Last year, I took an “AI for Business” course delivered by Wharton Online (via Coursera). The course covered foundational concepts in AI, machine learning, business applications of AI, and governance considerations.

The course reinforced something important:
Technical performance is only one dimension of AI deployment. Equally critical are decisions about data quality, optimisation trade-offs, explainability, and regulatory exposure

Equally critical are decisions about data quality, optimisation trade-offs, explainability, and regulatory exposure.

1. AI Has a Data Dependency Problem

Machine learning depends on trained data. New initiatives often face a paradox:

You need data to improve the model, but you need users to generate the data first.

Possible approaches include:

• Partnering with organisations that already have data
• Purchasing data through marketplaces
• Starting with rule-based systems before moving to ML

What struck me is this: AI initiatives often fail not because of weak models but because of weak data foundations.

Business implication:
Before investing in AI, companies should assess data readiness. AI is as much a data strategy decision as it is a technology decision.

2. AI Can Amplify Bias

AI systems learn from historical data. If past decisions contained bias (e.g., in hiring or lending), the model may replicate or even amplify those biases.

For example:

• A hiring screening model may favour certain gender or racial groups if historical hiring data reflects past imbalance.

Mitigation approaches include:

• Assessing how representative the training data is
• Adjusting weights for underrepresented groups
• Implementing an algorithm audit process that reviews (i) Input data quality and bias, (ii) Model fit and stress testing, and (iii) Output fairness and explainability

Business Implication:

AI governance must go beyond performance metrics. Companies need structured audit processes to ensure fairness and accountability.

3. “Choose Your Pain” — Optimisation Means Trade-offs

AI models are built to optimise specific objectives.

They inevitably create trade-offs, particularly between false positives and false negatives.

A classic example is fraud detection in banking:

• False positive (Type I error): Flagging a legitimate transaction as fraud (false alarm), which impacts customer experience
• False negative (Type II error): Missing an actual fraud case, which increases financial loss

Businesses must decide what they want to optimise (e.g., Customer experience or risk control?)

Business Implication:

Leaders must clearly align the model objective with business priorities and understand the consequences of different error types.

4. Explainability vs. Precision

Customers, regulators, and stakeholders may demand explanations for AI-driven decisions (e.g., loan rejections). In some jurisdictions (such as the EU), regulations increasingly require transparency and the “right to explanation.”

Highly complex models may deliver better predictive performance — but are often harder to explain.

In some cases, businesses may need to sacrifice some precision for better interpretability (e.g., by using simpler decision trees rather than black-box models).

Business Implication

To create compliance and maintain trust, organizations should understand key drivers behind model outputs (not treat AI purely as a black box).

5. Data Privacy and Governance

Data privacy regulations such as GDPR introduce additional complexity:

• Right to be forgotten
• Opt-in consent requirements
• Restrictions on cross-use of data

These constraints can affect how models are trained, updated, and deployed.

There are also practical challenges:

• Should trained data be deleted if a user withdraws consent?
• How does cross-training data across multiple models affect compliance?

Business Implication

AI initiatives must be designed with privacy architecture in mind from the beginning. Legal and compliance considerations are not afterthoughts, but they shape model feasibility and long-term scalability.

Final Reflection

One of the biggest takeaways from this course is that AI is rarely just a technical project.

It is a combination of:

• Data strategy
• Risk management and Regulatory awareness
• Clear business prioritisation

Deploying AI effectively requires aligning model objectives with business objectives — and understanding the trade-offs, limitations, and risks that come with it.