Photo: free stock photography (Unsplash licence) — see imprint
The clear winner: product content at scale
If you have 30,000 SKUs with a spec table and no prose, an LLM generating first-draft descriptions from structured attributes is an immediate, measurable win. The key word is draft — a human still approves, and the model never invents a technical value.
The quiet winner: support triage
Not a chatbot answering customers — a classifier reading incoming emails and routing them, with the order already looked up and attached. It removes a genuinely boring hour from someone's day and fails safely: a misrouted ticket costs a minute.
Where RAG earns its keep
Retrieval-augmented generation over your own datasheets, manuals and past quotes lets sales answer a technical question in seconds instead of an afternoon. The retrieval quality — not the model — decides whether this works.
What we would not build yet
An autonomous agent that changes prices or places orders. The failure mode is not an embarrassing sentence — it is a wrong invoice sent to 200 customers. Keep AI on the draft side of anything that touches money.
- Drafts, not decisions — especially near money.
- Product content and support triage pay back fastest.
- In RAG, retrieval quality beats model choice.
Frequently asked questions
Start where a mistake is cheap and the work is boring. First-draft product copy from your attribute data, or a classifier that routes incoming support mail. Both sit next to a human who approves the result, so a bad output costs a correction, not a customer. Anything that decides on its own can wait until the boring things are running.
It depends on volume, not company size. If you have thousands of articles without prose, or a person spending an hour a day sorting mail, the arithmetic works. If you have 200 products and two support mails a week, a person is faster and cheaper than any pipeline, and we will tell you so rather than build one.
Technically yes, and that is exactly why people underestimate it. A misrouted ticket costs a minute; a confidently wrong answer sent to a customer costs trust and sometimes a return. Triage fails safely, direct answering does not. If you do want an answering assistant, it needs source citations and permission to say it does not know.
Because the failure mode changes character. A weak sentence in a product text is embarrassing; a wrong price applied across a category is an invoice problem with 200 customers attached, and you find out days later. Near money, stock and tax, keep the model on the draft side and let a person press the button.
We do this for a living — Shopware, Node.js, React, ERP integration and automation for B2B.
Talk to an engineer