š Hi, Iām Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
Join our free virtual roundtable āThe Compounding Data Layer: Building an Edge When Everyone Uses the Same AI Modelā with Eight Roads, Offline Ventures, and Kruncher here
Brought to you by Affinity ā The CRM for private capital
Affinity's agent platform for private capital is here.
Every deal stage has work that eats into the hours your best people should be spending on judgment calls. Affinity's agent platform handles that work across sourcing, qualification, IC prep, and monitoring, so your deal team can get back to building relationships, evaluating management teams, and closing.
Join us July 22 for an exclusive walkthrough of the first agents shipping across sourcing, deal qualification, and deal monitoring.
Welcome to another Data Driven VC āInsightsā episode where we cover the most interesting research and reports about startups, VCs, LPs, AI & automation.
Agent Costs More Than Two Junior Bankers
Ori Eldarov, founder of OffDeal, on why his internal AI agent Archie has 10x'd in cost since January.
$420K Annual Run-Rate: Archie's monthly cost jumped from a few hundred dollars to roughly $35K as chat volume grew from under 200 to about 2,400 a month, now costing more than two full-time analysts.
Eval Score Drove the Spend: Pushing accuracy from roughly 25% to 75% meant optimizing purely for performance, and the team passed on a frontier model upgrade because a 3-4% eval gain would have doubled costs to $70K+.
Smaller Models Are Already Winning: Harvey's hybrid GLM 5.1 plus Opus 4.7 advisor beat Opus alone on both quality (18% vs 14%) and cost ($368 vs $954), and Cursor's Kimi K2.6-based Composer 2.5 cuts coding costs roughly 10x.
āļø KEY TAKEAWAYS
Model routing across specialized models is becoming standard practice for AI-native companies. Diligence should now track token spend trajectory alongside headcount. Companies still paying frontier prices for tasks a cheaper model could handle are bleeding margin a routing layer would protect.
How to Build an AI GTM Team
Kyle Poyar, from Growth Unhinged, breaks down how monday.com's AI agents generated millions in pipeline last quarter.
24 Hours to 2 Minutes: Inbound agent "Amanda" handles 100% of English-speaking contact-sales requests and runs five-minute qualifying calls within 60 seconds of a form submission.
2.5x Conversion Lift: Trial activation agent "Jax" logs 3,000+ monthly calls, with half of users returning for a second session and converting to paid at 2.5x the control group rate.
Weeks to Minutes: Outbound research agent "Oscar" compresses one to two weeks of account planning into roughly five minutes across a 1,000+ person sales org.
Speed-to-lead and trial-to-paid conversion are now concrete AI ROI metrics worth tracking. monday built this as an internal startup with a dedicated owner, suggesting org design matters as much as the tooling. This is one of the clearest public-company proof points yet that AI GTM agents move beyond pilots into core revenue infrastructure.
The āTalent Engineeringā Breakdown
Rich Zou, founder of Bo Le Capital, breaks down the distinction between two emerging hiring roles most companies confuse.
Talent Is a Magnet: A "talent" person attracts and holds deep relationships with smart people, technical or not, and the test is whether extraordinary people already cluster around them.
Talent Engineering Means a Real Engineer: The role only works when filled by someone who could be shipping core product but chooses people tooling instead, which is exactly why it's so rare.
Cursor Treats Every Hire Like an Executive Search: Two people spend 50-60 hours a week hand-mining referrals, and technical staff spend roughly a quarter of their time recruiting.
āļø KEY TAKEAWAYS
Stop judging hiring functions on pipeline metrics and start asking who personally closes candidates. Firms that hand recruiting entirely to a function tend to plateau at "above average." The same test applies to a fund's own talent partner: who they spend time with matters more than their resume.
Join 1789+ investors in our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
Your Agent's Memory Is Making It Dumber
Matt Van Horn cut his AI agent's memory from 218 files to 6 after his 46-kilobyte memory index started silently failing to load.
The Index Was Already Failing: Past Anthropic's recommended 200-line limit, the harness was silently dropping content every session without warning.
Most "Bloat" Was Misfiling: Only 14% of files were true trash; roughly 39% were single-skill lessons that belonged inside that skill as a pull request rather than sitting in global memory.
Six Survivors Replaced 218: Only hard safety rules, formatting standards, and stable repo paths made the cut, shrinking the index to about 1.4 kilobytes.
āļø KEY TAKEAWAYS
Any team running Claude Code at scale should audit whether standing instructions still earn their place each session. Push memory loads automatically and degrades performance past a threshold; pull memory is retrieved only on demand. Treat this as a recurring operational audit and not a one-time cleanup.
Origins of VC Content
Laurie Owen (Refining VC) traces how Fred Wilson's AVC blog became the template every VC content strategy still copies.
Two Decades of Near-Daily Posts: Fred Wilson blogged almost every day from September 2003 onward, building a visible record of changing his mind in public, including reversing his Bitcoin stance between 2014 and 2017.
MBA Mondays Spawned the Genre: Launched January 2010, the weekly series taught founders real-time problems rather than demonstrating expertise, and it's the direct ancestor of every "VC explains X" thread since.
A House of Named Voices: USV built partner-owned blogs around a central hub and recently launched the Librarian, an internal tool indexing 20+ years and roughly 15,000 articles.
āļø KEY TAKEAWAYS
Content compounds only when tied to a firm's actual judgment process. Pick one consistent, multi-year format rather than chasing channels and cadence. The firms doing this well, including USV, Slow, and Massive VC, treat their thesis as a living public document that keeps evolving.
The Herd Mentality Eating VC
Dan Gray (Odin) argues venture capital's reliance on consensus narratives erodes the intellectual diversity the industry depends on.
A 135-Paper Research Directory: Odin's new directory spans four decades, from the 1980s through 2026, built to counter an industry running mostly on anecdote.
Processing Fluency Rewards Simple Stories: The piece ties venture's short-form bias to the spiral of silence, where people suppress non-consensus opinions when they expect disagreement.
Outsiders Show Aggregate Outperformance: Distance from the herd preserves the idiosyncratic models needed to spot unconventional opportunities. ZIRP-era fund convergence and its disappointing returns made the case.
āļø KEY TAKEAWAYS
Treat genuine idiosyncrasy in an emerging manager's worldview as a feature worth underwriting. This is a useful gut check for any fund leaning on podcast circuits and shared narratives to form conviction. The filter: did this thesis come from visiting a factory or analyzing proprietary data, or just absorbing other investors' takes secondhand?
Thatās it for today!
Stay driven,
Andre
PS: Join our free virtual roundtable āThe Compounding Data Layer: Building an Edge When Everyone Uses the Same AI Modelā
