Speak then to me…

8. Disillusion


And it seemed as if a voice
(Sweeter far than by harp or by psaltery
Is breathed) called out, ‘Oh rats, rejoice!
The world is grown to one vast dry-saltery!
So munch on, crunch on, take your nuncheon,
Breakfast, supper, dinner, luncheon!’

And just as a bulky sugar-puncheon,
All ready staved, like a great sun shone
Glorious scarce an inch before me,
Just as methought it said ‘Come bore me!’
— I found the Weser rolling o’er me.”

Robert Browning (The Pied Piper of Hamelin)

In previous articles, I traced the origin of screen-integration technology, the evolution to its modern form, and its morphing into UI Automation with the primary intent to automate business processes that were subject to the inefficiencies of software-driven-labour.

The product-market-fit for this technology had existed for a very long time; but buyer behaviour, which I correlate with human IQ, yet again proved Geoffrey Moore’s technology adoption lifecycle theory. Tired of the notoriously dogmatic and slow buying behaviour of IT departments in large enterprises, we had moved our focus on the people who actually wanted to get things done— the business and operations people. This target group needed a non-technical basis for relating to the product.

While we (Inventys) were busy with the acquisition of our company and adjusting to the new owners, others who had joined the race went ahead and set up necessary marketing foundations for promoting UI Automation among business users. Thus the category called RPA was formed.

Once the buyers’ inertia was overcome, it did not take long for organisations to realise that this type of automation could give them reduction of 10%-30% in software driven labour cost very easily. Some process automations could even eliminate 100% human intrerventions. That scenarios were good enough to get any CFO to sit up and pay attention.

Between 2016 and 2019, the market went berserk with RPA: dozens of new product companies released competing products. Organisations created “centers of excellence” and appointed people to designations like “Robotics Lead”.

I can still manage to hold my calm when I see or hear the phrase “robotic automation”; but the term “robotics automation” is like fingernails scraping on chalkboard. Even mainstream vendors and supposedly high-end consulting firms are guilty of using this malapropism. The word “robotic” in “robotic automation” is an adjective used as a metaphor to describe the quality or type of the automation being described. It signifies a type of automation that can be likened to a robot using a keyboard and mouse to perform software-driven-labour. On the other hand, “robotics” is a noun in singular form that denotes the type of technology that deals with creation and usage of robots: real robots, not metaphorical robots.

As with most things that people do without thinking, the adoption of UI Automation — now called RPA — fell into a cognitive bias syndrome.

Dunning-Kruger Effect

The Dunning-Kruger Effect is a cognitive bias that causes people to feel extremely confident about a topic as their knowledge about that topic increases by a small extent in the initial stages. This confidence reaches a peak when people move from zero knowledge to little knowledge about some subject, which some authors call “Mt Stupid”. As people get wiser through further knowledge about that subject, their confidence drops. They encounter a barrage of failures and exception scenarios, and they realise that there is much more to the subject than what they had initially assumed. From this valley of despair, through patience, further study and analysis, people gradually climb the slope of enlightenment to reach the plateau of sustainability.

A few articles and blog posts were published, starting in 2020, that “debunked the Dunning-Kruger Effect”. The original study by David Dunning and Justin Kruger did not trace journey of individuals or groups, from ignorance, to little knowledge, to wisdom. Instead, it recorded the self-assessments vs actual results of individuals about university examinations that they had taken. While this is true, Dunning et al have subsequently published the result of six empirical studies, from which they concluded that “… although beginners did not start out overconfident in their judgments, they rapidly surged to a ‘beginner’s bubble’ of overconfidence.”

From my own experience of observing myself as well as people that I interact with, I continue to believe that the DK-effect is a valid cognitive bias.

It is widely believed that this concept has been used as the basis, by a well known technology research and analyst firm, to produce what they call “hype curves” for the various technologies that they “cover”. Their aim is to show that often, most technologies go through a journey characterised by the shape of the hype-curve. Initially, people (users, beneficiaries) are unaware of the technology. After a few successful implementations, there is a wild increase in their expectations and intentions to adopt the offering. Vendors seize this wave and blitz the space with exaggerated narratives. But soon, the buyers realise that the product does not meet all of their expectations, resulting in a plummeting of their enthusiasm for using the product. This shock that affects both vendors and buyers, is followed by a maturation phase characterised by gentle rise in adoption based on a combination of product improvements and genuine understanding of its capabilities.

New technology introduction scenarios often cause hype cycles to form because marketers will attempt to flood the mind-space of prospects with various positive aspects about the product being introduced. Very often the product makers are themselves victims of the “beginner’s bubble” syndrome that Dunning et al describe. Initial successes of the product in prospect/client organisations cause their stakeholders to also enter their own beginner’s bubble. Thus the product makers’ initial overconfidence, coupled with the clients’ initial overconfidence, catalysed by marketers’ and research firms’ spins, creates a powerful concoction that results in a market-wide expectation of miracles from the technology. People within enterprises who wait for transformative events to occur, also jump into the game and promote the hype so that transformations that are favourable to them can be implemented.

The trajectory of RPA’s adoption is a stark example of Dunning-Kruger effect and hype-cycle. The marketing companies had excelled at what they do best. The Pied Piper had composed a mesmerising tune called “robotic process automation”. The enchanted masses followed the Pied Piper merrily up the hype-curve.

Soon they were staring down at the “valley of despair”; and there they saw a set of hard problems that were going to impede the accelerated deployment of RPA in their organisations. While there were a few processes that were fully automated (causing the above mentioned exuberance), automations for a large set of manual processes failed and got delayed due to the following reasons:

  1. Most processes were triggered by documents, from which data had to be extracted. These documents arrived in semi-digitised or un-digitised forms.
  2. Many organisations did not have properly documented software driven labour process steps. Discovering these and configuring them created huge projects of their own and the returns on investment for such automations either did not exist or were severely delayed.
  3. Software driven labour processes themselves were highly inefficiently operating at the point when RPA was being attempted on automating them. The “as-is” processes were erroneous and missed various special conditions. Blindly applying RPA on such processes only magnified the errors and caused wasted efforts to refactor the processes.

Inventys had encountered some of these issues in 2010 — before the RPA blitzkrieg. Some of our customers had forced us, with good intentions, to implement data entry automations. We pioneered the hybrid use of optical character recognition (OCR) in UI Automations due to these compulsions. Most of the data entry scenarios required reading scanned documents, extracting specific data from these documents, and orchestrating them into enterprise applications using UI Automation. Our OCR was accurate in converting the required data into digital form, but the clients wanted us to compete with data entry operators who spent most of their time transcribing data from image viewers into downstream applications. They microbenchmarked our OCR + UI Automation against career-data-entry-operators. We failed miserably. This was because when a human looks at a single page of a scanned document, they can simultaneously perform multiple rule evaluations in parallel and can tailor their data entry actions accordingly. For example, within first glimpse the human would be able to decide if a certain row with a certain value was present on the page. Based on this they could tailor the data entry steps to achieve minimum latency. In the early days of implementing OCR + UI Automation, our workflow followed the generic rule set provided by the subject matter expert for the given flow. This required us to apply OCR and then iterate field-by-field and apply the flow rules that were given to us. Today, perhaps, this would be optimised by applying a two-pass mechanism of computer-vision (CV) and then OCR and then configuring additional heuristics that experienced data-entry operators would use.

Even though we failed in microbenchmarking of data entry automation, we were successful in promoting the use of OCR for scanned image digitisation. The gist of the argument was that our software runs twenty-four hours non-stop and therefore can score a better throughput on a macro scale as opposed to humans who work in shifts and have to take various short breaks during their working hours.

Despite several successes, I recall that standing in BPO offices among senior leadership and operational excellence managers, talking about OCR and CV was my “Mt Stupid” moment. This was perhaps year 2013 or 2014. I felt that something was not right about this entire concept of process automation that we had got ourselves into. We realised that a significant proportion of business processes in BPOs and SSOs require data extracted from scanned images to be used in the flow. Avoiding processes that needed document data extraction was leading to show-stoppers from a sales perspective. We had initially implemented CV-OCR to get over that acceptance hurdle. While it helped us to press ahead with more deals in BPOs, this concept was causing much discomfort to me from a product roadmap perspective. OCR’ing documents in back offices is a can of worms. There were other pure play OCR companies that were generations ahead of us from an OCR perspective. Their products could do advanced actions such as noise removals and advanced handwriting recognition. Since we declared that Fusion had a built-in OCR, several customers wanted to use that feature only for OCR work. There were a few disappointments when we explained that our OCR did not have advanced noise removal features such as reading text from below coffee cup stains and rubber stamp marks. The mainstream OCR products were formidably expensive for many back-office scenarios. I did not want to spend time building out advanced OCR capabilities. So, I attempted to contact the mainstream vendors to form a partnership. Unfortunately those vendors were too fearful of losing control, and kept dillydallying to point that I lost interest. This happened in the pre-RPA hype days. Later the same vendors went on to acquire or build their own RPA-like solutions; but none have reached any level of maturity and acceptance as pure-play RPA vendors.

By around 2015 I was back to the state that I had found myself in 2002-2003 — I did not align with the “trend” that had set in. In 2002 it was the idea that integration and portalisation require transformational projects and everything had to be rewritten to comply with the integration and portalisation middleware. That deviation from mainstream thinking had caused me to breakaway and lead the creation of Inventys Fusion UI Automation tool. In 2015 it was the notion of trying to “fix” human actions through screen-integration automations that got me to readjust my thoughts once more. Screen-integration based UI Automation was being taken in the wrong direction. I was back to working as an employee of the new owners of Inventys; therefore extreme directional changes could not be taken without approval.

There was a period of general disillusionment relating to RPA setting in. The consulting firms produced various “studies” that showed adoption issues with RPA — by geography, by industry, survey results, and all the other usual stuff.

But before things got worse, a few companies managed to get Series A funding and one even went on to a small IPO. Flush with cash, these companies embarked on creating new jargon and some genuine improvements in the functional capabilities of their respective RPA products.

Meanwhile, the Pied Piper who piped the original RPA tune was also getting frustrated. The creation of the RPA category and the positioning of UI Automation as RPA really helped these startups and “pivots” to achieve high valuations. I think the Piper did not get his due, and the RPA companies, like the Mayor and Corporation in Browning’s poem, might have exclaimed:

A thousand Guilders! Come, take fifty!

At which point the Piper might have thought:

And folks who put me in a passion May find me pipe to another fashion.

The promoters of the top two RPA companies would have made a fortune, which even I think was primarily due to the “RPA wave” played and popularised by the Piper.

Based on the view from “Mt Stupid”, and taking cue from the many failures that RPA projects were experiencing at that time :

Once more he stept into the street And to his lips again Laid his long pipe of smooth straight cane; And ere he blew three notes….

RPA Is Dead, piped the Piper.

About the author

Madhav Sivadas

Add comment

Leave a Reply

Speak then to me…
%d bloggers like this: