Are Analytic Applications what’s next in enterprise software?
Professor Diego Klabjan invited me to deliver the inaugural seminar at his Center for Deep Learning at Northwestern University. This is an excerpt.
As someone who has worked in enterprise software for a lot of years, I’m always interested in thinking about what’s next. Sometimes looking in the past is a way to predict the future.
In the 1st era of enterprise workflow software, infrastructure software companies emerged like Microsoft and Oracle, which focused on developers. These software engineers used Microsoft Visual Basic and the Oracle database to build custom workflow applications for the enterprise throughout the 90s. The focus was to automate workflows like order-to-cash, purchase-to-pay, or recruiting-to-hiring.
By the late 90s the 2nd era of enterprise workflow software began with the creation of packaged on-premises enterprise workflow application. Companies emerged including PeopleSoft, Siebel, SAP and Oracle. As a result enterprises didn’t need to hire programmers to develop these workflow applications, they only needed to buy the applications and manage them.
The 3rd era of enterprise workflow software began in the 2000s with the delivery of workflow applications as a cloud service. Examples abound including companies built for the cloud such as Salesforce, Workday and ServiceNow as well as companies, which transformed to deliver their software as a service such as Microsoft and Blackbaud. This era eliminated the need for the enterprise to hire operations people to manage the applications and has accelerated the adoption of packaged workflow applications. While you could hire software developers to write a CRM application, and operations people to manage it, why would you?
Now let’s look at the world of analytics, whose purpose is not to automate a workflow but to use data to discover deeper insights, make predictions, or generate recommendations. Most enterprise analytic software is in the 1st era. By that I mean companies are producing infrastructure technology to enable enterprises to build custom analytic applications. The 1st era of enterprise analytic software assumes the enterprise will hire software developers, data engineers, ML and AI experts along with operations people to manage the applications and data. Companies and products in this area include classic enterprise software like Business Objects, Cognos, Teradata and Oracle along with newer players like Amazon Redshift, Google Big Query, Microsoft SQL Azure, MongoDB, Cloudera, Databricks, Snowflake, R/Studio, Tableau, Anaconda, Pivotal, TensorFlow and Kinesis.
Most of this software is being used to to build “wrong rock analytics”. By that I mean the endless cycle of asking a question, and that question begets another question in each an every enterprise. Each of these result in a custom analytic application much like the custom workflow applications we saw in the 1st era of workflow software. If the past is a predictor of the future, then when will have move beyond custom analytic applications to packaged analytic applications as a cloud service? So what is a packaged analytic application? What are the characteristics?
1. Analytic Applications serve the needs of the worker NOT the software developer.
Enterprise analytic applications focus on workers, not developers. The best example from the consumer world is Google Search. It’s an application focused on the worker, not the developer. Pull back the cover and you’ll see a ton of technology inside. In the enterprise, analytic applications will focus on the needs of the clinician, VP of HR, fraud detection department or the construction site manager.
2. Analytic Applications use historical data
Typical workflow applications throw away historical data. Once the transaction is committed it might be in the audit log in case you have to roll back, but workflow applications are ready to move on. On the other hand if you’re an analytic application you want all the historical data, everything you’ve purchased; or the the behavior of the wind turbine over the past year; or all of the pediatric echo cardiograms from all kids suffering a congenital heart disease over the past five years.
3. Analytic Applications use lots of data
Some of you may be familiar with the famous Jeff Dean graph, which shows with deep learning (neural network) technology you can nearly linear improvement in accuracy with increasing amounts of data and compute over any other methods. If you live in Silicon Valley you can’t avoid the variety of vans and automobiles outfitted with lidars, cameras, radars collecting more and more data to support autonomous driving. And if you’re spending time in the world of imaging and medicine you know that the single biggest obstacle to creating robust AI algorithms is lots and lots of diverse data. At the recent Texas Children’s Hospital conference on AI in Pediatric Radiology both Greg Zaharchuk at Stanford and Ron Summers from the NIH called out the biggest challenge to making progress in AI is large, diverse imaging data sets
4. Analytic Applications use heterogeneous data, inside and outside the enterprise
Early analytics were all focused on getting reports out of the database, which was generated by the workflow application. While these were meaningful as we headed down to more specialized application cloud services (e.g., Salesforce, Workday, Taleo, Successfactors, etc.) it’s more and more important to be able to pull data from a variety of source. Segment, recently acquired by Twillio, is an example of a platform for aggregating marketing data from a variety of sources both inside and outside of the enterprise.
5. Analytic Applications are NOT deterministic
Finally analytic applications are not deterministic. We’ve been much more familiar with workflow applications where purchasing an item on Amazon or a ticket on Stubhub works the same way every time. In the world of analytic applications this is not the case. Instead we’re going to have to learn how to build and tune applications, which have different false positive and false negative rates. In credit card fraud how many times will you declare the transaction fraudulent when it isn’t and how many times will you miss the fraud altogether? How many times will the application declare the patient has pneumonia, when she doesn’t? and how many times will the algorithm miss the diagnosis?
Summary
There are some examples of enterprise analytics applications. Qualtrics serves the marketing worker. Visier serves the HR worker. Yotascale serves the IT worker and Xometry serves the manufacturing engineer. But we are still in the early days. Will the first analytic applications be horizontal (finance, HR, sales, marketing, operations) or will they be vertical? When will we see a fraud detection or pneumonia diagnostic analytic application?
As we have seen on the consumer side it’s likely the enterprise workflow applications will begin to fade into the background and delivering information that is personal and relevant to me, a pediatric neurologist, a shrimp farmer, building manager or cloud operations manager. Just take a look at your Amazon.com home page and you’ll be hard pressed to find the workflow application — it’s the shopping cart. Most banking websites today look like a shopping cart, what could they look like in the future?
