Drifting from the original is an innate part of working with AI. Sometimes, the consequences are incidental, like a mislabelled photo or a wrongly-targeted ad. But when stakes are high, such as hiring or mortgage decisions, a miscalculation can have severe consequences.

In the last couple of years, consumer behavior has changed dramatically. This resulted in wildly inept AI predictions, such as:

• Instacart’s AI model accuracy for predicting product availability in stores dropped from 93% to 61% due to changes in shopper habits.
• The AI model did not have the correct data during training, resulting in poor decision-making.
• The data was biased, and the human team is not diverse enough; so, the AI outcomes are skewed.
• There’s no long-term data-management in place to adjust data sets to real-world behavior.
• There’s a lack of standards and governance whereby the AI can explain the outcomes.
• Business leaders are not committed to the AI project or are ill-qualified to monitor it.
• There’s no transparency or collaboration between the different teams using the AI system.
• People using the system lack the skills and competencies to work with AI.

Human-AI systems require more than basic health checks. The co-founder and CEO of Evidently AI notes that for human and AI collaboration to succeed, all stakeholders have to know how to work with the AI.³ She lists what question each stakeholder should ask:

• Data scientist: is the model drifting?
• User: is the prediction trustworthy?
• Data science manager: should the AI be retrained?
• Business user: how much value is the AI bringing?
• Product manager: what are the model’s limitations?
• Compliance manager: is the model safe?

Enterprises have to monitor their human-AI systems at both functional and operational levels and business leaders have to institute best practices for monitoring:

• Data
• Model
• Predictions or other outputs
• Alerts
• Log

Further, they have to set in place practices such as:

• Building a culture of open data.
• Establishing a cross-functional team for monitoring human-AI systems.
• Keeping the monitoring process lean by centralizing tools and decentralizing people.
• Initiating the monitoring process as early as the design phase to ensure more accuracy.
• Keeping the focus on your employees and what will benefit them during decision-making.
• Training your workforce to use AI, how to train it and how to document their usage.

Business leaders need to know that deployment is not the final step of the AI project. They need clarity on how to monitor their human-AI systems post-production.