Gamrat is a Polish manufacturer of pipes, hoses, and rainwater systems, based in Jasło. Its sales team prepares 30 to 50 offers a day. A complex offer can take two to three days because a salesperson has to bring together current prices, product availability, and delivery details before replying.
In June 2026, Soba Labs ran a corporate AI training for twelve members of Gamrat’s sales team, building the AI literacy they needed to use AI in their day-to-day work. That made it easier to draft messages, follow-ups, and other customer communication. It also exposed the limit of a general assistant: better wording does not help much when the team is still spending days finding and checking the numbers behind an offer.
Soba Labs designed a proposed assistant to address that work. It is not deployed, and the figures below are pilot targets rather than achieved results. This case study explains the business problem, how the assistant could fit into the sales process, and what a pilot should measure.
The business problem
Preparing an offer at Gamrat means resolving several connected questions before a salesperson can make a promise to a customer:
- Which products and quantities does the customer mean, especially when the request describes a need rather than using a catalogue name?
- What are the current prices for products whose costs depend on PE and PVC raw materials and related clauses?
- Is each item available, and what lead time can the team commit to?
- What delivery arrangement fits the order? Pipes from 4 to 12 metres need different transport from hoses shipped in coils, and the answer also depends on the destination and delivery terms.
The details live in different places and do not all change at the same pace. Gamrat publishes price lists for gutters and hoses, while current pipe prices are issued on request. A salesperson has to interpret the email, check the relevant information, resolve gaps, and bring the pieces together into one coherent offer. For a complicated request, most of the two to three days is spent on this preparation, not on writing the response.
That delay affects both sides of the sale. Quoting is one of the clearest places where AI pays back in manufacturing, as we explain in AI in manufacturing: automate the work nobody wants. A customer waits for a clear price and delivery commitment. Meanwhile, the people who know Gamrat’s products best spend time looking up figures and repeating checks instead of advising customers, following up, or developing an opportunity. McKinsey reports that B2B sales teams that automate this kind of back-office work see efficiency gains of 10 to 15 percent and spend more of their time in front of customers. If a number is wrong, the team must also spend time correcting the offer and explaining the change.
From AI training to a practical next step
The June training helped the team use AI to write more quickly. It did not give a general assistant access to today’s price list, stock, or delivery rules. Without approved company information, an assistant can produce a convincing answer while still getting a price or availability wrong. This is a hallucination, which NIST’s generative AI risk profile lists as confabulation: confidently stated but false content. For a commercial offer, a confident guess creates more work and risk than it removes.
The proposed AI agent therefore focuses on preparation. It would read a customer’s email, help match the request to Gamrat’s products, and assemble a draft using the relevant company information, retrieved at the time of the request in the way retrieval-augmented generation grounds an answer in approved sources. It would show the source for each quoted price and mark anything it could not verify, such as an unclear product match or missing availability. Those gaps would go to a salesperson for a decision instead of being filled with an estimate. Letting a model say it does not know, and requiring a source for every claim, are the first techniques in Anthropic’s guide to reducing hallucinations.
The draft would be presented to the salesperson for review in the team’s existing Microsoft 365 workspace. Whether that review step is built in Copilot Studio or as a custom agent depends on how deeply it has to reach into Gamrat’s own systems. The salesperson could correct the details, ask for clarification, or approve the offer before it is sent. The aim is to take repetitive information gathering off the team’s plate while keeping the commercial judgment and customer relationship with the person responsible for the sale. This human-in-the-loop step also settles who is accountable when an AI agent gets it wrong before the first offer is sent.
A safer way to improve the process
The review is also a useful source of learning. When a salesperson changes a product match, corrects a quantity, or rejects a delivery assumption, that feedback can show where the preparation process needs attention. Over time, the team could identify recurring gaps in product information, pricing, or delivery rules and decide whether to improve the underlying process as well as the assistant.
This is important because speed alone is not a good measure of success. A draft that arrives earlier but needs extensive corrections has simply moved the work. A useful pilot should track whether the assistant reduces preparation time while preserving the accuracy and quality of the offer. It should include the time spent checking and correcting a draft, not only the time to generate it. That is the core of honest AI ROI measurement: count the whole process, not the step the model performs.
Potential value to test
The assistant has not been deployed, so these are proposed pilot targets, not achieved results. The pilot should measure the complete time from customer request to a salesperson-ready offer draft. The target is two hours or less for eligible requests, with most straightforward requests taking an hour or less, compared with today’s two-to-three-day turnaround for complex offers.
- Complete draft within two hours for eligible requests, measured from the customer’s email to an offer ready for salesperson review. The target is an hour or less for most straightforward requests.
- 100% of quoted prices linked to an approved source. Missing or unclear prices should be flagged for a salesperson rather than guessed.
- At least 98% first-review accuracy across product, quantity, price, and delivery details, measured before the salesperson makes corrections.
- Less preparation time per offer, measured across information gathering, price and availability checks, and salesperson corrections.
- No increase in avoidable corrections. Track how often a draft needs a change because of a wrong product, quantity, price, or delivery assumption.
If the turnaround and accuracy targets hold without increasing avoidable corrections, customers could receive eligible offers in hours instead of days, while salespeople reclaim time for customer conversations and follow-up. The pilot would establish the actual time saved and show which types of request are suitable for assistance. These measures should be confirmed with Gamrat against its existing process before any results are presented as achieved value.
What happens next
The design is ready to move into a pilot once Gamrat can provide access to the approved price, stock, and delivery information needed to prepare an offer. That data readiness comes first, as it did for another Polish manufacturer in our Ryba Wooden Products case study. Gamrat operates its systems on local servers, so the next step also depends on fitting the assistant into the company’s IT environment and connecting it to the team’s existing review process.
The pilot should start with a defined group of eligible requests and a baseline from comparable offers. The NIST AI Risk Management Framework follows the same order: map the context and measure the risk before relying on the system in daily work. Salespeople would review every draft, while the team records preparation time, source coverage, corrections, and turnaround time. That gives Gamrat a practical basis to decide whether to expand the assistant’s role, adjust the process, or stop.
The business lesson
Offer preparation takes sales time when people have to collect and reconcile the same kinds of information across several sources. AI can help with the language, but the larger opportunity is to make those details easier to bring together and check.
A useful assistant would prepare the first draft from approved information, make uncertain details visible, and leave the final decision with the salesperson. The value is not an automated offer sent without review. It is a faster, more consistent path from a customer’s request to a well-supported decision by the person who owns the relationship.
Preparing quotes from scattered price, stock, and delivery data? Talk to us or see our other case studies.
Customer's own words
“Hands on, on our own tools and our own products. The sales team left with prompts they use daily, and we brought Soba Labs back to train the next departments.”
Jerzy Pachana, Board Member, Gamrat