By Gastón Arismendi


If you are reading this post, surely you have already heard about some of the main differences between AI and traditional software development when conducting project management. Beyond being a newer industry and not having the same level of maturity as the former, what differentiates AI as discipline could be pictured with only one word: uncertainty. This concept is present at many levels of an AI project. 

In this article we will walk through some of them and  reveal how we deal with this uncertainty at Marvik so as to keep delivering high quality results, empowering  our clients to be a step ahead through the power of AI. If you are a manager in this industry or one looking to join the team, this article is the right one for you!


Sources of uncertainty and painkillers

In this section we will describe some of the most common situations that bring uncertainty to the game when dealing with AI projects. For each, we will explain  how we manage to keep uncertainty as low as possible to ensure the project’s success.

I want my AI

When noticing the title of this section, those in their late 30’s or over may have heard Sting’s voice and Mark Knoffler’s legendary guitar riff resounding inside their heads (for the younger ones who do not have the tiniest idea of what I am talking about, please just search “Money for nothing” in Youtube, we can wait until you get to the 4:39 minutes to go on). 

I know this section title may sound as if having AI features is just a kind of whim, but that is not the idea at all. AI is on the spot right now, this is not a revelation for anyone related to the technology sector, and leaders from many other industries are also aware that they could boost their business by leveraging AI. However, in many cases they reach this conclusion  based on a high level of understanding of what this technology can achieve and need some help to define the specific use case that may suit them.

This type of scenario is one of our favorites, as it forces us to immerse ourselves in our clients’ business and gives us the opportunity to advise on the adoption of – and hopefully implementation of – AI capabilities at many levels. The golden rule here is to keep expectations clear from day one. On that note, with an old colleague we used to joke and refer to the following equation as the “equation of life”: 

Expectations – Reality = Frustration

The applicability in our case is quite straightforward to explain: if the final result is not as good as the initial expectations, the client will not be happy at the end of the project. How do we manage this? Well, we put a lot of care in the initial proposal. We do our homework researching the client’s business and matching its needs with state of the art papers that may be useful to solve each use case. Then we write everything down as clear as possible (in English, not in “Engineerish”), proposing a realistic final outcome. We make sure to state each and every attention point to take care and how we should proceed to avoid or mitigate them (after checking with the client for requirements) so as to achieve the desired results. 

For more daring clients, we offer the possibility  of working on a special kind of project called “Product Discovery”. If you want to find out more about our commercial strategy you can read more about it here.

To infinity…and beyond!

Let’s suppose we had a couple of meetings with a client, understood what they do and came to an unified vision of how AI could be applied to drive value to their business. Although this already is quite a breakthrough, it is far from being the answer; in many cases it works only as a north. While we know where we would like to get and have many insights that make us believe in our best knowledge that it is something possible to achieve, at this point we cannot ensure the final result. 

In some ways it is not that different from what Cristobal Columbusfaced before starting his famous discovery trip. The Kingdom of Spain needed an alternative route to get to India and other spices lands. Every available information at the time suggested that if you sailed west for many many miles you would get there, but was the sailing technology of the time enough to accomplish that? As we all know, Columbus never arrived at the desired destination, and instead he bumped into some Caribbean islands on the way that shortened his trip. The rest is history. 

In all my time working in the AI world I have never seen these kinds of involuntary fortunate events happen so, if we want to be successful, we need to control as much as we can the multiple factors involved. The way to do that begins by identifying the key factors that may lead to the success or the failure of the project. The approach then is to deal with the more risky ones first, in order to remove as much unpredictability as we can right from the beginning. There is no point in moving forward with the more certain steps if we are to face a giant rock a month away from the final delivery date and only then start thinking how we could get rid of it. Last but not least, keep everyone involved  in the project always on the same page, have agile communication channels inside the development team and with the client and please do not forget to validate the priorities once and again with a business value mindset.

Did someone say data?

One of the main reasons why AI projects fail is poor data quality. Many companies have lots of data they have been storing through years of operations and expect to use in their benefit,  for instance, to forecast sales. The thing is that not every set of data is sufficient to get valuable results. It has to be consistent, not biased, not outdated and, fundamentally, enough to build the required AI solution. If any of these conditions is missing we could end up in a garbage-in-garbage-out situation, leading certainly to a failure of the project. 

Andrew Ng, a man that needs no introduction for those related to the field, promotes a Data Centric AI approach, which in short means building AI systems with quality data with a focus on ensuring that the data clearly conveys what the AI must learn. Take online purchasing data as an example: if we consider the last 3 years as input data to forecast next quarter sales and we do not take into account the possible effect the COVID-19 pandemic may have had in customers behaviors, it is very likely we will miss by a lot on the prediction. So, the more quality data we have, the better, but never leave context out of the equation. 

Regarding data, another situation we have to deal with very often is when a client proposes a project where data does not even exist yet!!! At Marvik we develop end-to-end custom AI solutions and we go as far as required by our clients, and that includes building the dataset from scratch if we need to. These tend to be both the funniest and the more challenging types of projects we face, since we have to define what kind and amount of data will be needed and how the dataset will be collected as part of the project proposal. 

For such a critical matter as which data to use, we have lots of experience in performing comprehensive data analysis, curating the data and applying augmentation techniques if needed. Beside this, we have a fully dedicated labeling team with knowledge of diverse fields of applications in case we need to generate our own dataset.

Being in a cutting edge industry with innovative clients

For the sake of the AI industry, there are a lot of innovative people working in a myriad of sectors and pushing us harder to the edge over and over again. Very often we find ourselves working in solving problems that at most only a dozen other people over the world had faced before. This is very exciting and challenging, and has the extra pressure of having a client expecting tangible results within the agreed deadlines. 

Beside the “no surprises to anyone involved” approach we have already commented, we need to find a way to deliver real value sooner rather than later. In order to do that, we act on some principles that help us stay focused throughout the way. I share some of them below:


  • Less is more: iterative building mindset, trying to keep solutions as simply as we can to fulfill the requirements, diving into more complex ones as the problem demands it (a.k.a don’t use bazookas to kill mosquitos, use bazookas to kill tiranosaurios)
  • Do not reinvent the wheel: this one is tightly related to the previous.  Before starting to build anything be sure no one out there has already done anything like it. As engineers we get tempted to solve the problem by ourselves, but this may not be the best approach for delivering value as fast as we can to our clients.
  • Trust only in data: this is our version of the well known W. Edward Deming quote: “In God we trust, all others must bring data”. We do not base our solutions on sensations about how the data should or expects to behave, but on the real behavior once we get the chance to analyze them in depth. We have found several insights our clients would have not expected when started the project by using this approach. 
  • Always (always) apply common sense: personally I think this is the one that rules all the other principles and gives guidance on how we should act in general. It is pretty simple, but unfortunately if you do not keep it always in mind you are in serious risk of ending up asking yourself this question: Why did we do this? Furthermore, in this particular aspect we humans (still) outbeat machines, so this is where we should come in and make the difference.


What should you remember from this article?

When facing AI projects, uncertainty is part of the game and thus we must learn to manage it to get the best results. Developing an agile mindset and establishing effective communications channels with everyone involved is more than ever crucial for the project to be successful. If you are an AI manager, you must certainly have identified yourself with much of what has been said and hopefully could get a tip or two to apply in your next project. If you are to begin your career as one, you already should have a good idea of what to expect and I hope you are more motivated than ever to take the field.

Here at Marvik we have achieved a success rate way above the average, which we are really proud of. How did we get there? Well, it was a mix of building a first class team, with exceptional technical and human level and a special manner of doing things, which we pass down to every team member since day one. One thing I can tell you for sure is: if the project is “doable” we have one of the best AI crews to get the boat to the desired destination port.

Get in touch with one of our specialists. Let's discover how can we help you.
Training, developing and delivering machine learning models into production