In short: If a task can be described clearly, classic automation is usually the better choice: faster, cheaper, traceable and the same every time. AI pays off where language, context or unstructured content is involved. The strongest setup is often a combination: rules set the frame, AI handles the ambiguous part, and checks verify the result.
Since AI tools started showing up everywhere, I hear one question more and more often in projects: "Can't we do this with AI?" Sometimes the honest answer is: yes, but it would be the worse solution. In this article I show you how I decide.
How to tell that fixed rules are enough
Classic automation means a script, an import, a rule in the Shopware Rule Builder or a small plugin does the job based on fixed instructions. It is the right choice when:
- the same input must always produce the same result. A price, a shipping method or a stock status must not vary.
- values must be calculated or validated exactly. Taxes, discounts, minimum quantities and required fields are arithmetic, not language.
- every decision must be traceable. If you need to explain later why something happened, a rule helps more than a model.
- there are only a few clearly defined exceptions. Three special cases belong in a condition, not in a prompt.
- an import or an API mapping is enough. Moving data from A to B is mapping, not interpretation.
In all these cases, AI mainly brings new problems: cost per call, varying results and errors that are hard to find.
Example 1: finding missing product images
In a B2B workwear shop with more than 10,000 articles, the client wanted to know where images were missing. The images come from the ERP system, and sometimes the sync does not transfer everything.
That sounds like a job for image recognition. It isn't. The catalogue rule is clear: a variant's first own image is the product image, the second is the model image. Whether a variant has two own images is answered by a single database query.
So I built a check in the Shopware Administration: select it, run it, download the result as CSV. Result in the test shop: of just over 32,000 active variants, more than 1,000 had no own image at all and over 9,600 had only a product image. It takes seconds, costs nothing and returns the same result every time.
Example 2: cross-selling without AI
A second example from the same shop: every article should automatically show complementary items ("goes well with") and alternatives ("similar items").
Again, the answer was in the existing data. The catalogue has parallel category trees for industries and product types. Alternatives share a product type. Complementary items share an industry but belong to a different product type. On top of that comes which articles are ordered together.
The result goes into Shopware's standard cross-selling. Every recommendation can be explained: "same industry, different product type" or "ordered together". A language model would not have made it better or more traceable.
Where AI really helps
AI becomes interesting when a task can no longer be described unambiguously:
- Language: writing product descriptions, harmonising supplier texts, preparing translations.
- Context: recognising that two differently named attributes mean the same thing, such as "100% cotton" and "pure cotton".
- Contradictions: spotting supplier details that don't fit together.
- Unstructured content: extracting information from PDFs, emails or free text.
This is where fixed rules quickly hit their limits, because you would need to know every possible wording in advance.
The strongest solution: rules and AI together
In practice, it is rarely either-or. A good workflow often looks like this:
- Rules prepare. Data is cleaned, required fields are checked, unsuitable cases are filtered out.
- AI handles the ambiguous part. For example the text, the mapping or the summary.
- Rules check the result. Are standards, sizes and mandatory details correct? Are banned phrases in there?
- A human decides on anything unusual. Not every record, but every one that stands out.
This keeps AI inside a clearly limited frame, and you get its strengths without its typical weaknesses.
Checklist: rule or AI?
- Can the task be described in if-then sentences? → Rule
- Must the result be exact and identical every time? → Rule
- Is it about language, meaning or free text? → Consider AI
- Are there thousands of ways to say the same thing? → Consider AI
- Are errors expensive or legally relevant? → Rules as a check, even if AI is involved
Frequently asked questions
Isn't AI more flexible than fixed rules?
Yes, but flexibility is not always what you want. For prices, taxes or stock levels you need reliability, not creativity.
What does AI cost in comparison?
Once built, a rule costs practically nothing per run. With cloud AI you pay per request, with local AI for hardware and power. With AI you also always need checks.
How do I find out what makes sense for me?
Write the task down as precisely as you would explain it to a new employee. If you can do that with clear rules, you don't need AI. If you find yourself writing "it depends", AI is worth a look.