Garbage In --> Garbage out

Welcome to the new episode of #AutomationPills!

In the last episode we introduced a new framework to evaluate the automation potential of a given task.

Now we get into the first part of the framework, where we talk about the first step of every automation journey: the input.

The good, the bad, and the ugly

We find again the protagonists of our first episodes, John from the IT department, and a new provider of a magic automation solution, TrustMe.inc.

Let us imagine for a second that John and TrustMe are splitting 50/50 all the requests of password reset.

Some are good, some are bad, and some are ugly, but John and TrustMe have different tastes.

Imagine them receiving something like that:

{
   "requester_name": "John Doe",  
    "requester_email": "[email protected]",  
    "user_id": "12345",  "token": "a1b2c3d4e5f6", 
    "expiration_date": "2023-04-01T12:00:00Z",  
    "status": "pending"
  }        

It cannot get better than this. The request is structured, clear, and all the data can be easily isolated (for the techies, it's a JSON format).

This is deeply appreciated by both John and his machine counterpart.

However, we know that the world is far from perfect, and sometimes the requests arrive in weird shapes.

Imagine that to process the ticket, you must find a code hidden in the following text, and it consists of 10 digits after the word "kexokuv".

Now try to look for it:

buzinexyvofroqimyjoglokydhojybwypywykupebahuvmizudolobetabyvuhykupuralekebizifasorinazequzegifebivoxiloqyqehytuzezuloqevuzokepylujiluqozaxagopiqexafuzunutylegozujyxagobeporymuravinoxiwuzyzidodogysuzybelavugubapazymubezyzizuvikyvutozydobelifujymowotanixuwuluxoxexavijavoqequwitoqatuzepenabikaxyhezajafyvepytugezevyxulovoxuxuqokijakotybyqarabofodadupybuzyquyjiqypobokyjohunagezycejovybyvyrugypykojapubiybixitozykikyvobygezypypyzykexokuvybafyxuvoquzidupuxygiwopemyrofiwiqihamavoquzyp
a
        

The answer is "ybafyxuvoq", and I hope you used Ctrl+F to find it. However, the request is still structured, and suitable for automation. It is super easy for a computer, but very hard for a human.

(If you think it's an unrealistic case, you probably never had to debug some logging files manually, and you can consider yourself lucky).

On the other hand, imagine you receive a ticket like this:

Hey Stefano, do you remember that account we were talking about the other day? 
Well, I need it reset. Can you kindly do it asap please, I have a presentation in 2 hours.
Thanks
Best
[Evil Colleague]        

This is way easy for a human,assuming your colleague was not hallucinating. However, it is a nightmare for a machine.

NLP (Natural Language Processing) is becoming increasingly good and the interpretation of text is improving by the day. Nonetheless, it is still tricky for a machine to reason from context.

The situation gets even more complex if to solve you need:

1. Information that is not part of a public model, but rather of the very specific situation of that company/process/team. For example, if your company has a special policy, where you can reset passwords only at a certain time of the day
2. Translate from other languages, particularly from complex non-latin alphabet languages, like Arabic or Mandarin. Translation can be challenging also for humans but, considering the case of a fluent speaker, it is generally easier to grasp nuances. And most of all you can avoid a situation like the following:

Easy for you, hard for me

Where is this difference between John and TrustMe coming from?

There are several key parts to this question, but I usually recap it with a single concept, that I introduce here but it will accompany us throughout our entire journey:

Humans and computers reason in a profoundly different way.

Which brings to us the following corollaries:

What is easy for a computer, is not necessarily easy for a human

What is easy for a human, is not necessarily easy for a computer

You can picture it in the following way

It may seem trivial, but it is not. I have seen seasoned executives completely missing the point.

As a (very) generic rule, computers like data sets structured and well-defined, while inputs that are unstructured and ambiguous may not be as machine-friendly. Their "map of world" is very limited, and they can have problems with requests heavily reliant on context. We will elaborate the "context" part in the following chapters.

Computers struggle also with non-verbal communication. Humans are adept at picking up on subtle cues such as body language, facial expressions, and tone of voice.

On the other hand, computers have rapidity and scalability on their side. They can process vast amount of data, continuously and consistently, at a speed that humans simply cannot match.

Another advantage often overlooked is that their processing is bias-free. An artificial model does not have any preconceived notions or biases that could affect its analysis. A computer will never cherry-pick data points from a large set just to look smarter or confirm a previous intuition (this is called confirmation bias, and I am greatly simplifying, but you get the concept).

Elaborating a bit what we said, our Venn diagram now looks like this*:

Now that we have covered this key concept, we are ready to jump into a critical part of our automation journey, the interpretation of the request.

*Again, let's stress that this is only for non-physical activities. Dexterity of machines - and robotics in general - are out of scope for our discussion.