Back to Hub

Extinction event: Amazon Textract has just killed the OCR industry. Who’s next and who’s safe?

12/04/2018 By

Spend Matters welcomes this guest post from Doug Hudgeon, a business automation expert.

The annual Amazon AWS Re:Invent conference has just finished. The most interesting announcement in the conference was not one of high-profile changes to their serverless and machine learning platforms. The most interesting announcement was a three-minute video about Textract, a new OCR (optical character recognition) service from Amazon. This service extracts text and tables from documents and is priced at $1.50 per 1,000 pages.

The Amazon Textract OCR service is interesting for three reasons, each of which is worthy of an article in itself, but we’ll briefly look at all three here and the industries that are affected:

  • Textract has killed the OCR industry.
  • The service highlights where the big cloud providers are headed and who is next to go extinct.
  • It demonstrates that big cloud providers can dominate a new industry by using machine learning rather than by acquiring start-ups or established players.

3 Areas of Interest

Textract has killed the OCR industry

The OCR industry exists because companies receive repetitive documents such as invoices, statements and contracts that they need to extract data from. Now that Amazon has entered this market, if you’re in the document data extraction business, your business is now dead. Amazon has just lowered the barrier to entry to almost nothing. Instead of 25 competitors globally, you will soon have thousands. And much of the IP you have built to extract data from documents is now available to anyone at a very low price.

The reason AWS is able to move so decisively into the OCR business is that OCR is a problem that can be solved by applying a single machine learning approach. Let’s look at this in more detail.

Most data in repetitive business documents such as invoices, statements and claim forms is one of two types:

  • Header-level data — data that appears only once on a document, such as an invoice number or a claim number
  • Table data — data that appears multiple times in a document, such as lines on an invoice or statement.

The screenshot below shows Textract in action taking header and table data from a document:

With the right mix of machine learning models, the cloud providers can build a comprehensive solution that handles both of these types of data. And once they’ve built it, they have the infrastructure to allow anyone to use it at a fraction of the cost of current players.

If this is the type of industry you’re in, your business is worthless.

Implications for other industries

At next year’s AWS Re:Invent conference, there will be other industries facing extinction. Two obvious contenders are data/spend analytics and robotic process automation (RPA).

Data analytics companies, of which spend analytics is a subset, take in data from multiple sources and link it together with data from other sources at various levels of aggregation. There’s a lot of complexity to this task, but it is still the type of task that can be solved by machine learning. AWS has started down this path with their QuickSight product. It’s pretty rudimentary at this point but will get better fast — probably in time for Re:Invent 2019.

RPA companies are in the same category as analytics companies. When you take away the parts of RPA software that AWS, Google and Microsoft have already built (such as machine learning, workflow and orchestration) you are left with nothing but a computer vision problem: How do you recognize which button to press or what field to enter data into? RPA is a hot market, and I’d be stunned if the cloud providers are not trying to solve this problem.

If you’re in these industries, sell now.

Implications for start-ups

One of the broader implications of Amazon Textract, but an easy one to overlook, is what it means for start-ups. The vast majority of start-ups exit by acquisition. This means that, instead of becoming a public company by listing on a stock exchange or building a large privately-held business, the start-up is acquired by a bigger company.

In the OCR space, this is a tried-and-true exit strategy: In 2015, Coupa bought Invoice Smash for an undisclosed amount, and in 2018 Xero bought Hubdoc for $70 million.

But the Textract service demonstrates that the big cloud providers, where a problem can be solved through machine learning, don’t need to acquire a start-up or an incumbent to dominate. Machine learning is like a Swiss army knife — the same toolset can be used to tackle different problems. So, once they have the toolset and access to the data, they don’t need to buy a start-up.

If you run an enterprise start-up, make sure your exit strategy still applies in a world with Swiss army knives.

Who’s Safe?

So far, we’ve looked at the types of businesses that are at risk (OCR, analytics and RPA). But who’s safe?

There are three types of businesses that are safe:

  1. Companies with the ambition to become a serious competitor to AWS, Google and Microsoft. We won’t discuss this in any detail other than to note that AWS, Google and Microsoft will eventually wither and die in the same way that the Roman Empire, despite seeming invincible, eventually withered and died.
  2. Companies who incorporate the tools from the cloud providers into their own products and services.
  • While the OCR companies are screwed, the companies that provide accounts payable solutions will thrive because they can deliver solutions to their clients at a much lower cost.
  • While data analytics companies are screwed, the companies that help customers understand the data will thrive because their customers will have access to more data and will need help applying insights to their business.
  • While RPA companies are screwed, the companies that provide effective automation advice and services to their customers will thrive because getting the right mix of automation solutions requires a different approach for each customer.
  1. Companies with functionality that is difficult to unbundle and/or wide networks.

Procure-to-pay companies are reasonably protected from disruption by the big cloud providers because the problems they solve are largely organizational.

Today, for example, a company could buy pretty much everything it needs from Amazon (or a group buying agency) but most of them don’t because they value choice, self-determination, etc. This is not simply a machine learning problem. For Amazon to solve this problem, it needs to change the way companies do business. This will take time.

Doug Hudgeon is the author of a new book from Manning Publications “Machine Learning for Business,” a hands-on guide to build and deploy machine learning applications using Amazon SageMaker.