Section 1
Three different machines wearing one label
Plain-English versions, in the order they were invented. Automation is instructions a person wrote. If A happens, do B. It has no opinion and no memory. Machine learning is a statistical estimate built from past examples. Show it ten thousand past orders and it will estimate the chance this one is fraudulent. It has no rules, only patterns. Generative AI produces new text, code or structure from an input it has never seen. It is the most flexible and the least predictable of the three. For the ground-level definition of how these get assembled into working systems, see [What Is AI Automation? A Plain-English Guide for Founders](/blog/what-is-ai-automation-a-plain-english-guide-for-founders).
Section 2
Rules: cheap, brittle, and honest about it
Rule-based automation is underrated. It is cheap to build, it costs almost nothing to run, and you can read the logic and know exactly what it will do. Auditors like it. Regulators like it. Its weakness is that it only handles the cases you anticipated. Reality supplies cases you did not. A rule set grows exceptions until nobody understands it, which is how companies end up with a workflow tool holding four hundred branches that one departed employee designed. The compensating strength is that it fails visibly. It stops, or it errors, and someone notices. That is worth more than it sounds. [Robotic Process Automation (RPA) vs. AI Automation: What's the Difference?](/blog/robotic-process-automation-rpa-vs-ai-automation-whats-the-difference) goes deeper on this trade.
Section 3
Machine learning: accurate until the world moves
A trained model is a compressed summary of your history. That is its power and its expiry date. It performs well while next quarter resembles last year, and it degrades quietly when your customer mix, pricing or market changes. The industry term is drift. The business version is that your demand forecast was excellent right up until you launched in a new city. Machine learning also needs labelled outcomes. If you want a model that predicts which leads close, you need a clean record of which leads closed, including the ones your team never bothered to update. Most companies discover at this point that the constraint is not the model. It is the CRM.
Section 4
Generative AI: fluent, and most dangerous when confident
A generative model produces plausible output whether or not it has the facts. It does not know the difference between recalling and inventing, so it cannot warn you which one it just did. Fluency is uncorrelated with accuracy, and humans read fluency as competence. That makes it excellent for drafting, summarising, classifying and reformatting, where a person reviews the result anyway. It makes it hazardous for anything where the output goes straight to a customer or into a ledger without a check. The design question is not whether the model is good. It is whether a wrong answer has somewhere to be caught.
Section 5
Who carries the failure
In the NIST framing, risk sits under trustworthiness, design, evaluation and use, and the three technologies land in different places on that map. A broken rule is a bug with an owner. A drifting model is a maintenance obligation somebody has to schedule. A confident wrong answer from a generative system is a content problem that reaches a customer at full speed. Practically: put a review step wherever the output is generative and consequential. Schedule a check on any predictive model that informs money decisions. Keep the rules readable. And name a human owner per system, not per project, because projects end and systems keep running. [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation) covers how to communicate these changes internally.
Section 6
What to ask before you sign
Four questions cut through most sales conversations. Which of the three is doing the work in the part I care about? What happens when it is wrong, and will I find out? Does it need my historical data, and in what shape? What does it cost at ten times my current volume, and is that priced per seat, per run or per token? The answers sort vendors quickly. A company that can explain where rules end and the model begins is a company that has thought about your risk. One that answers AI to every question has not.