The Capital Thesis Needs an Update

Venture capital has bet overwhelmingly on labor displacement as the AI growth narrative. AI firms captured 61% of all global VC investment in 2025, pulling in $258.7 billion out of a $427.1 billion total market, according to an OECD analysis published in February 2026. That share climbed to roughly 80% of global VC in Q1 2026, driven by frontier-lab mega-rounds from OpenAI ($122 billion), Anthropic ($30 billion), and xAI ($20 billion). The implicit model: AI replaces headcount, margins expand, multiples justify. But Shipper's ground-level data from an organization that has done more automation than most suggests the actual dynamics are more complex, and the resulting opportunity set is different from what the displacement thesis implies.

Anthropic's own economists documented the gap between AI's theoretical and actual labor-market footprint in a March 2026 paper co-authored by head of economics Peter McCrory. They found that while AI can theoretically cover the majority of tasks in computer science, financial management, and legal work, observed Claude usage across enterprises is a fraction of that theoretical ceiling. The gap between capability and deployment is not a temporary adoption lag. It reflects a structural requirement that someone with relevant expertise must frame the problem before the model can work on it.

The Frame Problem No Benchmark Captures

Shipper builds his argument around what he calls the frame problem. Benchmarks measure how well a model performs inside a problem definition that a human has already supplied. On OpenAI's GDPval benchmark, which tests AI performance against expert-level tasks across occupations including compliance officers, lawyers, and software developers, Claude Opus 4.1 outperformed human experts 49% of the time. The headline number generated a round of displacement coverage. What it obscured: the benchmark prompts for those tasks came pre-loaded with precise confidence intervals, enumerated criteria, named entities to include, and output format specifications. An enormous amount of expert judgment was already encoded into the frame before the model ran a single token.

Shipper's in-house Senior Engineer benchmark makes the same point from the other direction. A coding agent given a clear instruction to perform a "clean first-principles structural rewrite" of a broken codebase scored 62/100 on its best run for GPT-5.5, roughly 30 points above competitors. Change the prompt to "solve all the errors that keep popping up," and the score collapses toward zero. The model's performance is inseparable from the quality of the frame a human constructed around the task.

This is not a bug the next model will fix. It is a property of how language models are built. Models train on the recorded outputs of completed work. They have no access to the present-tense judgment required to decide which problem to frame, why now, at what scope, and against which constraints. That judgment must come from somewhere. Under current and near-term architecture, it comes from humans.

The Abundance Cycle and What It Funds

Shipper’s second mechanism is economic; when a rare skill becomes cheap, demand for that skill expands. Operations staff at Every now issue pull requests they never would have attempted before: marketers produce video thumbnails in minutes, engineers draft product copy and the volume of work in each category explodes. But the default output of models trained on the same corpus trends toward sameness, and sameness becomes a commodity quickly. The result is increased demand for the humans who can identify what differentiates good output from adequate output in a specific context.

The pattern shows up in the cost of automation itself. One of Every's PowerPoint automation workflows involves 24 skills and 18 scripts and costs $62 in tokens per deck. That is a new class of infrastructure that requires ongoing human maintenance to stay calibrated. OpenClaw's open-source repository, referenced by Shipper as a proxy for the scale of AI-assisted development activity, had received 44,469 pull requests as of mid-May 2026, with nearly 4,000 in the first three weeks of May alone. For context, Kubernetes received 5,200 pull requests in all of 2022. The volume of AI-assisted work being produced globally has no historical precedent. Reviewing, directing, and maintaining that output is work that requires people.

What This Means for the Market

For investors, the practical implication is a market map that diverges sharply from the pure labor-replacement playbook. Companies that build around expert augmentation rather than headcount reduction, that sell into the review-and-calibration workflow rather than the task-execution layer, and that serve the growing infrastructure requirements of human-agent collaboration are positioned for durable demand regardless of how benchmark scores move.

The enterprise buyers who have moved fastest on AI deployment are not reporting empty org charts. They are reporting new categories of work: AI engineers who maintain agent workflows, senior practitioners who review AI-generated output at scale, and domain experts who translate live business context into the problem frames that make models useful. That is not the story that justified $300 billion in Q1 2026 VC. It may be the story that justifies the next $300 billion.

The harder question, which Shipper does not resolve, is whether the expert-augmentation layer generates enough economic surplus to offset displacement in lower-skill roles. Anthropic CEO Dario Amodei has warned that AI could eliminate up to half of entry-level white-collar jobs. The two claims are compatible: expert work expands at the top of the distribution while entry-level work contracts at the bottom. Which dynamic dominates the next decade is the most consequential open question in the labor economics of AI, and no benchmark yet built can answer it.