The lighting and electrical industry looks like any other e-commerce industry from the outside . In reality, it's one of the most difficult to automate: hundreds of product lines, nearly identical variant names, technical questions requiring knowledge of standards and datasheets – and B2B customers who expect a quick, precise response to an email sent at 7:30 a.m.
It is in this industry that Cognize built and implemented a working AI agent for Ostrowski Handel Internetowy (OHI), the company that operates the DobreGniazdka.pl, LuxMarket.pl, and OtoLampy.pl brands. This article describes the specifics of automation in this industry and the specific benefits of the implementation.
The lighting industry and AI – where is the real problem?
Companies in the lighting and electrical engineering sectors have several characteristics that make it difficult to scale customer service without compromising quality.
Firstly, the scale and complexity of the catalog. Manufacturers and distributors operate with thousands of SKUs: lamps have unique markings, but electrical accessories (sockets, switches, frames) are sold in series. A series can have dozens of variants, differing in color, IP protection rating, number of modules, and mechanism type. B2B customers order based on documentation or a working designation – and they don't always provide a catalog number.
Secondly, seasonality and volume fluctuations. Construction projects, office renovations, and installation contracts—inquiries are not evenly distributed. During peak season, the sales department receives several hundred emails a week, and it's physically impossible to respond to them as quickly as during the off-season.
Third, technical questions. B2B customers buy based on specifications: IP rating, color temperature, luminous flux, and protection class. A "we have it in stock" response isn't enough—the salesperson needs to know which product actually meets the design requirements.
This combination makes email support in this industry expensive, slow, and difficult to scale. Standard chatbots or simple automated systems respond poorly or not at all.
Don't miss out on more valuable content about the digitalization of B2B trade!
A challenge that stops standard tools
OHI handled approximately 3000 emails per month. Each required manual catalog searches, inventory checks, and the preparation of a quote or technical response. Response time depended on the salesperson's availability and product knowledge—which, given the large and rotating catalog, was not always straightforward.
Standard AI solutions failed to address a specific industry problem: the Berker Q series. Berker Q.1, Q.3, and Q.7 are completely different products – different dimensions, different mechanisms, different prices. But their names are identical. The search engine based on lexical similarity selected products incorrectly, and the costs of error (returns, complaints, lost customers) were too high to accept such a level of error.
The problem wasn't unique to Berker—it's a pattern repeated throughout the electrical equipment industry. A solution was needed that understood the technical nuances, not just minced words.
How does the AI Agent built for OHI work?
Cognize designed a solution based on the Agentic Workflow architecture – not a single language model, but a sequence of specialized steps, each with a single task.
The workflow is as follows:
– Extraction and normalization – The agent reads the email (including messages written in colloquial language, abbreviations or typographical errors) and extracts a list of products with the context of the order.
– Multi-variant search – instead of searching by name, the Agent uses Pinecone (a vector database) for semantic search. This allows Berker Q.1 and Berker Q.3 to be distinguished based on technical parameters, not just the name text.
– Verification – found products are verified for availability via API and MySQL database queries. The agent also notes low stock levels so the salesperson can suggest a replacement or provide information about delivery times.
– Preparing a response – The agent creates a shopping cart in the store, generates a draft message in Gmail with product tiles, prices, and a payment link. The draft is then sent to the salesperson.
Technology stack: n8n as the orchestration layer, Pinecone as the vector base, interchangeable LLM models (no vendor lock), HTML/CSS in email templates. Build time: 6 months – 3 months prototype, 3 months fine-tuning with feedback from the OHI team.
We describe more about the architecture of this type of solutions in the article on quotation automation in B2B e-commerce : Quotation automation in B2B e-commerce : AI tools supporting the sales department
Human-in-the-Loop: Automation that doesn't take away control
One of the key design decisions was the Human-in-the-Loop model. Instead of sending emails themselves, the agent prepares a draft, which the salesperson then verifies and approves with a single click.
This solves several problems simultaneously. First, it reduces the risk of error in an industry where product errors have real costs. Second, it allows the salesperson to add relational context (e.g., a fixed discount for a customer who isn't in the system). Third, it builds team trust in the tool: instead of "AI took over my work," it becomes "AI did the work for me, I'm just checking."
In practice, instead of spending 15 minutes processing a single email, a salesperson now spends 2-3 minutes reviewing a completed draft. Customer response time has been reduced by over 50%.
We describe this Human-in-the-Loop model in more detail in the context of the AI Agent for sales support: AI Agent for B2B Sales Support
Customer Service Agent: handling technical questions on a large scale
In parallel with the Sales Agent, Cognize implemented a second module for OHI – the AI Agent for the Customer Service Office.
Customer Service in the lighting industry primarily receives technical questions: what is the IP rating of this fixture, what are the dimensions of the box, is the lamp compatible with a dimmer, and where is the catalog card? Answering this requires access to the PIM (product information management system) and the ability to interpret technical specifications.
The Customer Service Center agent handles approximately 3000 messages per month. The result: 80% of draft responses are completed without any corrections by the team. Customer Service Center employees handle exceptional cases – complaints, non-standard project questions, and key account support.
You can read about how Agent AI deals with technical questions in B2B e-commerce in a separate article: Technical questions in the Customer Service inbox: how can Agent AI, using PIM data, relieve the team?
Results – what changed after implementation?
The implementation of the AI Agent for OHI has yielded measurable results in several areas:
– 50% faster response to B2B customer emails
– 80% of BOK drafts ready without corrections
– 90% product matching accuracy (including differentiation between Berker Q series variants)
– Standardization of offers – each message has the same format, regardless of which salesperson serves the customer
– Stock visibility – The agent automatically marks low stock levels, eliminating situations where the offer reaches the customer with an unavailable item
– No more browsing through catalogs – salespeople no longer spend time manually searching through the product database
Full implementation case study, with description of the problem and solution architecture: Case study Ostrowski Online Trade AI Assistant
Summary
The lighting and electrical industries place specific demands on automation: distinguishing between similar products, handling technical questions, and scaling without compromising quality. Standard chatbots and simple automation systems struggle to cope with this complexity.
The solution is agent architecture – a multi-step workflow in which each stage has its own responsibility, and the human retains control over the final decision. The implementation for OHI demonstrates that it can be built, launched, and its effects measured – even in one of the most challenging e-commerce industries.
If you run a B2B business with a large product catalog and want to see if a similar solution makes sense for you, schedule a free consultation.
