---
name: Machine Learning & AI for Business
description: Decide whether machine learning or generative AI actually fits a business problem, then frame, evaluate, and deploy it responsibly — with the right metric, honest baselines, and a human on the irreversible decisions — instead of adding a model because it is fashionable.
audience: product manager · growth lead · founder · analyst
---

# Machine Learning & AI for Business

## What this is
A method for applying ML and generative AI to business problems: choosing the problem framing (predict / classify / rank / generate), selecting an honest evaluation metric tied to the decision, and deploying with monitoring — or deciding, correctly, that a simpler rule wins.

## What this is NOT
- **Not "add AI to it."** The first question is whether ML beats a transparent baseline; often it does not, and the skill says so.
- **Not a research/MLOps replacement.** It frames and evaluates; training large models, building feature stores, and production serving at scale route to ML engineering.
- **Not a way to launder bias or dodge accountability.** Fairness, explainability, and a human decision-maker on consequential outcomes are requirements, not extras (EU AI Act, SR 11-7 where financial).

## When to use
Deciding if a problem is an ML problem; picking a metric; scoping a churn/propensity/recommendation model; evaluating a vendor model; introducing generative AI into a workflow with the right guardrails.

## Operating principle
The metric is the decision. Choose the evaluation measure from the business cost of being wrong (false positive vs false negative), always compare against a dumb baseline, and keep a human on any irreversible or high-impact call.

## Capabilities
- **Problem framing & baseline** — Goal: right problem, honest bar. Method: translate the business question to a task, define the label, set a transparent baseline (heuristic/last-value/majority-class), name the cost of each error type. Output: a framed problem + baseline number to beat. Quality bar: if the model can't clearly beat the baseline, the recommendation is "don't ship the model."
- **Evaluation & fairness** — Goal: know if it actually works. Method: decision-tied metric (precision/recall/AUC/MAE, or calibration), holdout/backtest, slice performance across segments for disparate impact, check drift. Output: an evaluation card with the metric, slices, and failure modes. Quality bar: performance is reported per segment, not just in aggregate; degradation is named.
- **Generative AI in the workflow** — Goal: useful, safe LLM use. Method: scope the task, ground with retrieval where facts matter, set human-review gates, label AI-generated output, watch cost/latency. Output: a workflow design with gates + a disclosure/guardrail note. Quality bar: hallucination risk is stated; the human signs the consequential output.

## A worked example
"Predict churn and auto-discount." → Frame as propensity; baseline = "flag anyone inactive 30 days" (recall 0.55); the model reaches recall 0.71 at similar precision — a real gain, so it ships to *rank* outreach, but the discount decision stays human (the auto-money-movement is the red line). Slice check shows weaker recall on new users, disclosed.

## Guardrails & escalation
Consequential automated decisions (credit, pricing, eligibility) → human-in-the-loop + governance (EU AI Act Art. 14, SR 11-7). Training/serving at scale → ML engineering. Predictions feeding marketing spend → pair with the Marketing Analytics Science skill. Never present model output as certainty.

## References
Standard ML evaluation practice; EU AI Act (Art. 12/14/50); NIST AI RMF; SR 11-7 for financial models; C2PA for AI-content provenance. Verify against current model capabilities and regulation.