Foundation · 7 min
What you will learn
- A side-by-side comparison of how each one behaves, costs and fails
- When to choose each, and what to watch out for once you have
- How to combine them so each handles the part it is suited to
- The three mistakes merchants make most often when choosing
The comparison at a glance
The useful differences are not technical. They are what each one costs, how it fails and whether you can explain it to a customer who disputes the result.
| Rules | Automation | Predictive AI | Generative AI | |
|---|---|---|---|---|
| Decides from | Logic you wrote | A sequence you drew | Patterns in past data | Patterns in language or images |
| Same input, same output | Always | Always | Usually, until data shifts | No |
| Can you explain it to a customer | Yes, exactly | Yes, exactly | Only in general terms | Rarely |
| Cost to run | Lowest | Low | Medium | Medium to high |
| Oversight needed | Test on change | Monitor for failures | Monitor accuracy and drift | Review before publishing |
| Fails by | Doing the wrong thing consistently | Stalling or firing twice | Being confidently off | Sounding right while being wrong |
| Use it when | The right answer is known | The sequence is known | The odds help a decision | A person reviews the output |
Read the “fails by” row first. It is the one that decides how much supervision a feature will need once it is live, and it is the row vendors talk about least.
Choosing between them
Rules
Choose it when: The correct action is known and fixed, and someone may have to defend it. Free shipping above a threshold. No checkout on an unavailable item. Manager approval above a refund limit.
Watch out for: Rule sets grow. Fifty interacting rules are harder to reason about than a model, and nobody remembers why rule eleven exists. Review them on a schedule and delete what is dead.
Automation
Choose it when: A known sequence runs on a trigger and currently runs on someone’s time instead. Abandoned-cart recovery, order status updates, low-stock alerts, review requests after delivery.
Watch out for: The failure is silent. A workflow that stops firing, or fires twice, looks like nothing happened. Alert on the absence, not only on errors.
Predictive AI
Choose it when: A decision genuinely improves when you know the odds, and there is enough history to learn from. Fraud scoring, demand forecasting, flagging customers likely to lapse.
Watch out for: A prediction with no decision attached to it is a dashboard, not a feature. Decide in advance what happens at each score band, including who reviews the borderline cases.
Generative AI
Choose it when: Something has to be written or summarized, the source facts are already correct, and a person approves the result. First-draft descriptions, review summaries, alt text, support reply drafts.
Watch out for: Fluency is not accuracy. Treat it as a drafting tool rather than a source of truth, and never let it make a claim about stock, price, policy or legal rights without a check.
Agentic AI
Choose it when: The task needs several steps and the system must choose them, not just suggest them. An assistant that interprets a request, searches the catalog and assembles a cart.
Watch out for: This is the only one that can act in your store, so the controls are the feature. Define what it may do, what needs confirmation, what it must never touch, how actions are logged and how a person takes over.
How they work together
Treat them as layers with different jobs rather than as competing options. A support flow shows the pattern clearly: automation receives the ticket and routes it, AI summarizes the history and drafts a reply, rules hold the limits so no refund above your threshold goes out automatically, and a person approves anything the rules flag as ambiguous.
Each layer does the thing it is good at, and the layer that can cause the most damage is the one bounded by explicit logic. That is a far safer design than asking one system to handle interpretation and limits at the same time.
Three common mis-picks
Reaching for AI when nothing was linked. Merchants buy AI recommendations when related products were never set up, the category structure is confusing, or search returns nothing for common terms. Check the simple causes first, because AI will work around them rather than fix them.
Using AI where the answer must be definite. Tax calculation, permission checks, refund eligibility and order-state transitions need explicit logic. A probabilistic system in a place that requires a definite answer adds variation you then have to explain to a customer.
Letting a rule set grow instead of becoming a workflow. The opposite mistake. When the logic is really a sequence with waits and checks in it, forcing it into stacked rules makes it fragile and unreadable. That is automation wearing the wrong clothes.
Key takeaways
- Rules decide from explicit logic. Automation executes a known sequence. AI infers, predicts, or generates.
- Agentic AI can act rather than only suggest, so its controls are part of the feature.
- Pick by how a feature fails, not by how advanced it sounds.
- Combine them by responsibility and let rules hold the limits around anything AI touches.
- The simplest option that works reliably is the right one.