AI is not a binary event
Mainstream media continues to position artificial intelligence as a step function event. According to the media, we are in the pre-AI era today. Then, one day in the future, artificial intelligence will finally arrive, and many of us will lose our jobs overnight. Talk of artificial intelligence stealing jobs overnight may make for good press, but that’s not how technological progress works in the real world.
In industrial automation (e.g., robotics used in car manufacturing), machines did not suddenly appear one day and displace human machinists. Industrial automation happened gradually, replacing little by little what human machinists did. In fact, the early days of industrial automation were quite boring and painstakingly slow. Human machinists scoffed at how little the machines could handle and how embarrassingly they could fail.
Yet, little by little, industrial automation progressed, replacing at first very little in the assembly line but gradually more and more of it. Now, automation has become indispensable at large, modern day factories.
To date, white collar jobs have been only modestly affected. Like industrial automation, workplace automation will not happen overnight. Workplace automation is happening very gradually right now and, for the most part, is utterly boring. But, workplace automation is chipping away at what humans do piece-by-piece.
Workplace automation bots are dumb, with no real form of intelligence. But, even though they are not intelligent enough to handle a wide variety of tasks, they are able to do simple, narrowly scoped out tasks more efficiently, cheaply, and consistently than their human counterparts.
If workplace automation software continues to improve and chips away at more and more white collar work, workplace automation will dramatically change the future of office work by the end of this decade.
Robotic process automation
This weekend, Will Knight at Wired wrote a piece called, “AI is Coming for Your Most Mind-Numbing Office Tasks.” He highlighted how robotic process automation (RPA) companies such as UIPath, Automation Anywhere, and Blue Prism have introduced software to do routine office work such as copying and pasting between documents.
It might not sound like much, but many entry-level white collar jobs require doing basic tasks such as copy-and-paste over and over again. For instance, think of just how many tasks in legal, finance, and accounting involve opening up documents and transferring the contents of those documents into a desktop application.
Phase one - simple software robots
In the first phase of automation, RPA companies developed software (known as “software robots”) to observe the mouse movements, mouse clicks, and keyboard strokes of humans doing routine work. Software robots then perform the routine tasks with the very same keystrokes as the humans, reducing the need for humans to do such mind-numbing work.
The software robots will consistently do the same task over and over again. As long as the task doesn’t change at all, the robots will do great work. These software robots aren’t very intelligent at all; they just mimic human keystrokes. But, if the nature of the work changes, the robots will be ill-equipped to handle the work.
That doesn’t mean these phase one software robots are useless. Software robots do simple, routine work consistently well and have replaced humans in performing some of the most basic data entry and data extraction work. Given the volume of such work, firms that have adopted software robots have seen a good return on investment.
According to Gartner, RPA is the fastest-growing segment of the global enterprise software market; RPA revenue grew by 63% in 2018 to $846 million and likely topped $1 billion in 2019. Investors are clearly backing the space. Two of the largest players in the space have raised significant sums; UIPath has raised $1 billion, and Automation Anywhere has raised $840 million to date.
Phase two - machine learning-powered automation
In the next phase of automation - the one currently underway - RPA companies are developing software that does not rely on hand-crafted rules to mimic human keystrokes. Instead, the software uses machine learning (i.e., learns a model from lots of data) to replace more of the work humans currently do.
One very hot space is intelligent document processing (IDP), which involves machines automatically extracting intelligence from unstructured documents. Companies such as HyperScience, Rossum, and WorkFusion have products that use computer vision and natural language processing to automatically extract data from unstructured documents such as invoices, legal documents, real estate leases, healthcare billing statements, government forms, and more.
Unlike the phase one software robots, the phase two machine learning-based products are more capable and more robust. They will start replacing more of the work humans do today. As with industrial automation, workplace automation is happening slowly but is automating more and more of the workplace as time goes by.
Why are these phase two machine learning-based products more robust than the phase one software robots?
The machine learning-based products are better able to handle unforeseen scenarios than the software robots. For example, in extracting data from invoices, software robots would do a great job if the invoice templates always remained the same. But, if the invoice templates change, the software robots would fail. Yes, you could develop a new software robot to handle the new template, but this becomes grueling over time; you have to keep developing new software robots every time the template changes.
With machine learning, the software learns from data how to handle a wide variety of scenarios. It is more robust and better able to handle data extraction from changing invoice templates, for example.
With machine learning products replacing more of the routine entry level work white collar workers do, what will the humans do?
Although machine learning products will do more of the routine work humans do, they cannot manage all of the edge cases that may occur. Yes, machine learning-based products are more robust to changes in the underlying work than software robots, but they are not as robust as humans. Humans will still have to handle some of the rare edge cases that occur.
In invoices, for example, humans will have to extract data from invoices that are malformed (e.g., maybe there are two invoices in a “single” invoice). Humans in these roles will also have more time to do higher value work that the machine learning products cannot (yet) manage. But, on net, companies will need fewer humans to do the same amount of work since the machine learning products will be cheaper and more consistent than the humans at performing the routine tasks.
Where do we go from here?
We are in the early days of workplace automation. To date, businesses have outsourced routine tasks to cheaper labor overseas. Businesses have not yet replaced this overseas labor with workplace automation software. But, that’s the natural progression.
The businesses that have begun to automate routine tasks have relied on simple software robots. Going forward, machine learning-powered workplace automation software will complement and/or replace simple software robots.
But, more and more white collar work will involve some degree of workplace automation software, and this is a trend we would be wise to embrace rather than resist. Humans working alongside automation software will perform tasks more efficiently, cheaply, and consistently than those without. The media’s version of artificial intelligence may not be here, but workplace automation has already arrived. Little by little, it will chip away at the work white collar workers do today.