The question
Can you better substantiate portfolio decisions with data, without someone having to read and copy-paste all documents into Excel for weeks?
Portfolio managers spend a lot of time collecting and reading documents, transcribing information into lists and scores, and working out scenarios based on priorities and constraints.
It often feels like a backpack you have to fill with projects, while each project has a different format, weight and importance. In practice, this is currently mostly manual work, with many subjective choices, making the optimal filling impossible to calculate.
The question for this PoC was simple: How far can we get if we let AI help with the preparatory work, so the portfolio manager has more time for the actual choices?
The data pipeline
Project proposals (.docx)
Unstructured Word documents with project information
LLM extraction
AI extracts relevant data: costs, duration, risks, strategic goals
Structured JSON
Uniform data structure with fixed fields for each project
Python script
Inject data from JSONs into spreadsheet
Portfolio toolkit
Excel with automatic calculations and filters
Dashboard
Interactive visualization with scenarios and management info
The dashboard
Pragmatic start
For this PoC, the flow was set up manually. In production, this can be easily automated with n8n or Azure Logic Apps, so new documents automatically go through the pipeline.
Human in the loop
Between each step there's a human check. AI can hallucinate and source documents aren't always complete. Garbage in = garbage out, that's why validation remains essential.
What this PoC showed
AI takes over reading work
You no longer have to manually translate each paragraph into columns.
Portfolio managers stay in control
AI does the initial filling, but humans determine what weighs heavily.
Scenarios are faster to explore
You can play with "what if" scenarios more quickly.
Visualization helps the conversation
Clear overviews make it easier to substantiate choices.
Limitations and considerations
Quality depends on source documents
Incomplete or outdated documents won't produce good input, even with AI.
Soft factors remain difficult
Things like politics, change capacity or culture are hard to capture in a score.
AI is not flawless
Every extraction must be verifiable. That's part of the process.
This is not a push-button solution
It's a tool to arrive at scenarios faster and better substantiated.
What you can do with this as an organization
For Ruysdael, this PoC was mainly a way to see what happens when you combine their expertise with AI and data. The result is not a product that can go live tomorrow, but it does provide good insight into what AI can contribute:
Less time spent
on preparatory work
More time
for the conversation about choices
Better substantiated
with numbers and scenarios
OpenAI GPT-4 with custom prompts
JSON schema, Python processing
Excel with formulas and pivot tables
Custom dashboard (React + D3)
n8n / Azure Logic Apps ready
Check out these cases
Similar challenge?
Do you have a portfolio with stacks of project documents and want to explore if this kind of tooling can help?
Then a similar proof-of-concept is a logical first step.