Robotic Process Automation, Machine Learning & Artificial Intelligence
Robotic Process Automation (RPA), Machine Learning (ML) and Artificial Intelligence (AI) get a lot of attention these days. After studying these technologies and finding ways to apply them appropriately in our business. It is clear that there are vast ranges of perceptions of what these tools do and how they can be applied by managers and by marketers. A lot of hype is being used (as expected) to over sell benefits, and very little discussion takes place about the costs connected to these technologies.
RPA is sometimes represented as AI. Some marketers are describing an RPA solution with ML capabilities as if ML is a standard part of RPA. They do not give any indication that the cost thresholds increase from RPA to ML to AI. I recently saw an advertisement for RPA with NLP (Natural Language Processing) presented in a way that seemed to indicate that you can ask the system for anything and it will do the work and render analytics you didn’t even know you wanted.
When I worked as a consultant, an expression I often used with my customers was: “If you can think of it, we can be build it. But at what cost?”. Every application of technology has costs and benefits that need to properly detailed to analyse the value. Nothing the marketers are saying is untrue. It’s all possible, but at what cost and with how many layers of technology (RPA, ML, AI, NLP. . . )?
If you are considering how to take advantage of these technologies for your business, you need to understand the differences in these technologies and where costs can quickly rise. Some things you may want to consider include:
- RPA is used for structured data (field to field mapping, standard data types, clear validation rules, rules for data transformations, sometimes called swivel chair work.)
- RPA is not always the best solution to your automation need. Make sure you have properly vetted any workflow capabilities inherent in your internal systems before considering throwing the “technology of the day” at it.
- RPA at its basic level, does not include any ability “to learn”, nor to make qualitative decisions that are aren’t programmed into a decision or validation set of rules, cannot decipher unstructured data (in coming emails, documents that might be scanned or images of scans, file attachment parsing…)
- RPA is the least expensive of the automation tools available and is often marketed as delivering the benefits of 5 to 1 or 10 to 1 in terms of FTE (full time employee) cost savings
- Machine Learning is used to describe systems that learn over time to recognize or validate data. It is not AI, in that it cannot learn on its own. It typically uses a pattern matching approach (in the case of invoice processing for example) to compare a data set to known examples. The system is trained by a user as to what constitutes a pattern match and what doesn’t. As the system increases it success rate, the human is eventually taken out of the loop and the machine is trusted to recognise and process the data appropriately
- Companies getting involved with this technology often do not appreciate what the volume of data is going to be, that’s required to train the system. The quality of the data may also play an issue here. Make sure your prospective vendor is clear on how many samples have to be reviewed before you get to your required success rate. How is the solution priced (cost per attempted match?). In an invoicing example, what is the cost structure to learn to pattern match for all your suppliers (not just the example that will be done in your proof of concept)?
- Are the ML costs quoted separate from the RPA costs?
- Natural Language Processing is a technology that allows you to speak or type to a system to ask a question or give an instruction. Common uses are to ask a question about a data set (how much did I spend with vendor X last year? How many vendors supply commodity Y and what was the spend last year). NLP could be used to decipher a text or an email, breaking it down to nouns, verbs, sentiment, for example – and passing the salient data points onto an RPA process or an AI process
- NLP is often bundled with another technology – as a part of an RPA offering, or as a part of an AI/analytics tool for example. Vendors will attach value to the NLP capabilities. Be clear with your vendor how it is being costed and if there are thresholds that might cause costs to rise. Will the vendor itemize this part of the solution so that you can asses the cost benefits?
- There are plenty of articles out there about narrow AI and deep AI. Deep AI is still an evolving vision and not ready for consumption in most businesses today. Narrow AI refers to an AI application that is configured to deal with a very specific focus. It will be programmed “to learn” how to interpret data in a process to arrive a decision/prediction/recommendation.
- Narrow AI solutions are being developed in a broad range of industries, and many for tasks that are common across businesses. This will lead to the commoditisation of some of these applications but it is still early days. An AI solution works with complex algorithms developed by data scientists, to qualify (weightings) variables that feed into a process. These variables can be limited to a few or as wide/vast as needs to be. The quality of the outputs of the AI obviously have a lot to do with how well these algorithms are constructed and tailored for your business
- As AI is still an evolving commercial opportunity for many vendors, the pricing models are also still evolving. Some vendors talk about a fraction of a cent per call to an AI API. How many calls would your process invoke in a month? Hundreds of thousands? Millions? How long will it take the AI to learn the narrow function you have defined? Have you defined the requirement properly? Speaking to a number of my peers that have been testing this level of AI, I am hearing anecdotal stories of costs exponentially spiralling.
- The easiest AI’s to apply to your business will be a component of an existing solution as opposed to building an AI function from scratch for your process. A good example is an “auto classification” engine within a purchasing system that recognizes the item you are sourcing or working with and automatically applies your taxonomy to it (or applies a known standard such as UNSPSC).
These technologies are unavoidable. We all come in contact with them on some shape or form every day. They will play a part in your business sooner or later so you need to understand how to leverage them to your advantage. As you do your research, keep a commercial lens focused on the prospective solutions. Press your solution providers to be clear on the capabilities of each technology and provide clear pricing and guidance on thresholds. You need to have confidence in the value proposition for your business and should have options in terms of configurations. Buyer Beware is as applicable as ever for these evolving technologies.