Three thousand emails a month. Each one requires opening, reading, catalog searching, inventory checking, and handwriting. Some are simple availability inquiries. Others are complex requests for quotes with lists of dozens of items. All go to the same people, who are also responsible for answering phones, attending meetings, and maintaining relationships with existing clients.
At some point, you're faced with the choice of whether to hire more people for your sales department. And if not, what should you do instead?
This is where an AI agent comes in. It reads emails, searches for products in the catalog, and writes a ready-made response before the salesperson even opens the message. What previously took dozens of minutes can now be ready in just a few seconds.
In this article, we show how offer automation based on the AI Assistant works – using the example of an implementation we completed for the DobreGniazdka.pl, LuxMarket.pl and OtoLampy.pl brands.
When does B2B bidding automation become a necessity?
There's a certain scale threshold beyond which manually handling requests for quotes simply isn't profitable—both financially and operationally. This can't be pinpointed with a single number, as it depends on the complexity of the catalog, the number of active customers, and the number of orders requiring individual pricing.
But there are signals that usually appear together: salespeople complain that they spend most of their day on emails instead of selling. Response times are slowing down. Some inquiries are getting "lost." You hire more people to handle them, but the volume grows faster than the team.
In the case of Ostrowski Handel Internetowy, that threshold was precisely the aforementioned 3000 messages per month. Each required manual analysis and manual response. Our task wasn't to replace the team, but to relieve them of this most repetitive part of the work.
How does an AI agent work in your email inbox?
The AI agent we built for Ostrowski Handel Internetowy runs directly within Gmail. Salespeople don't need to learn any new system—they work exactly as before, except that when they open an incoming email with a quote request, a ready-made response is already waiting for them.
The mechanism in short looks like this:
- Intention analysis. The agent reads the email and identifies what the customer wants: whether they're asking about specific products, about order status, filing a complaint, or have a technical question. This is the first stage of classification – it determines what the agent does next.
- Product identification. Customers rarely provide a specific product code. They're more likely to say, "Do you have Berker sockets in stock?" or use industry jargon. The AI agent must understand this and match it to the catalog. We'll describe how we technically solved this in the next section, as it was the most challenging part of the entire implementation.
- Live verification. The system connects to the store via API and MySQL and checks inventory in real time. If the product is in stock, it's added to the offer. If not, the AI Assistant will suggest a substitute or notify the salesperson before responding.
- Creating a basket and draft. The agent physically creates a finished shopping cart in the store's system, and a ready-made email draft appears in Gmail—complete with product pages, photos, and a direct payment link. The salesperson then reviews the email, adds a discount if necessary, or makes any manual corrections, and sends it with a single click.
This entire process happens in a matter of seconds.
Don't miss out on more valuable content about the digitalization of B2B trade!
How does the agent know which product you have in mind?
This was the biggest technical problem in the entire implementation and it is worth dwelling on it because it shows what you need to be prepared for when automating offers in B2B eCommerce with an extensive catalog.
The electrical equipment industry has its own unique characteristics. While lamps have unique names, socket series can differ only slightly. For example, Berker Q.1, Q.3, and Q.7 are completely different products, but their names are identical. Standard AI models selected products incorrectly. This wasn't because they were "wrong"—they simply weren't built with such precision in mind.
The solution was the so-called Agentic Workflow – instead of a single query to the AI, we divided the process into several stages:
- Extraction and normalization: The agent first extracts only the product features from the email (series, color, type) and translates the customer's everyday language into store terminology.
- Multi-variant search: The system repeatedly searches the vector database (Pinecone), taking a broad sample of potential matches.
- Verification: The final stage compares the results with the strict specifications from the email and eliminates incorrect selections.
The result: product selection accuracy is 90%. This isn't 100%, and that's something we need to be clear about, which is why the model must include a human factor.
Human in the Loop – why is there a human here at all?
We intentionally didn't build a system that automatically sends emails without any verification. This was a conscious design decision.
The AI agent makes mistakes—rare, but possible. In bidding, a mistake can be costly: sending a bid with the wrong product or price is a problem not for the system, but for the customer relationship. Therefore, each draft response first goes to the salesperson—as a draft in Gmail.
The salesperson doesn't have to search for products, check stock levels, or write a response. They just need to review what the AI Assistant has prepared, add a discount if necessary, or clarify any details, and click "send." This is the Human-in-the-Loop model – automation that reduces workload but doesn't take over control.
An additional element is internal notes from the agent – before the salesperson sends an offer, the system informs them, for example, about low stock levels of a given product. Something that previously required additional checking now automatically appears as context alongside the message.
What does this actually change in the work of the sales department?
The results of implementing the AI Assistant for Ostrowski Handel Internetowy clearly show that:
- Email response time reduced by over 50% – this is the real value provided by the Client;
- According to the OHI Team report – over 80% of emails prepared by the Agent are ready for sending without corrections.
- The system handles approximately 3000 queries per month – without the need to increase employment
- No more manual browsing through directories – the salesperson no longer has to search for products in the catalog with each email.
- The team has standardized answers – Every offer is delivered in the same format – professional, with product pages and a direct link to the shopping cart. Regardless of which merchant sends it.
It's also worth mentioning what this system doesn't do: it doesn't build customer relationships, it doesn't conduct negotiations, and it doesn't resolve difficult sales situations. Salespeople are still needed, but their time, freed from repetitive tasks, can be spent on something truly needed.
When should you start thinking about an AI Assistant for your sales department?
There is no single answer, but there are questions that help assess readiness.
Do you have sufficient query volume? An AI agent begins to pay off when you have a regular influx of repetitive email queries—several hundred or more per month. Below this threshold, implementation time may not pay off in a reasonable time.
Do you have structured product data? The agent needs something to work with. If your product catalog is incomplete, unstructured, or scattered across various systems, it's worth organizing it first. Without a good database, the AI agent will generate errors instead of saving time.
Is your team ready to verify, not create? This is a shift in workflow. Salespeople stop writing responses and start verifying them. For most people, this is a positive change, but it's worth discussing it with the team first.
Is email your primary customer contact point? The basic version of the AI Assistant we built works through email. If quote requests primarily come in by phone or through an integrated B2B platform with its own dashboard, that's a different problem to solve.
If you'd like to see what this could look like for your company, schedule a free consultation . Before we implement anything, we want to understand your processes and whether Agent AI is the right tool for you.
Summary
B2B eCommerce quotation automation isn't about replacing salespeople. It's about freeing up the time salespeople spend searching for products in catalogs and writing standard responses to focus on things that truly matter.
The AI agent we implemented for Ostrowski Handel Internetowy handles 3000 inquiries per month and has reduced the time it takes to respond to requests for quotes by more than half – without hiring new staff or requiring the team to learn a new tool. It works where salespeople already work: in their email inbox.
Implementation details can be found in the case study on our website . If you manage a similar volume of inquiries and would like to discuss whether this approach makes sense for you, please fill out the contact form below.
