KromAsistan: AI-Assisted Part Selection in Metal Surface Treatment

In metal surface treatment, a customer calling to say "I want to polish my part" has told you almost nothing. The material, the geometry, the finish expected from the surface, the daily volume, the existing line layout — only the answers to those questions lead to the right machine and the right process. That knowledge usually sits in a few people's heads, and those people are on the phone.

KromAsistan is the AI-assisted part selection assistant MountainTech+ built for Kromaş to run that selection process. This article covers the problem, how the solution was put together, and what it suggests to any manufacturer considering something similar.

The problem: the knowledge exists, in the wrong form#

Kromaş has a broad and technical product range. The issue was never missing product knowledge — it was there in abundance, in catalogues, in technical documents and in the experience of the staff. The issue was that it was not held in a reachable form.

That has three concrete consequences:

• The same questions repeat. The technical team spends a meaningful share of the day answering opening questions whose answers are already known.
• A catalogue alone is not enough. A visitor can see the products but cannot tell which one is right for their part. A catalogue offers a list, not a decision.
• First contact is tied to office hours. The site gets visited at night and in the evening; with nobody there to respond, a qualified enquiry quietly disappears.

This is a very common pattern in manufacturing: an entire technical selection process running on a handful of people's phone lines.

Why an off-the-shelf chatbot is not enough#

The first solution that comes to mind is bolting on a ready-made chat widget. Those bots move through predefined flows: "press this button, pick that option." The moment the conversation leaves the flow, they stall.

Part selection is exactly the kind of work that leaves the flow. A user asks in technical language and in everyday language alike; sometimes they state the material up front, sometimes they remember it in the third message; more often than not they change direction while describing their own requirement. A fixed decision tree cannot follow that.

So KromAsistan was built not as a flow bot but as a GPT-based assistant. It holds context across the conversation, asks for the missing information itself, and treats the product range as a knowledge base rather than a menu.

How KromAsistan works#

The assistant walks the user towards an answer step by step.

1. It understands the requirement. The user writes what they are trying to do in their own words; the assistant asks for what is missing — material, expected finish, capacity — inside the conversation.
2. It matches against the range. The information gathered is evaluated against Kromaş's machine and process range, and the user is pointed to the suitable option.
3. It states its reasoning. The user does not just get a product name, they get why that product was suggested. In a technical selection, the reasoning matters as much as the recommendation.
4. It hands the conversation over. When a question falls outside the assistant's scope, or the user simply wants to speak to a person, it passes to the team along with everything gathered so far — the user does not have to start again.

That fourth step was the project's most important design decision. The point of the assistant is not to cut the technical team out, but to leave them a prepared conversation instead of a cold one.

How the build ran#

We ran the project in the same four steps we run every project in.

Process analysis. First we listened to how the existing selection process actually worked: which questions come in, what information is needed to reach an answer, where people get pointed the wrong way. That is where the step worth automating was identified.

Preparing the content sources. The quality of a digital assistant is the quality of the content feeding it. Product documentation, scope boundaries and business rules were compiled into a form the assistant could read. This is also the step that really determines how long a deployment takes; with content ready, a typical deployment runs two to four weeks.

Configuration and testing. The assistant was configured to Kromaş's language, scope and limits. A large share of the testing was about not what the assistant says but what it does not say: that it produces no confident answer to something it does not know, stays inside its scope, and escalates to a person when it should.

Deployment and monitoring. Adding it to the site is a one-line embed. The real work comes afterwards: seeing where the assistant struggles in real conversations and updating the content and the rules accordingly.

What changed#

After the assistant went live, what changed was not the amount of work but its distribution. Repetitive opening questions stay with the assistant; the conversations that reach the technical team are the ones that genuinely need technical knowledge. On the visitor's side, the site stops being a shop window and becomes a surface that answers: a question arriving outside office hours does not go unanswered.

The real gain from this kind of deployment is a shorter path to a decision and fewer misdirections. Both are measurable — which is why it is worth tracking not only how many conversations an assistant handles, but how many of them turn into a qualified enquiry.

Where else does the same shape fit?#

What makes KromAsistan work is not metal surface treatment, it is that the selection requires expertise. The same pattern repeats when:

• A manufacturer's range is broad and technical, and customers get lost in the catalogue.
• The right product depends on parameters: dimensions, material, capacity, regulation.
• A B2B site gets mostly the same pre-sales questions, but the answers do not fit into one standard paragraph.

And it does not fit when the selection process has clear rules and few exceptions — there, the answer is a filter interface. If plain software solves the problem, we do not propose AI.

Frequently asked questions#

What happens if the assistant recommends the wrong product?#

The way to reduce that risk is to keep the scope tight. The assistant only speaks within the product range it was configured for, and where it is unsure it escalates to a person rather than guessing. Most of the testing phase goes into verifying that behaviour.

Does the assistant stay current when the range changes?#

The assistant is fed from a knowledge base; when a new product or a changed rule is written into that source, the assistant updates with it. No retraining is involved.

Where is our data stored?#

The content the assistant reads and the conversation logs are held in an environment reserved for the organisation. Where it is needed, an on-premise installation is also possible.

Does it work in more than one language?#

Yes. The assistant is multilingual, Turkish and English first, and detects the visitor's language to reply in kind. For an exporting manufacturer that means a second channel opened by a single deployment.

---

If your range is technical and your customers struggle to choose within it, the same setup can be built for your process. Look at our digital assistant solution or get in touch directly.