AI product advisor
Code decides, AI phrases.
Your customers type freely — the advisor finds the exact variant by model, colour and size, or asks a targeted question. Deterministic code makes the decision. A language model only phrases the answer, and a check layer verifies every wording before it appears in the shop.
Customers name products differently from how the catalogue lists them.
Type as if you were a customer.
An invented outdoor catalogue with 11 models and 133 variants — built with the same traps real shop exports contain. On the right you can see what the code decided before any text gets written.
Ask for a product — or pick an example.
Who decides makes the difference.
Both answer free-text requests. They differ in who picks the variant — and whether that can be verified.
Four decisions, no guessing.
Every request ends in exactly one of four cases. Plus two capabilities that real input needs.
Match
Model, colour and size are unambiguous: the advisor names the exact variant with a stable product ID that can be linked directly.
Size missing
Model and colour are known: a follow-up with exactly the sizes available in that colour. If there is only one, the question is skipped.
Colour missing
The model is known: a follow-up with that model’s colours. If the customer names a colour family such as “black”, the choice narrows to the matching shades.
Model ambiguous
“Backpack” fits several models: the advisor offers the most likely ones instead of guessing. Short words must match exactly — “Karabiner” does not hit “Kar”.
Typos & four languages
Typos via a dedicated similarity function, model words from German, English, Spanish and French (“Damen”, “women”, “femme”).
Several turns
Context carries across the follow-ups: model → colour → size. If the customer names a different model along the way, the new input wins.
What the check layer verifies.
In the pilot a language model phrases the answer — it only receives the finished result and cannot change it. Before the text reaches the customer, code checks every wording. If it fails, a fixed template answers.
Match: product name present
For a match, the full product title must appear in the text.
No invented sizes
Every number in a size question must be one of the sizes offered.
Required details per case
Size question: model and colour. Colour question: the model. Ambiguity: at least one of the options offered.
No truncated sentences
At most 600 characters, and the text ends with punctuation. That catches answers the model cuts off mid-word.
Failure means template
If the model does not answer — outage, timeout, quota — the fixed template comes in the customer’s language. The answer is never empty.
Visible who answered
The history shows whether the model or the template answered — the basis for the evaluation in the pilot.
Two scopes, both at a fixed price.
The fixed price is agreed in writing after the first conversation. All amounts exclude VAT. The prerequisite is a catalogue export (CSV or JSON) — the prototype needs no customer data. Live operation and hosting are not included; handover is code plus documentation.
The matcher on your catalogue, with tests and a usable demo — answers from fixed templates.
Adds a language model that phrases the answers in your shop languages — each one checked before it appears.
From catalogue export to advisor.
The catalogue decides, not the technology. That is why the analysis comes first.
Catalogue export
You send an export with titles, variants and status. The prototype needs nothing more.
Findings list
What gets in the way of recognition: missing IDs, stray spaces, confusable names, mixed size systems and languages. The list also helps with catalogue upkeep.
Matcher & tests
The matcher is adapted to your catalogue, test cases built from realistic customer input secure it, and the demo runs on a test URL.
Handover
Code, tests and documentation. In the pilot additionally: language model with check layer, integration on your staging and an evaluation of the follow-up questions.
Frequently asked.
Limits included — so you know when this is the wrong fit.
Why not just use a language model?
What happens if the model fails or writes nonsense?
Which shop system does this work with?
What data do you need — and what about data protection?
What can the advisor not do?
Is this already running in a shop?
How is this different from the rule-based website assistant?
More on the topic.
AI in the shop, chatbots and the question of when a language model is really needed.
AI Chatbots & Voice Agents
When a language model really pays off in support — and when rules are enough.
Read more
Agentic AI Workflows 2026
Where autonomous AI agents make the difference — and where they are overkill.
Read more
Transparent IT Consulting: Packages & Pricing
Fixed prices for audits, prototyping and automation — openly calculated.
Read moreWhere do your customers stumble over product names?
In 20 minutes we look at how your catalogue is structured and whether a product advisor pays off — or whether your shop search is enough.