Turning 5 years of scattered files into a 30-second research engine

More2Win wins work on the strength of what they already know.

Five years of municipal policy docs, partnership research, programme outlines, and impact reports lived across OneDrive folders, named differently per project and per team member. Every new proposal started the same way: a strategist opening folders, guessing at file names, and trying to remember which colleague worked on a similar pitch in 2022.

The institutional knowledge was there. It was just invisible at the moment it was needed.

The questions leadership couldn't answer:

How many billable hours per week is the team losing to folder archaeology?

Why does onboarding a new strategist take weeks, when most of what they need is already written down?

How do we reuse insights from past programmes without depending on the memory of whoever was in the room?

And how do we keep this knowledge inside the EU, on our own data, without feeding it to a third-party model?

The cost of inaction was hidden but real: every proposal carried a research tax, every new hire ramped slowly, and every piece of institutional knowledge that wasn't reused was effectively lost.

What we built is a private AI research engine, hosted in the EU, that runs on More2Win's own knowledge.

Ingestion pipeline. Connected to More2Win's OneDrive via Microsoft Graph. Documents are detected automatically as they're added, converted to text, chunked, and tagged with metadata: client, proposition type, document type, municipality, confidence score. Zero manual indexing.

Vector knowledge layer. All processed content is stored as embeddings in Pinecone, partitioned by client and region to prevent cross-context confusion (so a Utrecht proposal doesn't accidentally pull policy hooks from a Rotterdam project).

AI Research Agent. A chat interface where the team asks questions in Dutch or English. The agent retrieves the most relevant chunks, generates a grounded answer, and links every claim back to the source file. A human review layer flags low-confidence classifications so the knowledge base stays clean as it grows.

The cost story is where this stops being a search tool and starts being infrastructure.

That's roughly 640× cheaper per proposal, with the curve going the right direction: a sixth proposal in a busy week costs the same as the first.

The deeper point is ownership. The system runs on More2Win's data, in an EU-hosted environment, with no external model training involved. If a Pinecone competitor ships something better next year, the corpus moves. If What's Next disappeared tomorrow, More2Win still owns the data, the prompts, and the configuration. That's the difference between buying a SaaS subscription and building a strategic asset.

The human review layer is intentional, not a limitation. The agent surfaces and grounds answers. The strategist still decides what goes into the pitch. The system makes the team faster at thinking, not a replacement for thinking.

The pre-work that used to consume half a day per proposal now takes 30 seconds per question.

Operationally, the team's workflow has changed shape. Proposal writers no longer open folders to start a pitch. They ask the engine, get grounded answers with source links, and move directly to drafting. New strategists onboard against the corpus instead of against colleagues' availability. The knowledge base grows automatically as new files land in OneDrive, with no maintenance burden on the team.

Strategically, the team can now compete for proposals they would previously have declined on capacity grounds. Institutional knowledge that was effectively trapped in named-by-who-saved-it folder structures is now compounding: every new client engagement adds to the corpus, every new corpus addition makes the next pitch faster. The foundation is also live for Fase 2 (automated proposal drafting) and Fase 3 (intake and qualification), both of which plug into the same knowledge layer.

Total timeline: 12 weeks from discovery to production. The discovery half wasn't optional. It's why the build half scales.