Exponential Risk: Why AI/ML Projects are Different

There are two important take aways from this article:

1. If an algorithm is part of your project, you need us on your team to get it right.

2. Risk associated with AI/ML projects can sound similar to other types of development projects but algorithmic-thinking systems have exponential risk. Meaning, small issues turn into big problems, fast.

Tl;dr If your project includes an algorithm, you need us at the table. Contact Us now.

It is important to distinguish between different types of development projects because while all have risk, they are exponential in AI/ML projects. There is much more at stake to get it right the first time, making it essential to have our expertise at the table as early as possible in ideation and development.

A large portion of the project failures we've seen have originated by teams treating AI/ML just like software development or automation projects.

Three types of projects:

Software Development - building applications for users to engage and perform various tasks. These are classic development projects and while they could be involved in an AI/ML project they aren't always. The most common component within an AI/ML project is building an interface for a user to leverage the model directly without having to code.

Process Automation - these projects take a work task and codify it so it can be performed by a computer rather than a person. Often, these projects include explicit instructions and logic flows that the program follows to complete a task. You know exactly what the program will do in every possible situation.

It doesn't mean there aren't unintended outcomes or that these can't be complex. The key is that the program is explicitly provided the instructions to make a decision. It is a judgment call when you start to call that "Intelligence" verses just automation. Most of these bots are now called AI Agents.

Just like software development, this type of project is also included to some degree in AI/ML projects.

 Algorithmic thinking - this brings us to the final type of project and the one HSD is most focused on impacting. These projects leverage at least one algorithm as part of its decision flow. You can think about it like concentric rings, most all projects will have the other two types of development and a subset will include an algorithm.

Keep in mind that most platforms (ie. Email providers, CDP, Web Analytics) will have algorithms built into different modules. So, it is possible for algorithmic-thinking projects to be completed by business users or without any of the other types of developers. The algorithm component could be entirely black box or highly transparent. It may have been built in-house or could be built into a platform you are implementing (ie. Microsoft Copilot or Salesforce's Einstein).

 Software: this team is usually referred to as Developers.

Automation: this team is usually referred to as Engineers.

Algorithms: this team is usually referred to as Data Scientists. Like the one who started HSD.

Projects that involve algorithmic thinking are unique to other types of technology and automated systems because all of the outputs compound over time. Small issues can become large very quickly. Slightly nuanced inadequacies can quickly become glaring errors. That's why we see our work more as Quality Control than anything else. We are interested in the seemingly tiny defects in the system that grow to large fractures over time.

What do I mean by exponential? Let's say your system produces 2 bad decisions a day. In a traditional system, these add up every day. So by day 8 you have 16 bad recommendations that have occurred and the 2 per day continues indefinitely. If your system makes 500 recommendations a day, it's easy to write this off as acceptable error. It may even be a massive improvement upon human error.

This same approach could be quickly fatal in an algorithmic system where new recommendations are not only part of the loop but are also sometimes weighted more heavily within the algorithm. Here, we started day 1 with 2 bad recommendations. By day 8 we have not 16 but 256 bad recommendations. You are one week after launch and already your system is entirely unusable.

That is the nature of exponential risk. Contact Us to see how we can help.