Challenges
More2Win's team is fast-moving and proposal-driven.
Every new client pitch requires drawing on a wide base of prior work: old proposals, municipal policy documents, impact reports, funding research, and partner data.
All of this lived across OneDrive folders with inconsistent structures, named differently per project and per team member.
This made it difficult to:
Find relevant prior work quickly when writing a new proposal
Connect policy documents to the right client context or geographic region
Reuse insights from past programmes without manually searching through dozens of files
Give team members fast, reliable answers to research questions
The result: high-value institutional knowledge was effectively invisible at the moment it was needed most.
Approach
How we approach building a solid solution
Step 1 — Discovery and Knowledge Mapping
We ran a scoping session to map More2Win's full knowledge landscape: what documents existed, how they were organised, how team members searched for information today, and where the biggest friction points were. We identified five core content categories: internal reports, municipal and national policy docs, funding and partner research, programme outlines, and impact reports.
Step 2 — System Architecture and Ingestion Pipeline
We designed an automated pipeline connecting to More2Win's OneDrive via the Microsoft Graph API. Documents are automatically detected, converted to text, chunked, and passed through an AI classifier that tags each piece of content with metadata: client, proposition type, document type, municipality, and confidence score.
Step 3 — Vector Database and Semantic Search
All processed documents are stored as vector embeddings in a Pinecone database, partitioned by client and region to prevent cross-context confusion. This gives the system semantic memory, meaning users can ask questions in natural language and retrieve the most relevant content across hundreds of documents instantly.
Step 4 — AI Research Agent and Interface
We built a chat-based AI frontend where team members can ask free-form questions in Dutch or English. The agent retrieves the most relevant document chunks, generates a grounded answer, and surfaces direct links back to the source files. A human review layer flags low-confidence classifications for manual correction, keeping the knowledge base clean over time.
Final thoughts
We delivered a private, EU-hosted AI knowledge system built entirely on More2Win's own data.
The system automatically:
Ingests and classifies new documents from OneDrive as they are added
Tags content by client, region, proposition, and document type
Surfaces relevant prior work and policy hooks in response to natural language queries
Cites the original source file for every answer it generates
Flags uncertain classifications for human review before they enter the knowledge base
All data stays within the EU. The AI operates exclusively on More2Win's own content, with no external model training involved.
The system was designed as a foundation: built to extend. Future additions like an automated client intake and qualification engine or an automated proposal drafter can plug directly into the same knowledge layer.
Example queries the system can answer today:
"Welke vijf punten uit dit gemeentebeleidsplan zijn relevant voor meisjesvoetbal?"
"Welke programma's hebben we eerder ontwikkeld rondom vrouwensport in Brabant?"
"Wat zijn relevante beleidshaakjes voor een nieuw voorstel aan gemeente Utrecht?"
Results
More2Win's team can now access the full depth of their institutional knowledge in seconds rather than hours.
Proposal writers get instant context on what has been done before and which policies apply -- without opening a single folder. New team members can onboard to client contexts without relying on colleagues to surface the right documents. And the knowledge base grows automatically as new files are added to OneDrive.
What was previously invisible is now immediately actionable.
Key Results
Seconds to find relevant info
Of their sources covered
Manual steps to index
