I am a Machine Learning Scientist and a Machine Learning Skeptic. I have come to the same conclusions most of you have. HR and Recruiting Tech software companies are addicted to the Machine Learning label. At the same time, they are not as interested in delivering Machine Learning products.
HR and Recruiting do not need Machine Learning to function. The business will not fold tomorrow because those business units are not Machine Learning First. No one is getting fired for not buying more software. Budget is hard enough to justify.
Solutions require training and changes to how people do their jobs. Software can meet a business need and slow business operations. Some machine learning software vendors promise automation and deliver more work for everyone.
For these and a dozen other reasons, many companies should not buy machine learning based software solutions.
They Are Trying Way Too Hard
Value must be obvious and tangible. I use a weather analogy to help frame the need for data driven software. In most cases, we look at the weather or use a weather app in the morning before we get ready. Weekend planning requires the app to be accurate several days or a week out. Travel planning is driving some apps to publish forecasts for 10 days or longer.
The app provides easy to use information to help make better decisions. At the most basic level, better is evaluated by outcome quality (Did I bring an umbrella and jacket for a sunny, 90 degree day?) and time to outcome (How long did it take me to find the forecast I was looking for?).
Business decisions based on data are evaluated the same way. A good analytics dashboard can meet both needs. Machine Learning software companies sell the need for automation.
Can time to outcome be improved through automation? Yes. Will the automation improve outcome quality? This is where software with the machine learning label often falls short. In Recruiting, candidate discovery, resume screening, and workflow automation have a large field of machine learning software solutions. Workforce management, evaluation, training, retention, and many other HR functions do too.
They Have This Amazing Power Point Sales Pitch
This is where vendors wander into jargon and jazz hands. Deep learning and natural language understanding. Proprietary algorithms and data. Anti-bias.
Most machine learning software is using logic by another name. Under the hood, there are models trained with a lot of data. Functionally, those models do the same thing as traditional software. Keyword matching and generic, semi-targeted outreach (amongst others) are common to both types of software. Automation is automation no matter what we call it.
There are software solutions that use applied machine learning. Models can point out employees or potential new hires who have the wrong job title or are at the wrong level of seniority. They have been misclassified but a model can review their qualifications and put them in the right bucket. That can open up or pare down the candidate pool. It can reveal employees who have a senior title, but their skills and capabilities are associated with mid-level roles. In the opposite case where someone is under-titled, it can help determine who is a good candidate for promotion.
Organizational Development can benefit from models pointing out employees who can transition into hard to fill roles. Often a few months’ or less than a year’s worth of continuing education can move their careers forward within the company. Comparing cost of training versus cost of hiring can justify the budget.
All You Have to Do Is Change Everything You Have Been Doing
Use cases have real upside. Both use cases from above require new or greatly revised HR processes. They impact other business units. The concepts have merit but require operational changes. Are those changes and the new software’s expense justified by the business model?
No doubt there is ROI but is the transformation dictated to the business by the software or is the transformation driven by the business and enabled by the software? Most businesses are going through a digital transformation at some level. Most businesses are not making transformations at the scale to justify applied machine learning solutions.
This is the case for my thesis. There is no value in replacing traditional software with a new solution that is different in name alone. It is more practical to upgrade an in-place solution. Every vendor is upping their game to stay competitive.
There is ROI from migrating to an applied machine learning solution. However, there is not a business case for the operational changes required to support those solutions unless the business has already decided to implement those changes. The solution must enable changes driven by an evolving business model.
They Spend More Time Talking About Users Than Talking to Users
I am a Machine Learning Skeptic because there is a difference between monetizing machine learning and just launching a product. Most software companies are just launching products with no connection to you, real world users.
They are not building applied machine learning solutions that fit into the way people work. Their solutions are a bit condescending. The premise is that people are the problem. They are dinosaurs if they will not bend over backwards to accommodate the software. Fewer people somehow translates into better business.
How Is My Field Going to Fix This?
That is slowly changing. We are working on Decision Support Systems. Those give people information, just like a weather app, then gets out of their way. Those keep people in the loop and in control of key decisions. Decision Support Systems are also competency aware. They are trained to evaluate their own abilities in different situations. Essentially, they know when to speak up and when to shut up.
What do they look like? A dashboard. Under the covers it is measuring the quality of outcomes. Did that set of information lead to a better outcome? Did that set of information lead to a faster decision? Has the model learned enough to provide information to support new types of decisions?
There is complete transparency. If you click on a data point, the system can explain what decision it thinks that data supports. If you remove it from your dashboard or keep it, the system learns from that.
The whole design hides the complexity, so all the user sees is a dashboard that gets better on its own. The system does not treat the user, someone with years of expertise, like a speed bump.
Unfortunately, the practical solution is dragging a new buzzword behind it. Get ready for “Third Wave AI.” If I am being honest though, it sounds a whole lot cooler than “Machine Learning Based, Capability Aware, Decision Support Systems.”