AI Fundamentals
The vocabulary you need
AI in an automation context means using models that produce predictions, classifications, or generated text from data. Key concepts: training (fitting a model to examples), inference (using a trained model to produce an output), model (the learned artifact), prompt (the instruction given to a generative model), and confidence/score (how strongly a model favors an output). Distinguish classification (choosing among fixed labels) from generation (producing free text) from extraction (pulling structured fields out of unstructured input). Also know hallucination - a fluent but unsupported output - and human-in-the-loop, where a person reviews model output before it causes an action.
Traps to watch for
Expect questions that treat model output as fact. A generative model does not "look up" an answer; it produces plausible text, so it can be confidently wrong. Another common trap is confusing training with inference: retraining changes the model, while inference just uses it. Watch for the claim that a higher confidence score means the answer is correct - confidence reflects the model's internal preference, not ground truth. Finally, questions may imply AI is required for every automation; most automations are purely rule-based and need no model at all.
Rule of thumb
Use AI where the input is unstructured (free text, images) and the output can tolerate review; use plain rules where the input is structured and the logic is exact. Worked example: routing support emails by topic - a classifier can label each message, but a confidence threshold should send low-confidence items to a human queue rather than guessing. That threshold-plus-review pattern is the safe default and a frequent correct answer.