CampusLearn
← BIT 402 Artificial Intelligence in Business

Deploying AI and Ethical Considerations

Notes

From experiment to business service Deploying artificial intelligence means integrating a trained model into an operational decision process. A deployment plan specifies input validation, prediction interfaces, response times, access controls and fallback procedures. Version datasets, model artefacts and configuration so decisions can be investigated and releases reproduced. Monitor both technical reliability and business outcomes because an available service may still make poor predictions. Data drift is a change in input distributions, while concept drift changes relationships between inputs and outcomes. Retraining should follow an approved evaluation process rather than automatically replacing a model whenever new records arrive. Accountability and responsible decisions Responsible AI requires attention to fairness, privacy, transparency and meaningful human oversight. Historical records may encode unequal treatment, so compare errors and outcomes across relevant groups while protecting sensitive information. Explainability provides understandable reasons for predictions, but a plausible explanation does not prove a decision is correct. Collect only necessary data, restrict retention and provide channels for contesting significant decisions. High-impact applications need clear accountability, documented risk assessment and a route to human review. Commercial performance should be assessed alongside potential harms, applicable legal requirements and the needs of affected people. Practical example A Thika SACCO pilots a loan recommendation model on 200 applications, split equally between two groups for fairness evaluation. It recommends approval for 70 of 100 applicants in the first group and 50 of 100 in the second. Approval rates are 70% and 50%, a difference of 20 percentage points. This difference signals a need for investigation, not automatic proof of unlawful discrimination. Analysts compare repayment risk, data quality and false rejection rates before release. Loan officers review adverse recommendations and record reasons, while a weekly dashboard tracks appeals, overrides and subsequent repayment outcomes.

Quiz

1. What follows from the 20 percentage point approval gap?

2. Which deployment control supports investigation of past decisions?

3. What is concept drift?

Sign in to submit and securely grade your quiz.