Windows and doors are considered purchases. A question about glazing, dimensions or installation can decide what happens next. At MartijnKozijn, Robby helps visitors find their way through the company’s own information, with access to a member of the team when a question needs personal attention.
THE IMPLEMENTATION IN NUMBERS
- active conversations
- 1,562
- AI replies
- 3,946
- ready knowledge sources
- 22
1 August – 29 October 2026. Customer channels; no dashboard or prompt tests. Knowledge sources: inventory at snapshot. How we count ↗
01
A broad range of products needs a clear explanation
MartijnKozijn supplies PVC windows, doors and sliding doors to homeowners and business customers. Stock products sit alongside made-to-measure options. A visit that starts with a particular design can quickly turn into questions about specifications, delivery or installation. The information may already exist, but visitors do not always know where to look first.
That creates a useful role for an assistant. Comparing two products is a different task from finding the right contact for an existing order. Helpful support starts by recognising the context. Sometimes an explanation is enough. Sometimes the next step is a product reference, a specific knowledge article or a person who can look at the situation.
The setup starts with that practical distinction. Robby is there to help visitors move forward in their own buying process. He does not need to recreate a showroom appointment. He needs to explain the available information clearly and point to the right next step when the conversation reaches the limits of that information.
02
Built around MartijnKozijn’s own knowledge
The company’s own information provides the foundation. At the verified snapshot, Robby had 22 processed knowledge sources linked to him. That gives the assistant a defined starting point for questions about the business and its products. This is an inventory at the time of measurement, rather than a count of sources used during the reporting period.
The configuration emphasises information that can be checked. Product details should belong to the product being discussed; practical conditions should come from the relevant source. If a detail is missing, a focused follow-up question or a useful reference is better than an assumption. That matters when technical properties vary between different doors and windows.
The tone is configured for a purchase that deserves care: Dutch, clear and polite. Technical terms have a place when they help explain a choice, but the visitor should still understand what the answer means for their home or project. A useful answer connects the specification to the question, rather than simply repeating a product description.
03
From a product question to a useful next step
Consider a visitor looking at a door who wants to understand which details matter when choosing a version. Robby’s role is to clarify the question and consult the available product information. The visitor can then continue with a more specific question or the appropriate page. This describes the configured workflow; it is not a quotation from a customer conversation.
Practical questions also continue after the first product comparison. A visitor may need to find delivery information, installation guidance or the right contact route for an issue that needs individual attention. An assistant can bring that information closer to the question without suggesting that every request can be completed automatically.
In this setup, Robby provides advice and navigation. He does not change shopping carts or place orders. That boundary helps make expectations clear. Explaining a product and helping someone find information remain distinct from carrying out a transaction on the customer’s behalf.
04
1,562 conversations, recorded in Corthex
From 1 August to the snapshot on 29 October 2026, Corthex recorded 1,562 active conversations for Robby on customer channels. During the same period, 3,946 AI replies were recorded. Dashboard conversations and prompt tests are excluded. These are conversation activity counts, rather than orders or a count of verified individual customers.
The figures make the implementation tangible. They show an assistant providing answers across a substantial set of conversations, beyond a single demonstration. They do not, by themselves, prove that every question was resolved. For that reason, this story does not turn reply counts into a satisfaction score, revenue uplift or a percentage of successfully resolved questions.
Taken together, the activity and the knowledge setup show how Robby is being used: as another entry point into MartijnKozijn’s information. Reviewing conversations can help the team identify explanations that need work, missing sources and situations where personal attention remains important. Those observations can guide the next improvement to the assistant.
05
Personal service stays part of the experience
Small details can matter a great deal when buying windows and doors. An unusual situation, a service request or uncertainty about an order may need a member of the team. On its public Robby page, MartijnKozijn explicitly tells visitors they can also reach an employee through the chat. The assistant sits within that wider service experience.
The conversation does not have to remain automated at any cost. Existing knowledge can support explanations, while employees retain room to assess situations that go beyond it. A clear question and relevant context give personal contact a better starting point when the visitor needs additional help.
The lesson for other webshops selling products that need advice is practical: start with the information visitors need, connect it to the places where questions arise and keep the route to the team easy to find. That is the basis for an assistant that fits the business and the decisions its customers are trying to make.
About the figures in this story
Source: Corthex analytics, verified snapshot on 29 October 2026 at 18:20 (Europe/Amsterdam). Counts are limited to this assistant and customer channels; dashboard and prompt-test channels are excluded. The final day is partial.
An active conversation contains a visitor or AI message within the period. AI replies are recorded assistant messages, excluding employee messages. Knowledge sources are processed, linked sources at the snapshot. Conversations are not unique customers; channel filtering does not exclude every possible test on a website. These counts do not measure verified resolution, revenue or satisfaction.
The described workflow was checked against the assistant configuration. Examples are illustrative; no customer transcripts or quotations are published.
Business information and imagery: MartijnKozijn ↗