all projects

Grant consultancy

AI grant application expert

Writing one grant application took a consultancy's industry expert several weeks. An AI system now reads the company's documents, researches market and competitors on the web, and drafts every section in Google Docs, where the expert steers revisions through inline comments.

weekshours

per application draft

grant-application-expert-ai · shipped 2025 · weeks → hours

The challenge

A grant consultancy prepares applications for innovation-hub residencies and government grant programs. Each application means working through a large stack of the client company’s documents, researching the market, competitors, and patents, and structuring all of it against the program’s evaluation criteria. The work sat with the consultancy’s industry experts, and a single application took an expert several weeks.

Two things made this hard to automate. Client documents are often incomplete, so the system has to notice the gaps and research them. And general-purpose AI assistants can’t hold dozens of sources at once; push them and they start inventing facts. A grant application with an invented number gets rejected.

The approach

The system writes the application section by section, directly inside Google Docs. It reads the company’s uploaded documents, researches market, competitor, and patent data on the web, and maps the content to the program’s evaluation sections.

The industry expert stays in the loop the whole way and never leaves Google Docs. They read the draft and leave inline comments, the same way they would review a colleague’s text. The system treats each comment as an instruction and produces the next revision.

Retrieval-augmented generation grounds every section in the company’s real documents and the researched sources. Every claim in the draft traces back to something the expert can check.

Architecture

  • Orchestration: n8n workflows drive document processing, web research, section generation, and the comment-handling loop
  • AI: OpenAI GPT-4 and Gemini Pro for content generation and document analysis
  • Integration: Google Workspace (Docs, Drive, Sheets) via a Google Apps Script web app and a Workspace add-on
  • Data: PostgreSQL with pgvector as the vector store for RAG

The comment loop took the most design work: the system reads the expert’s inline comments, decides which section each one affects, and rewrites only what the comment asks for.

Results

Drafting that took an expert weeks now takes hours of review. The expert’s time goes into judgment: checking facts, sharpening arguments, steering the draft through comments. The writing itself is done by the system, and every statement in it traces back to a document or a researched source.