Automation or Customization? You Can’t Have It Both Ways

In business there is a classic “trilemma” if you will: Fast, Cheap, or Good — Pick Two. If your company’s name is Netflix, then you have the luxury of picking Fast and Good. For those with much shallower pockets, the choice is thus reduced to either Cheap and Fast or Cheap and Good. In the realm of data analytics, Cheap and Fast in practice translates to automation of reporting and workflows, while Cheap and Good manifests as manual, time-intensive processes requiring more hands-on human interaction at every step in order to ensure the utmost quality.

As the title of this article posits, You Cannot Have Both. Everyone needs to understand that Automation and Customization are mutually exclusive concepts. Automation necessarily involves setting things up so that the computer executes procedures that run in the background without anyone needing to lift a finger, except at the very most to click a button. Customization is the polar opposite. There are some things that only folks with boots on the ground would know, and their input and expert review is absolutely critical to validate and operationalize any data analysis output. As front-line workers are the end users of reports intended to improve performance, then ultimately meeting THEIR needs is the main goal. If the data analysis product is not being implemented and integrated into actual workflows, then all you have is a Bright Shiny Object.

With that being said, there are times when you really need to be Cheap and Fast. In that case, compromises need to be made, i.e. there needs to be some blend of automation and customization. Although my background is in healthcare data analytics, this idea of compromising between Cheap and Fast and Cheap and Good is completely applicable to all sorts of verticals. Similar to what I did in a prior article, I will try to show how this data analysis concept can play out in a totally different industry.

I started my career working on the assembly lines in a factory that made beauty products for pets and farm animals. No joke. It was my job to screw tops onto bottles of doggie cologne and horse conditioner. Back then in the early 90s, data analysis was not a thing. But it sure is now. Let’s say the board of directors would like a report showing the productivity of assembly line workers. The data analytics department has come up with a super spiffy pdf file that automatically gets emailed to the different product managers on a monthly basis that looks something like this:

The horse conditioner manager loves the report because it makes him look like the most productive, while the doggie cologne and pig lipstick managers hate it because it makes them look like they are slacking. They complain to the data analyst’s boss that the report is not measuring the right thing. Of course, the horse conditioner has the most boxes packed per month — because bottles of horse conditioner are physically larger than doggie cologne bottles and pig lipstick tubes by orders of magnitude — it takes fewer units to fill up the same size box!

Therefore, the data analyst, who had no idea of the relative sizes of packaging for animal beauty products, comes up with a brand new automated report that has what she hopes is a more useful metric for measuring assembly line productivity — average number of units assembled per day compared to the same number the previous month:

Uh oh, now this report is making clear that the horse conditioner assembly line is performing worse this month than last month. Now the horse conditioner manager complains to the data analyst’s boss. The reason why there were fewer bottles assembled this month than last is that the star bottle assembler was taking a much needed two-week vacation after having no vacation in the last two years, so they had to hire a temp who needed some time to get up to speed. Can’t the report be modified to take such events into account?

So the data analyst goes back to the drawing board. Now the report is perfect because not only does it allow for a fair comparison among the different product lines (because each one has its own benchmark value by which the performance metric is judged), it also allows for the occurrence of an outlier here and there (because it is comparing against the 80th percentile, meaning that up to 20% of the time there can be exceptional cases that won’t count against the manager).

So, was that whole process Cheap and Fast? Not by a long shot. My animal beauty product example shows that it can take several iterations to arrive at a satisfactory data analytics solution. But was it Cheap and Good? Yes! The final solution gives an accurate picture of assembly line productivity. The automatic monthly report can catch any drops in productivity that are beyond the standard allowance. In this example, we built in a 20% allowance, but your organization could be more or less. However, it is unreasonable to insist on only a 1% allowance for example, or to be as lax as accepting up to a 40% allowance.

What is not visible to upper management is that the data analyst spent a lot of time to refine that performance metric to determine that on average, about 20% of the time there is a drop of assembly line productivity due to such things as employee turnover, equipment malfunction, seasonality, etc. Not 10% of the time, or 30% of the time. You just can’t arbitrarily pick a number and call it a day; everything needs to be evidence-based. And in order to find out all these legitimate justifications for drops in productivity, the analyst had to have meetings with HR, assembly line foremen, IT, and the data governance team to get access to numbers on staffing levels and equipment breakdown rates.

I hope that I was able to illustrate that in the end, you can have a fully automated data analysis solution (Cheap and Fast). It just takes a whole lot of customization (Cheap and Good) to get there.