AI Consultant for Choosing the Right Flooring
We built an AI consultant for a flooring manufacturer that doesn't just answer questions — it guides the buyer to the right product, warns about common mistakes, and works out the full budget. With no sales rep involved.
- Client
- Flooring Manufacturer
- Location
- Ukraine
- Timeline
- 3 weeks
- AI Model
- OpenAI
// Problem
What happens when the product is complex and you only have one sales rep
Flooring isn't just "I want wood or tile." Buyers arrive with dozens of questions: will this material work with underfloor heating, how thick should the wear layer be in a hallway, can laminate go in a kitchen, how much extra should I order, how do I work out the full budget.
A qualified rep spent several minutes on every one of those conversations — even though 80% of the questions were identical. Every day. On repeat.
Buyers also made the wrong call on their own: laminate in the kitchen, solid wood over underfloor heating, R9 tile in a shower. Sometimes the rep didn't catch it in time. The result was returns, bad reviews, and damage to the brand.
The catalog is large. Availability varies by region. The rep had to check every time: is it in stock, what's the current price, what are the lead times. And if a request came in on a weekend or overnight, there was no reply at all.
// Solution
What we built in three weeks
The AI consultant isn't one chatbot but two modules working on the same principle: understand the customer's problem, don't just answer the question.
Expert Q&A
Buyers can ask anything: the difference between SPC and WPC, whether laminate can be laid diagonally, which grout to use in a shower. If a question carries the risk of a common mistake, the consultant flags it in the first sentence.
Product Selection
A structured decision tree: two to four questions, and the buyer gets a specific product with an SKU, a price, an explanation of why that option, and the cost of the whole project.
Smart Escalation to a Human
If confidence drops below the threshold, it says so honestly — "let me pass this to a colleague" — and hands over the full context: what has been established, the room type, the budget, the buyer's location.
// Why Not a Standard Bot
We looked at a standard chatbot. And decided against it.
Our first instinct was an FAQ bot: load in the questions and answers, train it on the catalog, ship it. We built a prototype and tested it. It answered correctly about 60% of the time. But accuracy wasn't the real problem.
A bot answers questions. A consultant solves a problem.
The buyer asks: "Which laminate works in a kitchen?" A bot returns a list. A consultant stops and explains that laminate in a kitchen is a mistake whatever the grade — steam and splashes will destroy any laminate within two to three years.
A bot forgets context. A consultant holds the conversation.
Someone mentions underfloor heating in their second message, then asks about adhesive in their sixth — and the bot gives the standard C1 answer, having forgotten the heating entirely. The consultant keeps the whole thread in mind.
A bot quotes a price per square meter. A consultant works out the full budget.
The buyer sees 640 UAH/m² and thinks "that's fine." But they haven't counted underlay, adhesive, installation, skirting boards, and 10% waste. The real number is 1.8× higher. The consultant works it out on the spot, line by line.
A bot doesn't know what's in stock where. A consultant does.
A standard bot recommends whatever is in the catalog and never checks the regional warehouse. We built a four-level geolocation filter. A buyer in Kharkiv will never be shown a product sitting in a Warsaw warehouse.
A bot doesn't know its limits. A consultant escalates.
When a bot doesn't know the answer, it either goes quiet or makes something up. Ours scores its own confidence on every request. Below the threshold, it hands off to a human — honestly, and with the full session context.
// How It Works Inside
Tech Stack and Architecture
LLM: OpenAI
Strong handling of Ukrainian-language content and complex technical explanations. Temperature 0.3 keeps answers predictable and accurate.
Vector DB: Qdrant
Self-hosted, so the data stays in Ukraine. It supports country[] and risk_flags[] metadata directly in the index, which lets us filter by geography at search level.
Product map: JSON in Git
Deterministic logic, with no LLM at the core of the recommendation. Any catalog change is a pull request, not a retraining run. Versioning comes for free.
Confidence scoring
Below 0.70, the answer carries a caveat. Below 0.50, it escalates to a human automatically, with the full session context.
One language, one region
Ukrainian, one market. No unnecessary complexity. The Concept Agency rule: better narrow and reliable than broad and broken.
// Technology
Tech Stack
Every technology here was picked for a specific reason.
Why OpenAI?
We chose OpenAI for its strong handling of Ukrainian-language content and complex technical explanations. Temperature 0.3 keeps answers predictable and accurate, without unnecessary "creativity".
// Key Capabilities
What the Consultant Can Do
24/7 Availability
Handles consultations around the clock — a human is only needed for genuinely complex technical cases.
Mistake Prevention
Flags common mistakes in the first sentence, before the buyer has a chance to choose wrong.
Full Budget Calculation
Material, underlay, installation, skirting boards — line by line, with an explanation for each. No surprises at the end.
Four-Level Geolocation
IP detection, RAG filtering, regional branches in the decision tree, and a warehouse check — buyers only see what they can actually get.
Deterministic Selection
The LLM doesn't guess which flooring to recommend; it explains a recommendation that has already been decided, in plain language.
Smart Escalation
When confidence is low, it passes the full context to a human: what has been established, the room, the budget, the region.
// Results
What changed after launch
| Metric | Before | After |
|---|---|---|
| Time to choose a product | Several minutes with a rep | Seconds with AI |
| After-hours coverage | 0% — nobody available | 100% — the consultant runs 24/7 |
| Share of escalations | 100% of requests went to a rep | ~17% — complex cases only |
| Recommendation errors | Depended on the rep's experience | Deterministic logic — eliminated |
| From request to proposal | Several hours (by email) | Under 2 minutes (automatic) |
Reps stopped spending most of their day on repeat consultations. Buyers get an answer at any hour — a specific product, an explanation of why it's the right one, and a full budget broken out line by line. Wrong orders and returns caused by bad product choices dropped.
Ready to start?
A quick audit to pinpoint where you're losing time and money right now — and whether AI can fix it
