLP (natural language processing), and others to provide adequate classification, extraction, and validation. They often require multiple modules to effectively extract data from different document types.

In the Amazon portfolio this means not just Textract, but also possibly Comprehend, Rekognition, and others. Some of the vendors in categories one, two and three combine their engines and are improving the efficiency with which they can be used together. Few intelligent capture developers are doing the same with the Amazon or other engines in the fourth approach.

  1. They also fail at breadth and depth of relevant secondary capabilities.

The secondary intelligent capture capabilities include RPA, and an intelligent capture “backbone” that includes process orchestration, reporting and administration. Most of these are likely required for the auto loan capture processes and are almost certainly required by many of the other enterprise processes. The engine vendors lack most of these capabilities, thus requiring the proprietary incorporation of such components into the overall solution.

By contrast, most of the best vendors in categories one, two and three include vendors who either natively provide the secondary capabilities or have strategic and technical integrations with partners that do.

  1. Finally, they fail at deployment readiness and cost-risk performance.

Given the above factors – few similar capture operations in production to reduce R&D cost and risk; fewer competent, available solution providers; little productization; and solution incompleteness for both primary and secondary capabilities – we find category four, including Amazon, to be a poor solution for the large scale, complex capture requirements of large financial services operations. By contrast, the best solutions from categories one, two and three meet most, if not all, of these requirements.

Hard quantitative comparisons are difficult because every implementation we’ve seen has been at small scale. In POCs Amazon and similar vendors have done reasonably well in simple, small tests – but not in cases with significant variety and production volumes.

Conclusion

Large capture operations – particularly those belong to financial services enterprises – have unique needs that are not one-size-fits-all. If your organization is looking to increase the use of intelligent capture, our team can help. Reach out and we’ll schedule an exploratory conversation.

How to Select Enterprise Software

Rich Medina
Rich Medina
I’m a Principal Consultant and co-founder of Doculabs, and the resident expert in using ECM for information lifecycle management.