1. Determine AI Ownership and Responsibilities
The unpredictability of AI’s results means that teams must work together to determine which team is responsible for which task, and that communication flows between the teams. In essence, problems start to happen in the gaps between departments; for instance, the team that has governance over the system and the team that deals with security are often operating in silos.
A fix for this is simply open communication and establishing ownership of roles and responsibilities. Identify the individuals who will have to report and answer questions if, for instance, the output of the AI system is unexpected, if the system is being misused or manipulated, or if its use strays from what it was designed to do. The sooner these ownership questions are answered, the better the project’s chance of success.
2. Ensure Data Governance Is Being Properly Implemented
The old computing adage “garbage in, garbage out” definitely applies to AI. If the data that the system learns on isn’t of high quality, results can be unreliable. Other issues can crop up if data governance isn't properly implemented; for example, the system can be left vulnerable to cyberattacks, it can run up against regulatory restrictions, or the AI itself can degrade over time.
Data governance problems pop up when organizations don't take a holistic approach to making sure data is sound. Accountability and oversight need to be identified, and human factors need to be part of how the data is maintained. If multiple teams in an organization aren't included in the data governance equation, there will be information missing in the data, and that can lead to those problematic results.
Integrate AI Into the Workflow Instead of Making It an Add-On
AI needs human interaction to guide it, refine its output and fix any mistakes that it makes. Making AI oversight a task that’s handed to a team without the team knowing the system that they're working on is a recipe for failure.
In cases where AI implementations were successful, oversight teams, whether in-house or outsourced, have knowledge of the entire system. They are trained to know how to react when the AI makes mistakes or outputs information that strays from the standards that have been established by the organization. If AI is integrated into this workflow, problems can be corrected efficiently because of the oversight team's knowledge of the system and its communication with the teams that implemented it.
