AI Pushes Venture's Edge Toward the Decisions Nobody Can Score, Standard Metrics' Melas-Kyriazi Argues

24 minutes ago

4 min read

What's New

Every venture task that can be measured against a benchmark will be automated, and the human value in the job migrates to the decisions where no benchmark exists. John Melas-Kyriazi, co-founder and CEO of Standard Metrics, argues this in a podcast on How I Invest. Parsing cash flow statements and board decks, building founder dossiers, and getting a generalist up to speed on semiconductors all have evals: you can run a thousand documents, measure accuracy, and tighten the loop. Judging a founder over hours of conversation, supporting a company after the check, and deciding how to deploy the 30 to 50% of a fund held in reserve do not. Firms that treat AI as a research tool will free up time; firms that redirect that time to unscorable judgment will keep the alpha.

Why It Matters

The consensus story is that AI compresses diligence and sourcing, and that speed itself is the advantage. Melas-Kyriazi's argument implies the opposite for differentiation: once every firm can generate the same dossier in minutes, the dossier is worth nothing at the LP pitch. On the other side sit firms building their own full software stacks with coding agents, and anyone who assumes follow-on decisions can stay a quick partner-meeting nod. He sells portfolio management software to venture and private equity firms, so the buy-versus-build and reserves arguments serve his product. He also concedes the build barrier has fallen in the coding-agent era.

Big Picture Drivers

Automation follows eval quality: Where a workflow can be scored, models and agent harnesses improve relentlessly. Where it cannot, humans stay. Melas-Kyriazi notes software engineering hiring rose even as agents took over more coding, because the freed capacity moved to taste and judgment problems.
Workflows are still bottoms-up: Most AI use inside firms is individual experimentation, path-dependent on how each investor first picked up a tool. The workflows that become policy, such as red-teaming every investment memo with an LLM, spread only after someone proves them.
MCP turns software into an agent layer: Model Context Protocol lets an investor prompt "build me a report on this company" and have the tool make the underlying API calls. Vertical products stay relevant as the data and ontology under the prompt, an iceberg beneath the chat window.
Single-player versus multiplayer decides build or buy: Internal research tools that only the deal team uses are good candidates to build. Anything portfolio companies and auditors also touch needs a verifiable audit trail from a trusted third party, which a vibe-coded internal system cannot give.
Reserves are the under-scrutinized alpha: Initial investments get customer calls, references, memos, and committee debate. The next check in the same company often gets a sentence at a partner meeting, despite reserves making up 30 to 50% of many funds.
AI growth at the top decile is unprecedented: Across more than 12,000 companies on the platform, AI companies grow faster at median across revenue bands and far faster at the top decile, helped by capital access that let mediocre or negative gross margin businesses leap the valley of death.

By The Numbers

30 to 50%: Share of many venture funds held as reserves, deployed through decisions that get far less rigor than the initial check.
12,000+: Portfolio companies on the Standard Metrics platform, the base for its cross-sectional analysis.
$100 million+ revenue: The band where, per a recent report with Emergence, AI companies' revenue per FTE crossed above non-AI companies for the first time, after lagging for years because cheap capital funded over-hiring.
68: Standard Metrics employees, with the CEO still interviewing every hire to screen for mission and values.
4 of 4: Engineering managers who started as individual contributors on the team.

Key Trends to Watch

Follow-on decisions getting initial-investment rigor: Watch whether firms adopt net-buyer or net-seller discipline at every round, with reserve planning treated as a scored process rather than a pro-rata default.
Firm-level AI mandates replacing individual experimentation: The memo red-team is the template. Expect more grassroots workflows to harden into required steps, and more firms to hire full-time AI engineers, already common in Melas-Kyriazi's customer base.
Data vintages losing shelf life: 2021 data is irrelevant to 2023, and 2023 top-decile growth is irrelevant to 2026. Benchmark refresh cycles will shorten as rate and technology shocks reset the curves.
AI-native services as the middle category: Businesses with venture-like growth and private-equity-like models will test whether human accountability, which a model cannot carry, remains the reason customers pay for services on top of software.

Memorable Quotes

"The frontier of where human beings create the most value continues to push more and more toward workflows or activities where it's very difficult to build a good eval." The mechanism behind the whole argument.
"AI can't be punished for making a mistake." Why responsibility, and the services that carry it, survives automation.
"The fastest growing AI companies are growing faster than any companies ever in human existence." What the top decile of the dataset shows, and why capital concentrates there.
"Hire slow and screen relentlessly for mission and values alignment." The advice he would give himself six years ago, and the cultural precondition he sets for an AI-native firm.

The Wrap

The thesis holds if firms that automate document parsing, research, and reporting redeploy the recovered hours into founder evaluation, portfolio support, and disciplined reserve decisions, and if those firms show measurably better fund outcomes than peers using AI only for speed. It also holds if multiplayer tools with audit trails keep winning over internal builds as coding agents keep lowering the cost of building. It fails if evals improve enough to score founder quality and follow-on timing, collapsing the judgment frontier further than Melas-Kyriazi expects, or if firms that build everything in-house prove auditors and portfolio companies will accept it. The 2024 to 2027 fund vintages, the first raised and deployed with these tools in place, will show which firms turned time into judgment.