Another reason why AI projects fail is that the primary business problem wasn’t an AI one.
Let’s look at an e-commerce example. Let’s say an AI-based recommendation engine has been deployed and it hasn’t made a significant difference to revenue from the site.
What went wrong? It’s possible that the business doesn’t understand the shopper’s journey. If they had tracked and studied the majority of shoppers’ journeys, they might have discovered that the majority of the shoppers were purchasing additional items because they were using the store’s excellent search engine rather than the product’s page or even a recommendation engine.
It might make sense here to have a project that focuses on further improving the search engine, which will provide a better return on investment.
This isn’t a failure of AI per se, but in this instance, the primary route to success was not an AI-based solution.
It is not uncommon for product owners and data scientists to believe that AI projects hold the key to digital transformation. If the AI team continues to deploy AI products that don’t work well, this can lead to major issues within an organization.
For example, the models in production are degrading and acting strangely. Perhaps the AI team isn’t mature enough, and the models aren’t robust enough to deal with real-world data. Deep-learning AI models are not always simple to understand.
After a series of failures, organizations may return to legacy systems, where at least if something went wrong, the system could be troubleshot.
The management team’s confidence and trust in AI, in general, may dwindle. The issue is that this may cause a company to miss out on the powerful impact that AI can have on their business.
How do we get around this?
Set correct expectations. Understand that some failure in AI projects is to be expected. Pick projects that are small, well-defined, and in line with the team’s ability to deliver. And as the AI team grows in experience and confidence, pick more complex and challenging projects.
And finally, ensure that there are enough data science resources to monitor and manage models after going into production.