AI assistant for tender documentation

An R&D prototype of an AI assistant that prepares commercial proposals, technical rationales and tender documentation based on a corporate knowledge base

The goal was to build an AI assistant that helps specialists prepare tender and commercial documentation faster. In a typical process most of the time goes into searching for similar bids, copying fragments from past documents, adapting wording and checking mandatory requirements.

We needed a solution that works not as a regular chatbot but as a corporate tool: it searches for relevant materials in the knowledge base, uses them as context for generation, produces a draft document in the required structure and highlights the points that need review by a specialist.

Project format: R&D prototype of an AI solution
Document types: commercial proposals, technical descriptions, explanatory notes, questionnaire responses
Test knowledge base: 250+ documents
Index size: about 3,500 text fragments
Test scenarios: 35
Goal: reduce the time of initial document preparation and lower the risk of missing requirements

We chose a RAG approach: the system first runs a semantic search across the knowledge base, finds relevant fragments from past documents and only then passes the context to an LLM to generate a draft. This reduces the risk of abstract or irrelevant answers and makes it possible to reuse corporate wording, templates and standards.

We also implemented a requirements check module: the AI assistant matches the draft structure against the original tender conditions, flags missing items and highlights the sections that need manual work.

We designed a prototype of an AI assistant that works with the corporate knowledge base and helps a specialist assemble the first draft of a document. Documents are uploaded to the system, cleaned up, split into meaningful fragments and indexed in a vector database.
The user sets the parameters of the new document: bid type, procurement subject, requirements, deadlines, scope of work and additional conditions. The system then finds similar materials, builds the context and generates a draft in the given structure: introduction, solution description, technical compliance, advantages, timelines and clarifications.
The final document is never sent automatically. The specialist receives the prepared draft, reviews it, edits questionable parts and regenerates individual sections when needed.

The result is an R&D prototype of an AI assistant for working with corporate documentation. The test knowledge base holds 250+ documents, about 3,500 text fragments were indexed and 35 scenarios of tender document preparation were tested.
In the test scenarios the average time to prepare a draft was 12–18 minutes instead of 2–4 hours of manual work — a reduction of up to 70% in initial preparation time. The accuracy of relevant fragment retrieval was 82% and the share of correctly filled mandatory sections was 88%.