[case] AI agent + CRM for a service center · water heater manufacturer · AQUAТЕРМ (під NDA)
An AI agent for the claims department: −2 staff and ₴730,000 saved per year — AQUATERM case study
Project name changed due to NDA
Telegram AI agent (LLM + computer vision) + web CRM · AQUAТЕРМ (під NDA)
[01] challenge
What the project had to solve
A water heater manufacturer. The claims department — three operators who handle a stream of inquiries every day: a customer writes or calls, the operator finds out the model, purchase date, and symptom, checks the warranty, identifies the fault by the error code on the display, and resolves it remotely or dispatches a technician. The issue is that most inquiries are alike. The same error codes, the same questions, the same answers. The operator spends time on routine that can be formalized. Some technician visits are unnecessary: the problem could have been solved with advice in chat. And scaling meant one thing — hiring more operators. The client wanted to take the routine off people without losing quality of handling.
[02] solution
What we did — and why
An AI agent, "Bohdan," in Telegram, trained for the product. This isn't an off-the-shelf bot with scripts: the agent knows the client's model lineup, error codes, and warranty terms, and behaves like an experienced service engineer. A natural dialog that collects the data. Instead of a form — a conversation: the agent asks for the model, purchase date, and symptom, prompting with buttons. The customer answers the way they're used to writing in a messenger. Error code recognition from a photo. The customer photographs the display — the agent recognizes the code (for example E04), identifies the fault (temperature sensor) with a probability, and immediately says whether it's a warranty case or not. Resolve on the spot or dispatch a technician. The agent closes typical inquiries itself; where a visit is needed, it dispatches a technician and creates a ticket. Unnecessary visits are filtered out right in the chat. A web CRM for the department. All inquiries from Telegram and the service center — in one list with an AI diagnosis, statuses, and technician assignment. Analytics for the manager: a fault map, the most problematic models, a load forecast, the share of inquiries handled without an operator, and recognition accuracy. How we trained the agent (and this is exactly what we offer as a service — not "hook up a bot," but train an agent for a specific business): we loaded it with data (models, error codes, warranty rules, typical solutions), wrote out its behavior (how to greet, what to ask, when to close on its own, when to call a technician), and trained it together with the department's staff on real inquiries until the quality reached the level of a live engineer. How it works technically: the customer writes in Telegram → the agent holds the dialog and collects data → computer vision recognizes the code in the photo → the agent checks the model and warranty, makes a diagnosis → gives a solution and closes it or dispatches a technician and creates a ticket in the CRM. An operator steps in only where the agent isn't sure.
project frames
[03] result
→ The claims department shrank from 3 to 1 person in 3 months — the one who remained checks the agent's answers and coordinates information for other teams → ₴730,000 saved per year → The agent closes 67% of inquiries without an operator — typical codes and warranty cases → 94% accuracy recognizing the error code from a photo → Unnecessary technician visits are filtered out right in the chat The agent could have closed the last role entirely too, but the client deliberately kept one person on quality control — the depth of automation is decided by the client, not by the limits of the technology.
Department
3 → 1
Savings / year
730 000 ₴
Closed without an operator
67%
Photo accuracy
94%
Need a project like this in your niche?
We'll show how we'd solve it for your business. The estimate is free.
← All cases