Failure of AI proof of concept industrializations, a really bad recurring joke!
My oh my ! I must say ! Various sources are reporting that up to 85% [1] of Artificial Intelligence/ Machine Learning proof-of-concepts (POCs) are failing and many articles, posts or comments online intend to analyse the reasons. However, it feels like in these posts only symptoms are addressed and not underlying deseases…. if I am allowed this expression. Consequently, I thought it was time for me to howl with the wolves and expose my opinion on the matter.
Therefore, let us come back to what a proof of concept is. Despite the general idea that it is a way for consulting firms to enter customers premises hoping to sell other projects or products, a proof of concept intends to demonstrate that an idea, a methodology, an algorithm is somehow valid. It does not mean in any way that the methodology used is the one that should be industrialized, far from it….. POCs are usually carried out considering “pseudo” research approaches, but these might often be of substandard rigour. I am not blaming anyone here, it is merely an observation, and the reason might be related to underlying timelines, budgets, etc., unfortunate and most of the time exogenous elements that we cannot necessarily overcome (the key word here being necessarily). Beside, POCs are usually leaning on the “fail fast, fail early” philosophy, and it is not that simple: what does mean failing ? What does mean failing fast ? What does mean failing early ? Is that getting a precision of 97% when a precision of 98% was expected? Was that our expectations were not set at the right level ? Was that because the flow of information changed at the last minute ? And so on, and so forth….
In summary, one may wonder if the whole problem is not just related to the entire process, I am also wondering if the criteria set for success of an experimentation are suitable at an industrialized level. It feels like I am labouring the point, but how can it be otherwise…. The objective is not to write a paper with that experiment, but to assess the suitability of the solution over time, which mean anticipate the issue related to various components such as the connection of the incoming data flows, the way people are going to use them, etc... We are still doing POCs on topics that have thousands of academic papers published addressing the matter of interest. I am barely exaggerating saying that either we are trying to prove over and over again well known results, or companies do not trust or do not know the published papers, or it is just a way for some companies to start a project at a lower cost. Though most arguments can be justified, this does not favour the success of the projects. I guess the worst in all that, is that we are not actually experimenting with POCs, we are merely replicating, because a proper experimentation plan would implicate that we are bringing something new, though most of the time we are not…. in other words we would need to get off the beaten track, and I wonder if one problem is not company cultural risk management, indeed considering the risk associated with experiments, managers are likely to be reluctant to allocate resources….. and if experimentation implies risk, and the culture is too risk averse, then needless to say that we end up in a “slightly” contradictory situation. Another aspect is that we might be doing POCs where we should do pilot project, as a pilot refers to an initial roll-out of a system, considering a lighter version of the final solution, but capturing all impacting or influencing elements into the implementation plan, and that, in a production environment.
Drawing a parallel between medical trials, economics, and artificial intelligence, I would like to quote Banerjee et al. 2017 [2] (paper written with Esther Duflot Nobel Prize in Economics Sciences in 2019): “In medical trials, efficacy studies are usually performed first in tightly controlled laboratory conditions. For the same reasons, it often makes sense to verify proof of concept of a new social program under ideal conditions — by finding a context and implementation partner where all the necessary steps for success are likely to be taken […]. However, the results of such a program tested on a small scale, while informative, are not necessarily a good predictor of what would happen if a similar policy were to be implemented on a large scale. Indeed, it is not uncommon that larger-scale studies fail to replicate results that had been established in small […] randomized controlled trials elsewhere.”
Reluctancy, power of habits, spill-over effects, unexpected non-linear dependences, human errors, etc. are as many things that can lead to the failure of industrialisation of AI POCs. In other words, we cannot say, I know how to drive the remote controlled car of my 5 year-old, therefore, I should be able to drive a Ferrari on a race track…
So, now that I have pointed out some issues, I believe that it is time to be slightly more constructive, and provide some ideas on how to deal with those.
· First, now that we are aware of these issues, and if we are really interested in carrying out a POC, I would recommend revising our expectations, but that is only the bare minimum,
· Second, a more important question is: what do we want to learn with POCS ? Which aspect are we interested in testing? And then, what am I ready to compromise on when we industrialise the solution? What timeline should we consider to get the whole solution functioning? How can we gradually open to new aspects such that the development is controlled and that if an issue arises, we are more likely to catch it and address it quickly. If I see that I am more interested in the result of the solution functioning in a production environment, then I would recommend a pilot.
· Third, coming back to the “fail fast” aforementioned, I guess it is time to understand that fast does not necessarily mean immediately, and the important part related to failure is to learn something… Learning something out of a POC is not a failure, it is just another step towards success.
In summary, I would recommend not to jump to conclusion with respect to the results of POCs, and if you want to analyse the behaviour of the industrialised solution, immediately start with a pilot project addressing the identified problem as a whole, it might cost a bit more upfront, but in time it might prove extremely cost efficient.
[1] https://www.datanami.com/2020/10/01/most-data-science-projects-fail-but-yours-doesnt-have-to/
[2] Banerjee, Abhijit, et al. “From proof of concept to scalable policies: Challenges and solutions, with an application.” Journal of Economic Perspectives 31.4 (2017): 73–102
