Automating routine processes using large language models can deliver measurable economic results within just six months. The experience of IT company Cynteka, which develops software for the construction sector, clearly demonstrated that delegating part of the business logic and development to LLMs allowed them to reduce the manual document processing team from 80 to 35 specialists, despite growing data volumes.
Invoice processing: 65% of the flow without human intervention
A key automation focus was processing incoming financial documentation from suppliers. Previously, every line item was manually reconciled against initial purchase requests. Now, the algorithm operates in several steps:
- the system recognizes the incoming document and evaluates data completeness;
- the neural network matches invoice line items with the request nomenclature, checking quantities and units of measurement;
- with high model confidence, the document is processed automatically (taking between 5 and 30 seconds);
- uncertain items are forwarded to a human specialist.
Currently, about 65% of documents are handled without human intervention, with the error rate staying around 1%. The main technical challenge remains complex orders—cases where a single item in a request (e.g., a door unit) is split in the invoice into hardware, door leaf, and molding. Such scenarios are currently intentionally left under manual control.
Accelerating development: one-week MVPs and one-minute code reviews
The adoption of LLMs directly influenced the company's engineering workflows as well:
- Accreditation service in 7 days. A full-fledged module with authentication, counterparty profiles, and a role-based access model—which previously required several months of manual development—was built in two weeks of half-day work with active AI assistance.
- Preliminary review in one minute. The
review.aitool is integrated with GitLab. When a task is submitted for review, the service inspects the diff and provides the developer with a list of potential issues before a human team lead steps in. - Smart automated test maintenance. The team migrated their testing infrastructure to Playwright and introduced assistants to analyze test failures. If an element selector breaks after an interface update, the model analyzes the markup and suggests a valid locator fix.
- SQL assistant for technical support. Feeding database schemas and relationships into an LLM enabled tier-2 support engineers to localize incidents faster and craft targeted queries.
RAG challenges and the transition to isolated agents
Initial attempts to delegate test plan generation to neural networks failed: base models struggled with domain-specific terms and internal product names. An attempt to ingest all available information (Wiki, issue trackers, regression checklists) into a single RAG database caused context conflicts and degraded response quality.
Simply expanding the knowledge base does not work: if sources contain contradictions or varying degrees of relevance, model performance starts to deteriorate. Fine-tuning source prioritization was required.
Ultimately, these developments evolved into a chat widget powered by a multi-agent architecture. A user request is received by an LLM orchestrator, which delegates subtasks to highly specialized agents: one interacts with the backend within the user's permissions, the second manipulates the interface, and the third retrieves data from documentation.
Monetization and matching purchase requests
AI mechanisms also extended to interactions with external users of the "Zakupay" platform. The developed service matches supplier price lists against the historical nomenclature of open orders. During the pilot launch, the system automatically generated 19 commercial proposals for construction material supplies, one of which resulted in an actual paid order.
The main takeaway from six months of testing is that generative models do not replace specialists entirely, but act as a powerful productivity multiplier when their roles are isolated, context is structured, and final control remains in human hands.