The AI Toolbox, Unpacked
"AI" gets thrown around as if it's one thing. It's not. It's a grab bag of different tools, each with different strengths, limitations, and price tags. Here's a quick reference card.
Roughly ordered from simple/predictable to complex/probabilistic.
Rule-based Automation (sometimes called RPA - Robotic Process Automation) If-then logic. "When a support ticket contains 'refund request,' route it to the billing team." Or: "When an order exceeds €10,000, flag it for manager approval." Same input, same output, every time. Best for well-defined processes where inputs are already structured (form fields, database entries, spreadsheet rows) and rules are clear.
Operations Research / Mathematical Optimization Finding the mathematically best answer given constraints. Route planning, shift scheduling, inventory allocation. Old-school (1950s+), battle-tested, provably optimal.
Classical Statistics / Regression Finding patterns in historical data to predict numbers. Demand forecasting, pricing, trend analysis. Highly explainable ("sales rise 12% when temperature drops below 10°C"). Your finance team can audit it.
Classical Machine Learning Algorithms that learn from labeled examples. Show it 10,000 "fraud" and "not fraud" transactions, it learns to spot the difference. Good for classification and segmentation. Garbage in, garbage out.
Deep Learning Neural networks. Especially powerful for images, audio, and sensor data—visual inspection, speech recognition, anomaly detection. Needs lots of data. Less explainable.
Traditional NLP Purpose-built systems for specific language tasks—extracting names, dates, amounts from documents. Can be more reliable and efficient than LLMs for narrow, well-defined extraction.
Large Language Models (LLMs) The ChatGPT/Claude category. General-purpose reasoning about text. Summarization, drafting, Q&A, flexible extraction, translation. Probabilistic—same input can give different outputs. Can hallucinate. Best with human review.
Agentic AI LLMs with tools and autonomy to execute multi-step workflows. The newest category. Powerful, but still earning its stripes in production.
These tools sit on a spectrum from deterministic to probabilistic, narrow to general. The right choice depends on your problem. Sometimes that's an LLM. Sometimes it's a simple optimization model or plain automation. Knowing the landscape helps you ask better questions.
