---
title: "Borgh now answers every product question across 15,000+ SKUs in seconds"
description: "How What's Next built Borgh an AI agent that answers any product question across 15,000+ Milwaukee and Makita SKUs in seconds."
canonical: "https://www.whatsnext-ai.com/cases/borgh-answers-every-product-question-in-seconds-across-15000-skus"
published: "2026-08-28T00:00:00.000Z"
updated: "2026-09-11T15:42:12.712Z"
client: "Borgh"
---

# Borgh now answers every product question across 15,000+ SKUs in seconds

How What's Next built Borgh an AI agent that answers any product question across 15,000+ Milwaukee and Makita SKUs in seconds.

![Construction worker in a steel-framed building holding a tape measure, wearing a hard hat and tool belt](https://fgzcpjbyiakhjifaciaj.supabase.co/storage/v1/object/public/media/borgh-construction-worker.jpg)

## Overview

Borgh is the fastening and tool specialist for construction, industry and installation, active since 1978 with 200+ people across Europe and South America. In a branch, a rep can pick the right tool for a job in a minute, but online that same question fell into a slower email loop through a specialist. What's Next built an owned AI recommendation agent that answers any product question across 15,000+ Milwaukee and Makita SKUs in seconds, in stock.

## Details

- Industry: Tool & Fastening Distribution
- Region: Europe & South America
- Company size: 200+ employees
- Timeline: 3 months

## Challenge

A professional tiler messages Borgh: "I need an M18 SDS drill that can hammer and chisel to strip bathroom tiles." That is a precise question. Brand, product line, two functional requirements, and it has to be in stock. Get one part wrong and the answer is worse than useless, because the person on the other end is standing on a job, not browsing.

Borgh has answered questions like that well since 1978. They are the fastening and tool specialist for construction, industry and installation, 200+ people active across Europe and South America, and their whole reputation rests on advising customers toward the right solution. In a branch, a rep does that in a minute. Online, that same question fell into a slower loop: a message to the team, an email to a product specialist, a back-and-forth with follow-up, and the answer relayed back, sometimes only the next working day.

Handing this to an off-the-shelf chatbot was never an option. Across an assortment of 15,000+ Milwaukee and Makita SKUs, a generic assistant invents part numbers, recommends phased-out or sold-out machines, and cannot tell a bare tool from a full set. For a professional buyer, a confidently wrong part number is a risk to the brand, not a convenience.

Leadership wanted four questions answered. Can we give every online customer the same "right product for the job" advice our reps give in person, without hiring extra advisors? Can an AI handle 15,000+ SKUs across two brands that structure their data completely differently, without ever inventing a part number or recommending something we do not have in stock? How much of the advice work that now runs through the team and a specialist could instead be answered directly? And can Borgh lead with AI in its market, instead of watching a competitor do it first?

Left alone, the very thing Borgh sells on, expert advice, stayed locked behind office hours and a manual specialist loop the moment a customer went online.

## Approach

This started as an investigation, not a build project.

Step 1: Map the data honestly. We looked at everything available: the daily XML product feed, the live stock link, and Milwaukee's sector content on nl.milwaukeetool.eu. The finding that shaped everything: Milwaukee and Makita structure their product data in incompatible ways. SKU formats do not match. Category trees do not line up. One brand says "impact driver", the other uses different terminology. And every machine breaks into variants (bare tool, with case, full set with batteries and charger). The result was a clear map of where the data was reliable (brand, price, stock) and where it was messy (descriptions, specs, categories).

Step 2: Define what "correct" actually means. Together with Borgh we pinned down what a right recommendation is: correct brand, correct line (M12 vs. M18), the stated functional requirements, in stock, and never an accessory, which Borgh as a strict rule does not sell here. These became rules the system has to enforce structurally, not politely request.

Step 3: Deliberately break the obvious approaches. Pure semantic search returned Makita when the customer said Milwaukee, and a €450 machine against a €250 budget, because the descriptions looked alike. Hard category filtering up front missed products, because the two brands file the same item under different categories. Letting the model write freely produced output the webshop could not reliably process. Every failure proved the same point: this called for an architecture where each layer does one job and fails cleanly, not a clever prompt.

Step 4: Scope the phased build. We agreed that the first thing built would be the recommendation agent that answers "the right product for the job", designed as the foundation for a broader e-commerce agent later, with observability and governance built in from day one.

The research half was not optional. It is the reason the build half never had to guess.

## Solution

What's Next built Borgh an agent that behaves like their best advisor. When a question is specific, it recommends straight away. When it is vague ("I need a drill"), it first asks one targeted question, brand, category or budget, exactly as a rep would. It gives at most three real, in-stock options, each with the SKU, the key specs, the price and who it is for. Underneath sits a stack of four layers with one guiding principle: the AI never improvises, it builds queries against Borgh's own assortment.

An agent that clarifies first: the model does not chat, it decides. Parse the question, clarify or search, build the conditions, retrieve, filter, format. That imposed structure keeps it accurate across a four or five message conversation, without losing the customer's earlier budget or brand.

Hybrid retrieval that respects the messy data: SQL filters handle the precise properties (brand, price, stock). Semantic search over embeddings handles the fuzzy ones (descriptions, specs). It searches narrowly within a category first and automatically widens if it finds fewer than three matches. That two-step approach solves the category mismatch between the brands instead of tripping over it.

Guardrails built into the structure, not the wording: telling a model "do not recommend accessories" does not hold. So filtering out accessories, the three-option limit and the SKU-first format are enforced in the architecture. The agent can only return codes that exist in the live assortment.

An assortment that maintains itself: a daily delta sync from Magento re-embeds only what changed, and the live stock link makes sure nothing sold out is ever shown. The advice is always current without anyone having to touch it.

Full observability and governance: every conversation is versioned and traced in LangFuse, with automated checks on every change (Is it relevant? SKU-first? At most three options? Accessories filtered out?). That is regression testing for AI: adjust the prompt, and see exactly what breaks before customers do.

Here is what the economics of that look like, per conversation:

ApproachHuman time per questionCost per conversationCost at ~1,500 conversations/year*Manual specialist loop~15 min across 2 people~€10-25~€15,000-37,500Borgh's own stack (gpt-5-mini + infra)none~€0.005~€8

*The volume figure is illustrative (roughly 3 to 5 conversations a day) and gets replaced by Borgh's own number. Working math: manual = ~15 min blended across the people involved at a loaded ~€60/hr. Even at a conservative €5 manual, the owned stack is 1,000× cheaper per conversation.

The deeper point is not price. It is ownership. The guided-selling SaaS tools in this category (Neocom, Rep AI, Zoovu and similar) charge a fixed monthly subscription whether you use them or not, and scale against you as your assortment and traffic grow. Borgh's agent is an asset Borgh owns: their data, their embeddings, their prompts, their traces. If What's Next disappeared tomorrow, Borgh would keep a working system, not a cancelled subscription.

The precision is deliberate too. The agent is built to ask when a question is genuinely ambiguous, and to decline and refer when a question is off-topic, rather than bluff. For a professional buyer, a clarifying question beats a confidently wrong answer every time.

And every conversation adds up. Borgh now sees what customers actually ask, where the assortment data is thin, and which products get compared against each other. That is the foundation for phase two, the broader e-commerce agent.

## Outcome

A professional buyer asking for "an M18 SDS drill that can hammer and chisel to strip bathroom tiles" now gets three real, in-stock options, with the right specs and SKUs, in seconds, at any hour, instead of the old loop of a message to the team, an email to a specialist, and a back-and-forth with follow-up until the answer came back, sometimes only the next working day. Each answer costs Borgh about half a cent.

Operationally, the advice that used to sit behind office hours and a specialist queue now runs on its own, correctly, around the clock. The team is freed from routine "which one do I need" questions to pick up the genuinely complex ones.

Strategically, Borgh can now offer its personal advice standard online, 24/7, without a single extra advisor, and it owns a manageable, transparent AI system it can extend rather than rent. Leadership bet on leading with AI in its market, instead of watching. This is the proof that they can build production AI that clears a professional buyer's accuracy bar, not a demo that looks good in a screenshot.

From research to live production in three months. The research half is the reason the build half never had to guess, and both were equally essential to deliver an agent a professional buyer can actually trust.

## Results

- **1,000×** cheaper per conversation than manual specialist advice (~€0.005 vs. manual)
- **Seconds** to a correct in-stock recommendation, down from a back-and-forth supplier email loop
- **24/7** specialist advice across 15,000+ SKUs, same team, no extra advisors

## Quote

> What's Next knows their craft, and at the same time they really get inside how we work, as if they are on the shop floor with us. Breaking the project into smaller deliveries gave us visible results fast. And something I also really like: this AI genuinely talks like a Borgh person, not like a generic tool, but with Borgh's own standards and values.
> — Remco Nagtegaal, Marketing Manager, Borgh