In my previous post (Intro to ML #6), I worked through the data exploration, feature engineering, and model-building steps of an employee attrition case at a pharmaceutical company.
This post picks up where that left off: once you have a prediction model, how do you actually use it? And what are the risks you need to think through before you do?
Two Very Different Approaches to the Same Problem
When a company faces high employee attrition, there are two fundamentally different ways to respond using data.
① The analytical approach
Aggregate the data, look for patterns, form hypotheses about root causes, and roll out organization-wide initiatives. “Junior sales staff have high overtime → reduce overtime and add headcount.” “Female managers are leaving due to childcare pressures → expand parental leave and flexible working options.” This produces uniform interventions aimed at population segments.
② The AI prediction approach
Build a model that scores each individual employee’s likelihood of leaving — and direct resources specifically toward high-risk individuals. “Employee A has a 78% predicted attrition probability → flag for manager, schedule a 1-on-1, discuss a role change.” This enables intervention at the level of individual names.
Neither is inherently better. They suit different situations.
Three Advantages of the AI Prediction Approach
What the AI prediction approach offers that traditional analysis cannot is scale — in three dimensions:
- Generalizability: Once a model is built, it runs at consistent quality regardless of who’s using it. No dependence on a skilled analyst’s intuition.
- Real-time responsiveness: As input data updates, predictions update. You can detect when an individual’s attrition score spikes month-over-month — automatically.
- Economies of scale: More employees, more interventions, more data — the value compounds without proportional cost increase. Delivered as a SaaS product, the model becomes even more powerful.
A quote attributed to Jeff Bezos in 1998 appeared in class: “If we have 4.5 million customers, we shouldn’t have one store. We should have 4.5 million stores.” AI-powered personalization at scale is exactly this idea applied to prediction.
The Risk: AI Is Not Neutral
But “AI said so” is not the same as “objective.” This was the part of the session I found most thought-provoking.
Bias enters AI models through two routes:
- Algorithm design: Deciding which variables to include is a human judgment — and those choices carry assumptions.
- Data: If historical data reflects past discrimination or structural inequality, the model learns and reproduces that inequality at scale.
The Amazon recruiting tool case is a well-documented example. Amazon built a hiring model trained on historical hiring decisions — decisions made in a male-dominated environment. The model never used gender as a direct input, but it learned to penalize resumes that contained signals correlated with being female (certain word choices, patterns in educational background). The result was a model that systematically disadvantaged women — and looked objective because a machine produced it.
AI learns from the past. If the past was biased, the model will reproduce that bias — and the appearance of algorithmic objectivity can make that bias harder to see, not easier.
The Ethics Problem: Cambridge Analytica
For a larger-scale example of AI prediction misused, the Cambridge Analytica case was covered in class.
During the 2016 US presidential election, a data analytics firm harvested the profile data of approximately 50 million Facebook users (without proper consent) and used it to build psychological profiles. Those profiles were then used to deliver individually customized political messaging — designed to shift behavior at the individual level.
This is, technically, the same capability as “AI-adjusted insurance pricing” or “attrition prediction for HR.” The underlying mechanism — predict individual behavior, act on the prediction — is identical. What differs is the purpose.
What’s technically possible and what’s ethically permissible are two different questions. That gap needs to be part of every conversation about deploying predictive models on people.
What the Research Says: AI and Employment
The course also introduced a 2025 working paper from Stanford and NBER by Brynjolfsson, Chandar, and Chen: “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.”
Key findings:
- Since ChatGPT’s release, employment among young workers (ages 22–25) in AI-exposed occupations (software developers, customer service, etc.) has declined by nearly 20%
- Experienced workers in the same occupations have seen relatively smaller impacts
- Workers in low-AI-exposure occupations (e.g., nursing aides) have seen stable or growing employment
The “canaries in the coal mine” framing is deliberate: early-career workers in AI-adjacent roles are showing up as the first warning signal of a broader labor market shift.
I’ve been in the IT industry since 1997. These numbers aren’t abstract to me. Infrastructure work is increasingly something AI can assist with or automate. At the same time, the value of people who can frame the right prediction problem, interpret results in business context, and communicate findings to non-technical stakeholders is going up. That’s what this entire course series has been building toward.
Putting It All Together: Posts #1–#7
Looking back across this series:
- #1 (Titanic): Conceptual intro to ML — classification, confusion matrix, overfitting
- #2–#3 (Social lending): Business application of classification — thresholds, recall, precision
- #4–#5 (Auto insurance): Business application of regression — predicting continuous values
- #6–#7 (Pharma attrition): Solo end-to-end practice — analysis vs. AI prediction, bias, ethics
The through-line: being an AI-capable business leader isn’t about writing code or knowing the latest algorithms. It’s about asking the right prediction question, making ethical judgments about how to act on the output, and explaining it clearly to people who weren’t in the room when the model was built.
→ [Intro to ML #8 — coming soon]