5.3 When Machine Learning Adds Value

Microsoft's AB-731 bullet is identify scenarios when machine learning adds value. It sits next to generative AI in the same skill area because leaders now default to a chatbot for every problem. That default is expensive and often wrong. Artificial intelligence (AI) is the broad idea of machines performing tasks associated with human intelligence. Machine learning (ML) is a subset of AI: systems that improve a predictive task by training on examples. Generative AI is a subset of AI that uses techniques such as deep learning to generate new content — text, images, or audio — usually from massive pretrained models plus your prompts. You do not need to train a neural network on this exam. You do need to tell a steering committee when to fund an ML model instead of (or in addition to) Copilot or a Foundry chat agent.

Microsoft's Azure Machine Learning overview of the learning process is still the right leadership picture: feed data into an algorithm (often with features the business already understands), train a model, test and deploy it, then consume predictions — call the deployed model and use the scores it returns. That loop is built for stable, evaluable outputs: a class label, a number, a rank, a probability. Generative systems are built for open-ended language. Both can create business value. They fail in different ways.

The ML jobs that show up in real functions

Classification

Classification assigns an item to a category: spam versus not spam, invoice type, support-ticket queue, likely-to-churn versus not. A contact-center director should fund classification when thousands of tickets must land in the right team in seconds and you have labeled history. A generative assistant can draft the reply after routing; it is a weaker primary router if you need a measurable accuracy rate on a closed set of queues.

Forecasting

Forecasting predicts a future numeric series: units demanded next week, call volume next Monday, cash collections next month. Operations and finance live on these numbers for purchasing and staffing. The output must be a stable numeric forecast with an error metric (for example mean absolute error) that a planner can budget against. Asking a chatbot to "guess next week's orders" from a paragraph is not a forecast process. Seasonality, promotions, and lead times belong in an ML or statistical forecasting design.

Anomaly detection

Anomaly detection flags points that do not match the learned normal range: fraud-like transactions, a spike in factory vibration, unusual badge-swipe patterns. Time-series anomaly models take timestamps and numeric metrics and return whether a point looks abnormal — a classic ML job. Security and operations VPs want precision (not drowning in false alarms) and a playbook for what a human does when the flag fires. Generating a poetic description of the shift is not anomaly detection.

Ranking

Ranking orders items by predicted usefulness: search results, product recommendations, next-best offer, which resume to show a recruiter first. Retail and HR feel ranking quality immediately. Generative AI can explain why an item was suggested. The core ranker is usually an ML model (or a hybrid of lexical plus vector relevance, as in Azure AI Search) because you need comparable scores across candidates, not a one-off essay.

Structured prediction

Structured prediction fills fields or graphs: named entities on an invoice (vendor, amount, due date), medical codes from a note, parts on a claim. Microsoft discusses named-entity recognition as turning text into a pre-specified class that can land in a schema. That is ML (or a specialized Language feature in Foundry Tools), not "write me a paragraph about this invoice." Generative extraction can prototype the idea; production finance still wants field-level precision, recall, and an exception queue.

When generative AI is the wrong primary tool

Two exam-ready tests:

1. You need a calibrated probability. Credit pricing, medical triage thresholds, and inspection sampling often need a number that means "about 12% chance," not a fluent paragraph. Language models predict next tokens. Their "confidence" is not automatically a calibrated probability on your business event. You can ask a model to emit a number; that does not make the number a validated risk score. Use a predictive model (logistic regression, gradient boosting, and similar families — names you need only at a conceptual level) with a proper evaluation set, then optionally let generative AI explain the drivers to a human.

2. You need a stable numeric forecast. Planning systems break if Tuesday's "forecast" jumps because the prompt changed, the retrieved context changed, or the foundation model version changed. ML forecasting with versioned data and a frozen model (until you retrain) is the tool for replenishment. Generative AI can still write the commentary that sits next to the forecast in a board pack.

Other mismatch signals: the gold output is a single label from a known list; regulators will ask for a reproducible score; you have lots of labeled rows and almost no need for open-ended prose; latency and cost must be milliseconds and fractions of a cent per prediction, not a long reasoning trace.

When generative AI is the right tool (so you do not over-rotate to ML)

Drafting, summarizing, translating, answering questions over documents (retrieval-augmented generation), and assisting with code or meeting notes are generative strengths. If the user input is messy language and the valuable output is also language, start with Copilot, Copilot Studio, or Foundry agents — with the security bar from the previous sections. Do not stand up a custom classifier to "detect the vibe of this email" if a grounded Copilot answer plus a human send button is the actual job.

Hybrids leaders should recognize

Many high-value systems are two-stage:

• ML (or Search) retrieves and ranks; generative AI writes.
• ML scores risk; generative AI drafts the analyst narrative; a human approves.
• A Language feature extracts fields; generative AI explains exceptions to accounts payable.

Scenario. A CFO wants "AI" to set next quarter's inventory buy. If the decision is a purchase quantity, fund forecasting ML with a defined error budget and a planner in the loop. If the decision is a narrative for the board about why demand shifted, fund grounded generative AI on the same numbers. Funding only a chatbot for both jobs usually delivers a confident paragraph and a bad purchase order.

Leader questions that pick the tool

• What is the output shape: paragraph, class, number, rank, or filled form?
• Is there labeled or sensor history that matches that shape?
• Must the number be calibrated or stable under a frozen version?
• What error is acceptable, and who is accountable when it is wrong?
• Do we need language at all, or are we decorating a predictive job with a chatbot because chatbots are fashionable?

AB-731 will not ask you to pick PyTorch versus scikit-learn. It will ask whether you recognized an ML-shaped business need. If you can name classification, forecasting, anomaly detection, ranking, and structured prediction — and you know when a generative model is the wrong instrument — you have the skill.