We build AI into the deal and portfolio workflows of Dutch private equity firms, so sourcing, IC memos and value creation move faster and stay consistent. The systems below already run in production inside a Dutch PE firm.
Founder, Poolman Ventures · Delft, NL
Most AI consultants have never sat in a deal team. I spent three years inside Dutch PE (Vortex Capital Partners, then embedded at Lexar Partners), building the systems above from the inside, on an engineering and econometrics foundation (TU Delft, VU Amsterdam). Poolman Ventures is founder-led and PE-native by design: you work directly with the person who builds, grounded in how the workflow really runs rather than a generic AI playbook.
Four production systems built and shipped end-to-end inside a Dutch private equity firm, across competitor intelligence, portfolio value creation, deal-flow analytics and LP reporting. Plus a sourcing engine proven at cross-sector scale.
An operating cockpit for the firm and its portfolio companies: portfolio management, a valuations engine, quarterly roll-up, an 8-pillar maturity scorecard and playbook-driven task management. Production SaaS with single sign-on and role-based access for staff, operators and admins.
→
→
→
Scrapes 10,000+ PE deal announcements, classifies each against a 23-firm competitor map, enriches it with LLM research and scores it for thesis fit (1 to 5). A BI dashboard surfaces competitor activity, advisor league tables and the relevant deals the team never saw. Five quarterly intelligence cycles shipped.
→
→
→
The firm's sourcing funnel read straight out of Affinity (leads, appointments, NBO, term sheet, close, split platform vs add-on, season on season) on live dashboards. Plus an automated pipeline that builds the quarterly LP advisory-board slides end-to-end: pull the export, parse it, render the deck, send it for review. Every quarter since 2023.
→
→
→
An automated pipeline that maps an entire market: every company found, enriched, filtered and scored against the target profile, with per-company web research as the quality gate. Proven at cross-sector scale on a B2B market-mapping run outside PE: 93,497 companies across 12 EU countries, narrowed to a research-qualified shortlist.
→
It works for simple tasks (summarise, draft, reformat). On anything complex it stalls at 70-80 percent, and tomorrow starts from the same blank prompt.
ChatGPT Claude
Copilot
Works for simple tasks (summarise, draft, reformat). On anything complex it stalls, and nothing carries forward.
It cannot see Affinity, the data room, or last week's IC memo, and remembers none of your workflows, preferences or corrections. Every session starts from zero.
It picks the structure, the weighting and the criteria without checking. You catch the mismatch only in the final draft.
What comes back is the average of the internet: fluent, plausible, and the same answer any fund would get. It reflects none of your thesis or judgment.
An LLM is a brilliant new hire on day one: sharp, well-read, fast. What it lacks is everything a person absorbs over years inside a firm. We onboard every human, then hand AI a prompt and expect senior work.
The systems hold the data, but not in a shape an agent can use. Meaning lives in people's heads, and every run is driven by hand.
Two moves, and they only work together. Each run produces reviewed, reusable artifacts that make the next run faster, more accurate and more aligned.
Decompose each workflow into an abstraction hierarchy. Each layer produces a distinct artifact, and each artifact gets reviewed before the next layer proceeds.
A layer on top of your existing ops stack that holds persistent context (the data every run reads and writes) and packages reusable workflows, for both humans and agents.
Both are required. Together they create a compounding system. A correction or new signal enters at the layer it belongs to and re-flows from there, so you fix the cause rather than rerun the whole job.
Today, the only structured step is the lowest-value one. What "fit" means, the target profile, never gets written down, so there is no place for an agent to plug in.
Filtering is the only structured step, and the analyst already does it by hand. What "fit" means is locked in the associate's head: no artifact for an agent to match against, no checkpoint for the partner to validate.
Every layer owns one artifact. Write the target profile down once, and an agent can finally plug in. Hover any layer to open its artifact.
Analyst + Investment Manager. Scope on the table, design, build, blockers.
Partner + working-session group. Roll-up of decisions, scope, risk, budget.
Full sponsor group + IC member. Kickoff, mid-engagement, pre-handoff sign-off.
We know PE firms handle highly sensitive data, so we build on a choice of models, each with its own data residency, risk profile and tradeoffs. You set the posture per workflow, with identity, access and audit always in your own environment. Hover a posture to see how it runs.
Frontier model via enterprise API (Anthropic, OpenAI), or Azure OpenAI inside your own Azure tenant.
Leaves the tenant under a zero-retention contract. Never used to train a model.
Market maps, public company research, general drafting.
Open-weight model on EU cloud, ideally your own cloud tenant.
Stays in the EU and your cloud boundary. Never reaches a model vendor.
Portfolio KPIs, internal memos, archive retrieval.
Local open model on a dedicated workstation inside the firm.
Never leaves the building. Air-gapped option.
Live deal data, data-room contents, a live IC memo.
Your IdP / SSO governs every tier
Across all three tiers, always
Who asked what, over which documents
What are your biggest operational challenges?
What are you already doing with AI?
Where does your data live, and how structured is it?
Where do you see the highest-ROI use cases?
Which workflow would you want to tackle first?
What would make a first pilot a success?
A short Discovery sprint maps where AI plugs into your deal and portfolio work, and gives you a prioritised roadmap to act on.