Brought to you by FORESIGHT - Portfolio Monitoring and Deal Intelligence for Investors
You're chasing board decks, cap tables, and closing docs across your portfolio. Manually pulling KPIs. Translating ownership and liquidation preferences into models. Rebuilding spreadsheets every quarter.
Automate all of that with Foresight's Portfolio Monitoring and Deal Intelligence.
Portfolio companies submit financials and board materials through one-click links. Foresight turns them into clean KPIs and extracts ownership, rights, and economics from closing documents and cap tables. Run scenario models and fund reporting directly on the platform - no spreadsheets required.
🚨New Paper Altert
Exciting news: Our paper "DIALECTIC: A Multi-Agent System for Startup Evaluation" has been accepted at EACL, the flagship conference for computational linguistics🚀
It’s the result of exceptional team work initiated by our previous Earlybird engineering team intern Jae and supervised by TU Munich researchers Simon, Joyce, Georg, and myself from the Earlybird side.
Today’s episode is a quick TL;DR with some clear takeaways and implications for investors.
The Problem: Too Many Investment Opportunities, Too Little Time to Evaluate
Over the last couple of years, starting a company has simply become easier. Vibe coding, AI copilots, and off-the-shelf infrastructure have collapsed many of the traditional barriers to entry.
Hereof, the supply of startups has exploded - and that’s amazing! More founders, more products, more demos, more decks.
At the same time, investor attention doesn’t scale. And that’s not so amazing.
As a result, limited attention faces a growing number of opportunities, and the average time per assessment drastically decreases.
BTW: Did I tell you that I love data over anecdotal stories?🤓
According to DocSend reports, VCs spent 2 minutes and 42 seconds on average per pitch deck in 2022. This dropped by almost 20% to 2 minutes and 12 seconds in 2023 and by another 12% to 1 minute and 56 seconds in 2024 (source, source).
That’s hard data confirming what all VCs complain about: Too much noise, too little time.
But even worse, put yourself in the founder’s position: You’ve spent weeks polishing Slide 8, refining the narrative arc, and tuning your charts. An investor opens the deck between meetings, swipes for under two minutes, and closes it before ever reaching your “big moment.”
That’s not a judgment on the founder, it’s a reflection of a market where attention, not capital, has become the scarcest resource.
To help bridge this gap, my co-authors and I developed DIALECTIC, an LLM-based multi-agent system designed to scale early-stage startup screening.
How Does It Work?
Most attempts to apply AI to VC start from the same assumption: that the hard part of investing is prediction.
If we could just predict outcomes better (e.g. who raises a Series A, who becomes a breakout etc.), we could automate the messy, human parts of the job.
Screening. Prioritisation. Even conviction.
That’s also why most prior work in this space optimizes for predictive accuracy. Better features. Better labels. Better models.
And yes, I’ve also spent several years down this path and published papers such as “Human versus Computer: VC Investors and ML Algorithms for Startup Screening”.
But our new paper takes a different starting point. It optimizes for reasoning quality.
Instead of asking “Can we predict startup success?”, it asks something more subtle:
“How do investors actually form investment decisions, and can that process be modelled?”
That shift matters.
Inspired by how deal flow calls and investment committees operate, DIALECTIC doesn’t just provide a one-shot answer or score but simulates an expert group discussion:
Fact Collection: The system doesn’t just ingest text. It decomposes the company into questions a real investor would ask about team, product, market, positioning, etc. It builds a structured fact base before any judgement is made.
Simulated Debate: Multiple LLM agents generate pro and contra arguments, which are then iteratively critiqued and refined through a "survival-of-the-fittest" debate. This mirrors how humans surface uncertainty.
Strategic Ranking: The system produces numeric decision scores that allow investors to rank and prioritize opportunities more efficiently. Interestingly, the performance improves when the system is allowed to argue with itself. But only up to a point. Too many rounds and quality degrades. That feels uncomfortably familiar.
The Results
In a backtest on 259 early-stage startups evaluated by five different VC funds, DIALECTIC achieves the same precision as human investors when predicting which companies later raised a Series A or beyond.
It does not meaningfully outperform VCs but produces a full ranking across the entire opportunity set rather than a single operating point.
That matters because early-stage screening is not about finding a perfect classifier; it’s about deciding where scarce attention should go when almost everything looks uncertain.
More revealing than the headline metrics is how performance changes as the system “thinks more.”
Allowing the agents to iterate through two rounds of argument and critique improves results; pushing beyond that makes them worse. Argument quality scores keep rising, arguments get longer, and more facts are cited, yet predictive performance declines.
This looks a lot like over-thinking in human investment committees: more words, more sophistication, less signal.
The result suggests that the value of debate is not infinite.
Join our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
Implications for Investors
For practice, the implication is not that VCs should outsource decisions to debating machines, but that screening tools should consider both scoring and reasoning.
A system like this is most useful at the top of the funnel, where the constraint is not prediction accuracy but attention. By generating a small number of defensible pro and contra narratives and turning them into a ranked list, it gives investors something they can react to, challenge, and override - much like a junior investor’s first pass, but at scale and with consistency.
The more subtle lesson is that iteration itself should be treated as a design variable: a little structured debate improves judgement, too much degrades it.
That insight applies as much to human IC processes as it does to AI systems.
Stay driven,
Andre
PS: Reserve your seat for our Virtual DDVC Summit 2026 where expert speakers will share their workflows, tool stacks, and discuss the latest insights about AI for VC
