✍️ How to Automate Deal Sourcing…

1. Key Themes

Theme: Data quality and trust are the foundation of AI-driven sourcing and scoring

Fix the data before the scoring

A sophisticated score built on bad data destroys internal credibility, and that trust is hard to rebuild.

"A sophisticated score built on unreliable data can quickly cost you your investment team's trust. We learned this at Earlybird when we launched scoring too early. Once trust is gone, it is very hard to win back."

AI has collapsed the cost of data plumbing

Entity matching and deduplication, formerly a multi-year project, is now a sprint.

"Entity matching and deduplication used to take years. Today, a one-week sprint quickly gets you beyond 90% accuracy, the same range that once took several years."

Measure sourcing coverage with a hit rate

Benchmark against peer funds' new investments to quantify what you're missing.

"Pick the peer funds you consider relevant and track their new investments through public registers. Your hit rate is the share of those companies you had already met."

Theme: Scoring systems must be governed like products, not dashboards

Two-layer scoring with learning from misses

Hard filters first, then a preference-learning layer, plus structured post-mortems on missed deals.

"Hard filters (geography, stage, company age, funding raised) come first. A second layer learns what each investor, and the team as a whole, likes, based on the reasons people give when they reject a recommendation." "When a great company slips through and you only hear about it from a peer, it goes into a structured review: what do these misses have in common, and why did the scoring drop them?"

Single ownership and periodic review

Centralized control of weights prevents model sprawl.

"Only engineering can change the scoring. Investors send feedback, which keeps the model from turning into a puzzle nobody can untangle." "Don't change the scoring every day but review batches of misses and survey the team on which characteristics matter most right now."

Scores are rough guides, not precise rankings

"A 76 vs a 81 of 100 deserve equal attention. A 40 vs a 76 of 100 do not."

Theme: The conversation has shifted from AI adoption to organizational operations

Governance, permissions, and review gates

Role-based access via a central MCP and a review process for shared AI "skills."

"Use a central MCP with role-based rules that decide who sees what. A partner can read colleagues' email content, while an analyst only sees that a partner was in contact with someone." "Anyone can submit one. Engineering checks it for data leakage and relevance before rolling it out firm-wide, and a weekly team call shows everyone what's new."

The macro shift

"Since our last breakfast in Berlin in June, the conversation has moved from adoption to operations. Who can change the score? Who sees which data? Who triages the top 100? Which parts of the memo are written by humans? These are organisational questions."

Connectors and dedicated technical talent are the unlock

"One fund in a larger regulated group ran an enterprise chatbot without connectors for a long time. Since connecting their tools, people there describe being able to do about 5x more. Their next step is a dedicated CTO-level hire, because an investor building on the side eventually hits a ceiling."

Theme: Post-deal automation, from portfolio data to memos and LP updates, delivers the top ROI

Portfolio data collection should be automated, not requested

"Portfolio data collection is a behavioural problem. Founders are busy, and sometimes they don't want to send numbers." "Portfolio updates are CC'd to a dedicated email address, and AI extracts the data, attachments included, with accuracy they report at 97%+. With their MCP server, an analysis like follow-on vs. initial-ticket performance takes few minutes in Claude."

Memos and LP updates compress from days to hours

"Memos that took days now take hours, built from data sources, CRM notes and call transcripts, and updated after every founder call until the IC. One fund now often runs Series A and B deals with two people instead of three."

KPIs must serve a purpose

"Portfolio data matters when it helps you show up better for your portfolio companies and your LPs."

Theme: Edge migrates from tooling to proprietary data, judgment, access, and network

Where alpha lives when the stack is commoditized

"Most of the stack is or will be commoditized. The room pointed to what remains: Proprietary first-party data: transcripts, memos, and above all portfolio data." "Judgment: it takes many repetitions, and some people simply have it. Two investors can leave the same meeting with opposite reads of the same founder." "Network leverage: a platform one or two hops away from almost anyone in its market has an advantage that's hard to copy."

Venture increasingly resembles sales and brand

"As technology democratizes, venture becomes more about brand and looks more like sales. Using the fund's internal knowledge, like the right intros or structuring an M&A process at exit, is part of that pitch."

Network graphs are the next frontier

"Several funds see their collective network as one of their biggest untapped assets and want a warm-intro graph that surfaces the strongest path to any founder. The building blocks are there, but relationship data from email and calendars still needs to get more reliable before it can carry that weight."

2. Contrarian Perspectives

AI-assembled memos can make IC papers worse, because nobody reads them

Faster memo generation isn't a clear win. The bottleneck shifts from writing to reading, and one fund's analysis showed most content was redundant.

"One attendee said that since their memos became AI-assembled, people are reluctant to read them in full." "Their own analysis of the past 12 months found over 80% of the content in their IC papers repeated information from decks, data rooms and websites. Now deal-team commentary on key risks and 'what we need to believe' sits in a different format in every section, so readers can jump straight to what the humans think."

Scoring should upweight team and momentum precisely because moats are unpredictable, even at the cost of bias

Against the instinct to lean on defensibility or market analysis, one fund shifted weight toward softer signals, while admitting the downside.

"With moats so hard to predict right now, they significantly upweighted team and momentum. They were also candid that pedigree signals penalise underdog founders."

Walled-garden tools are not where the market is heading, and your sunk-cost tools may be hurting you

Closed portfolio-data UIs are viewed as a dead end, and in-house tools often persist for emotional reasons.

"He also doesn't think walled-garden tools that keep data inside their own UI are where the market is heading." "Many funds are stuck with an average in-house tool because it's someone's baby. Buy when something better exists." "Walled gardens are slowly opening. Pressure from customers is pushing more data providers to offer MCPs and integrations."

3. Companies Identified

Earlybird

Description: European venture capital firm; Andre Retterath's firm
Why mentioned: Case study in launching scoring too early and in being a long-time customer of Vestberry
Quotes: "We learned this at Earlybird when we launched scoring too early." / "we've been happy customers at Earlybird for years."

Vestberry

Description: Portfolio monitoring and data collection platform for funds; event partner and one of DDVC's first partners
Why mentioned: Example of automating portfolio data collection with AI extraction and an MCP server
Quotes: "Portfolio updates are CC'd to a dedicated email address, and AI extracts the data, attachments included, with accuracy they report at 97%+." / "They were one of DDVC's first partners."

Granola

Description: AI meeting copilot/transcription tool (newsletter sponsor)
Why mentioned: Sponsor; positioned as a bot-free meeting transcriber
Quotes: "There are no meeting bots. Nothing weird joins your call. Granola transcribes directly from your computer or phone audio."

Claude (Anthropic)

Description: AI assistant used with MCP integrations
Why mentioned: Used to run portfolio analyses on top of Vestberry's MCP server
Quotes: "With their MCP server, an analysis like follow-on vs. initial-ticket performance takes few minutes in Claude."

Harmonic

Description: Startup data/sourcing platform; co-host of the next Builders Breakfast
Why mentioned: Partner for the Helsinki event
Quotes: "Join our next DDVC Builders Breakfast at Slush in Helsinki on November 18th, together with Harmonic and Goodwin."

Goodwin

Description: Law firm; co-host of the next Builders Breakfast
Why mentioned: Event partner
Quotes: "...together with Harmonic and Goodwin."

4. People Identified

Andre Retterath

Description: Author of Data Driven VC; Partner at Earlybird
Why mentioned: Host and author; shares the firm's scoring lessons
Quotes: "We learned this at Earlybird when we launched scoring too early." / "Every big brand I've spoken to is doing something here. Many just aren't allowed to talk about it."

No other individuals are named in the article. Event attendees are referred to only anonymously (for example, "One investor," "One attendee").

5. Operating Insights

Give one owner control of scoring weights

Route investor feedback to engineering, and review misses in batches rather than continuously.

"Only engineering can change the scoring. Investors send feedback, which keeps the model from turning into a puzzle nobody can untangle."

Restructure memos to separate human judgment from aggregated facts

Standardize a distinct section for deal-team views so readers can skip the redundant material.

"Now deal-team commentary on key risks and 'what we need to believe' sits in a different format in every section, so readers can jump straight to what the humans think."

Protect engineering capacity with clear OKRs

Internal requests silently drain technical teams.

"LP data requests, DD support, or compliance questions quietly eat engineering capacity. Clear OKRs keep the team shipping and aligned."

6. Overlooked Insights

IC dynamics are predictable and mood-dependent

This hints at an opportunity to model partner preferences, but also at noise in decision-making.

"Knowing each partner's preferences, you can often forecast the vote. Votes also shift with energy and mood."

Individual IC ratings enable measuring partner decision quality over time

A quietly powerful use of first-party data that most funds don't systematically exploit.

"Individual IC ratings let you look back at each partner's decision quality."