RPA, ML, and AI is all the rage these days, with RPA being the most mature technology. But just because it’s a mature technology, that doesn’t mean it’s a mature technology offering from the vendor you are considering, as powerful as they are purporting it to be, or even as general purpose as you might expect (especially if it relies heavily on ML or AI techniques).

In fact, like early spend classification technology, which was usually 60% auto-class and 35% behind-the-scenes manual-class by the hundred interns in the backroom (in India, Poland, or another outsource locale with a relatively high percentage of English-as-a-second-language speakers), a lot of the RPA technology being promoted today is in fact supported by, if not done by, humans behind the scenes.

This is especially true when natural language processing is involved, and doubly true when interpretation is involved. And it even comes in to play with something as simple as calendar scheduling. For example “book me an appointment with John Russell” next Wednesday is not often straight forward. John Russell the person, or John Russell the company? And the next calendar Wednesday, or the calendar Wednesday in the next week? (English speakers typically refer to the next calendar Wednesday as this Wednesday and the Wednesday in the following week as next Wednesday, but certain European cultures always refer to the next calendar Wednesday as next Wednesday.)

So imagine how much human intervention is required behind the scenes if you want to do document analysis or contract interpretation! Quite a bit. When it comes to document and contract processing, it’s one thing to break it up into sections and annotate what’s in the document, it’s another thing to interpret what each section means, and yet another to determine whether or not its enforceable, or even allowable, against a regulation or law.

Advanced RPA / ML systems can analyze a document or contract and break it up into relevant sections, identify the constituent components of each section (party, address, obligation, description, explanation, etc.), and make it easy to determine whether or not a section, entity, or value is contained within. With sufficient ontological definitions, training, and tweaking, these systems can get highly accurate.

But when it comes to interpretation, that’s different. It’s easy to determine that a document contains the phrase “the receiving party is bound to provide the sending party with a hold payment to be applied against the obligation of the sending party upon transmission”, but harder to figure out precisely what that means if the hold payment, receiving party, sending party, and obligations are specified elsewhere in the document. And then if you want to determine whether or not that obligation is in line with organizational policies or contract law in the jurisdiction of choice, that’s yet another level.

Really good tech might be able to sift the document and make probabilistic guesses as to what the hold payment is, what the obligation is, and maybe even what transmission means (providing to a courier, being received by a local courier, showing up at the recipients door if its a physical good, or confirmed receipt / acknowledgement if an electronic IP deliverable), but chances are it will be wrong a good percentage of the time and require human confirmation. And when it comes to interpretation, frankly, unless a human is reviewing the clause and given the most likely scenario, a random number generator mapped to an outcome table is likely to be just as accurate. (In other words, trusting RPA means you are rolling the bones.)

Thus, any RPA system that performs an advance task is likely not true Robotic Process Automation but in fact Redistributed Process Automation, even if the vendor doesn’t advertise it as such. But if you are curious, there are tells. How long does the system take to perform the task? An hour or two to process a document? Definitely RPA of the second category. Fifteen minutes or more to schedule that appointment? If both sides were using true RPA of the first type, it would take seconds, maybe a few minutes if there was limited bandwidth and email delay. And so on. Look at service times, customer counts, and what’s being heavily promoted. The truth is under the covers.

But redistributed process automation is not necessarily bad. It’s probably the most efficient use of your organization’s time, especially if the vendor has RPA-lite algorithms that can quickly determine what needs to be done by a human and what can be automated. Anything that saves your organization time and money while improving outcomes is a step forward, and as long as the vendor continues to reinvest its profits into system development, the system should get better over time.

But don’t buy RPA with eyes wide shut. Otherwise, you might not get what you are expecting. Or put too much faith into the system.