AI-Assisted B2B Platform Development: 46,000 Lines of Code in 1.5 Months

Previously, managers at the company used 4-5 different services: email clients, spreadsheets, separate CRM systems. Every incoming inquiry required manual processing: identify request type, create contact, set up deal, track status. With hundreds of inquiries per day, this created a bottleneck in the business process.

We developed a custom CRM system with an AI core.

The system parses incoming inquiries, extracts key information (contacts, company, request essence) and automatically creates entities in the system. Works with both structured and free-form formats.

Based on deal context (correspondence history, status, tasks), the system suggests next actions: change status, create task, propose response text. Manager confirms or adjusts.

Deal kanban board, activity timeline, prioritization system, filters and search. Everything that was previously scattered across 4-5 apps is now in one interface.

The system analyzes attached files, extracts information and adds it to the deal context.

During the pilot, the manager acts as a controller of AI suggestions. But over time, once we see that AI system development has reached a certain level, the process will be fully automated (without a manager).

The manager will only engage with deals when the situation requires it. For example, at the finish line, or if a client asks for a callback in their email. In such cases, we assign a task to the manager.

This way, our AI assistant will participate at all funnel stages.

Stack: NestJS, NextJS, ShadCN UI, PostgreSQL.
Development Tools: Claude Code as the primary code-writing tool.

• Decomposed documentation for AI (separate md-files for each module)
• "Agent" system with specializations (backend, database, integrations)
• Iterative development within one context per task

• 6 weeks

  to working version

• 2-3 weeks

  for feedback-based improvements

• Processing time per inquiry reduced several times
• Managers work in one interface instead of five
• AI suggestions reduce cognitive load during high volume
• System launched and in production use