Reasons Why Some A.I. Projects Fail


At the end of 2019, it would have been a fairly common occurrence to stumble upon a tech article like this ZDNet one, that used statements such as, “AI jobs are on the upswing, as are the capabilities of AI systems.” Fast forward about 6 months and this statement is still accurate in some ways, but highly conditional in others. As COVID-19 continues to act as a threat to the US and global economy, businesses are being forced to reevaluate not only their current and prospective AI projects, but also who they hire and who they retain to help execute these projects.

Many business leaders will likely feel pressure to quickly finalize AI projects and bring them to market, but in the process they may succumb to one or more of the major pitfalls outlined below, rendering their results invalid and even potentially harmful to the public.

Bad data

Data should often be the starting point for your AI project, as it will come to represent the underlying fuel for your market offering (platform, solution, etc.) to thrive. It is important though to carefully choose the right data because the early decisions made around data can make or break an AI project. The data examples we typically define as “bad,” tend to garner this label for one of the following reasons:

Issue with size

More often than not, having a data set on the smaller side relative to other projects is cause for concern. AI platforms may still perform necessary functions with this data, but the accuracy and statistical significance of their results will face certain limitations from the beginning.

Issue with source

Having a body of data with a problematic source can spell out disaster as well, such an example could be a source that is not available at the time of prediction for models (simple yet commonly occurring problem). Careful assessment of potential sources is essential prior to integrating them into machine learning models and other AI frameworks.

Issue with environmental context

During periods of unexpected change, like the COVID-19 pandemic period, environmental context becomes very important to the creation, deployment, and ongoing management of AI solutions. Failing to account for this factor can jeopardize AI project outcomes. Gary Marcus, a cognitive scientist and professor at NYU, was recently quoted for saying, “top algorithms are left flat-footed when data they’ve trained on no longer represents the world we live in.”

Not having the right data science team

To properly launch your AI project, you need to strike a balance through the right mix of technical creators (system designers, product developers, etc.), data scientists and engineers and relevant C-level staff. Recruitment of talent is an essential prerequisite to success, but so is the management of teams throughout the project. We can look at each level to gain understanding, but we also need to look at how teams inter-collide and work together throughout the project.

Senior leaders (management)

Whether the AI project is at the forefront of business operations, or an enticing side note, it is important to achieve a sufficient level of approval from the core executives at any business. While it may seem natural to say that a failed system was caused first and foremost by failures of its system designers, oftentimes the blame comes just as much from above. According to a recent survey by O’Reilly Media, the biggest obstacle to AI’s adoption was reportedly a lack of institutional support. Beyond conveying their support, in mindset and in financing, leadership needs to also display support by fostering the appropriate level of team collaboration. It is also their responsibility to spot potential organizational silos and help prevent them from undermining projects.

I have seen this firsthand and it can be as simple as lack of communication between the execs and the AI teams, which then renders the finished AI projects useless as the execs were not on the same page to begin with, and adoption of those projects is non-existent.

Shooting for the moon outright

It is important to avoid getting stuck on moonshot projects as a starting point. As the technology and its target markets continue to mature, most businesses entering AI (or already in it) no longer need to heed the siren call of moonshot projects. Many businesses can instead find success in fine-tuning existing solutions or deploying new, practical solutions that solve problems within arm’s reach. In this current period, where we face a major world health crisis AND a major economic recession, we need to be even more wary of the potential desire and risk of moonshot AI projects aimed to resolve these two issues.

One example of this issue, for example, is creating a machine learning model that predicts how many first responders to allocate to each region. This would be immediately applicable and helpful today. A moonshot example of that would involve mobilizing drones to collect up-to-minute data, satellite imagery, and weather data as well as being integrated and finding some reason to use crypto, 3d printing or GAN’s, just because they are hot right now. All this to maybe increase the accuracy of a prediction from 84% to 85%. This would tremendously increase the cost and time for the project and decrease its probability of ever becoming a reality. Going from 0% to 84% is a much bigger deal and should be pursued first.

Platforms are there to help

In the world we live in, anything that can be easily integrated into someone’s workflow to help facilitate AI projects is something to be considered. Platforms that can reduce the time needed, the cost, the frictions, and most, if not all, of the problems listed above should be top of mind for executives wanting to get a return on their AI investment.


* indicates required

Comments are closed.