It may sound like the classic consumer warning of “caveat emptor” but it seems like a lot of enterprises that should know better are getting swept up in the dizziness associated with machine learning. Hey, cars are about to drive themselves so shouldn’t everything have 100% automation at the highest levels of precision? Not so fast!
Marketers love to stretch the boundaries. Have you seen those commercials showing drivers taking their hands off the wheel letting the car change lanes? It’s quite dramatic stuff, but the tech isn’t approved for anything other than as a driver assistant, yet people will undoubtedly start using it in other ways.
I was reminded of this disconnect in an email exchange with an industry analyst the other day. This unbridled enthusiasm in marketing is the same for enterprise software (yet thankfully not as deadly). At issue is the belief that acquiring software that includes machine learning will allow operations to sleep at the wheel so to speak, with no need to pay attention to how systems are trained and configured and no need to monitor results. After all machine learning can take the wheel while data analytics and dashboards provide assurance that the system is delivering the goods.
When it comes to document automation, words like “out of the box” and “pre-trained” elicit visions of easy-button automation where documents go in one end and reliable, accurate, structured data pops out the other. But that is far from reality. Most of these pre-built skills for various documents is the equivalent of a kit that has a lot of things defined but isn’t anywhere production-ready. Using one in production “out of the box” will result in a lot of exceptions.
That’s not to say that offering pre-configured skills for various documents like invoices, remittances, and forms has no value – quite the contrary. A lot of tedious work goes into document automation projects just to define what data you’d like to get out of a particular type of document, such as an EOB. I consider these “functional configurations” in that they describe all of the work that needs to be automated, but it isn’t optimized.
And if the idea is to use machine learning, because there is no magic, the more input information about targeted data you can provide, the better the results. So anything that offers a configuration that provides as much of that “priming context” as possible can shave hours if not days off a project. Just be sure that you know what you are getting and are prepared to put in the effort to move from a functional configuration to one that is highly optimized.
Just don’t think that a pre-trained skill will work for you at high levels of automation and precision. And if you need something not available in the skill, you still have some work to do.
Ultimately reality will set in and organizations will realize there is still no silver bullet, even with pre-built or “out of the box” skills. Still, having a leg-up in the form of a pre-built configuration provides value, and if it is coupled with the ability to optimize it through training, that represents almost a step function in terms of optimized automation without the traditional exorbitant costs. The upshot – things are definitely getting better.